Project 2.

2
Equipment Scheduling and Cycling
Task 2.2.3
Automated Diagnostics ased on Start!"p Signals
Deli#era$le 2.2.3%a&
De#elopment o' a (unctioning Centrally )ocated Electrical!
)oad *onitor
Su$mitted to Architectural Energy Corporation
"nder the
Cali'ornia Energy Commission+s
Pu$lic ,nterest Energy -esearch %P,E-& Program
Energy!E''icient and A''orda$le
Small Commercial and -esidential uildings
Cali'ornia Energy Commission Contract .//!00!/11
2. D. )ee3 ). 2. 4or'ord and S. . )ee$
*assachusetts ,nstitute o' Technology
*ay 253 2//3
De#elopment o' a (unctioning Centrally )ocated Electrical!)oad
*onitor
T6,S -EP7-T 8AS P-EPA-ED AS A -ES")T 7( 87-2 SP74S7-ED 9 T6E CA),(7-4,A E4E-:9 C7**,SS,74
%C7**,SS,74&. ,T D7ES 47T 4ECESSA-,)9 -EP-ESE4T T6E ;,E8S 7( T6E C7**,SS,743 ,TS E*P)79EES3 7-
T6E STATE 7( CA),(7-4,A. T6E C7**,SS,743 T6E STATE 7( CA),(7-4,A3 ,TS E*P)79EES3 C74T-ACT7-S3
A4D S"C74T-ACT7-S *A2E 47 8A--A4T93 E<P-ESS 7- ,*P),ED3 A4D ASS"*E 47 )E:A) ),A,),T9 (7- T6E
,4(7-*AT,74 ,4 T6,S -EP7-T= 47- D7ES A49 PA-T9 -EP-ESE4T T6AT T6E "SE 7( T6,S ,4(7-*AT,74 8,))
47T ,4(-,4:E "P74 P-,;ATE)9 784ED -,:6TS. T6,S -EP7-T 6AS 47T EE4 APP-7;ED 7- D,SAPP-7;ED 9
T6E C7**,SS,74 47- 6AS T6E C7**,SS,74 PASSED "P74 T6E ACC"-AC9 7- ADE>"AC9 7( T6E
,4(7-*AT,74 ,4 T6,S -EP7-T.
Overview of project
This report documents the development of a functioning centrally located electrical-load monitor.
The non-intrusive load monitor (NILM) is a personal-computer based instrument that samples
voltage and current at a single location and estimates the electrical energy drawn by individual
pieces of euipment. NILM results can also be used for detecting and diagnosing faults.
! "h.#. dissertation by $wangdu% #ouglas Lee constitutes this report. The thesis& titled
'(lectric Load Information )ystem based on Non-Intrusive "ower Monitoring&* is included in its
entirety. Mr. Lee ac%nowledges the contributions of "rof. )teven )haw at Montana )tate& +red
)mothers and Ted Ludwic%& both formerly associated with the test-site office building& and "eter
!rmstrong& ,obert -o. and -hris Laughman at MIT.
The "--based NILM& running the Linu. operating system& uses a data-acuisition card to sample
voltage and current at high speed& about /0 %12. It then +ourier transforms the data to obtain
real and reactive power at the fundamental and higher (up to seventh) harmonics. These data are
stored at /34 12 for subseuent analysis. !t the beginning of this pro5ect& the on-line NILM was
capable of detecting loads via analysis of the electrical-power transients observed during start-up
of euipment. !ll other analysis& notably detection of loads via changes in steady-state power
and detection of faults& was done off-line& via a combination of algorithms and intervention of an
analyst.
"ro5ect 3.3& '(uipment )cheduling and -ycling&* has focused on developing a more fully
automated NILM. The pro5ect has had three& yearly phases. The first phase was based on
detecting loads from step changes and made use of data collected at the Iowa (nergy -enter6s
(nergy ,esearch )tation. The second phase focused on fully automated detection of euipment
operation. This phase developed the algorithms documented in #eliverable 3.3.3c& 'Intermediate
,eport on !utomated #iagnostics.* The delivered report included e.amples of naturally
occurring faults that the NILM detected& and that would be very difficult to detect with more
traditional thermo-fluid fault-detection methods. 1owever& detection of faults must build on
automated detection of normal euipment operation. In particular& power oscillations associated
with such faults as a chiller vacuum lea%& a misaligned fan shaft& and under-damped feedbac%-
control loops& are detected with the same freuency-analysis techniues used to trac% variable-
speed drives. !ccordingly& the wor% in the third year& initiated intended to focus on automatic
detection of faults& instead concentrated on further developing load-trac%ing algorithms and
coding them for on-line use in the NILM. This report will be supplemented by a -# with NILM
code. The remainder of this summary highlights the ma5or advances documented in this report
and identifies future wor%.
Use of both steady-state load detection and analysis of start-up transients
This research has developed a load disaggregator that analy2es both step changes in steady-state
loads and start-up transients and automatically arbitrates the results in cases where the two
methods disagree. The step-change analy2er is significantly more powerful than that used in
previous versions of the NILM but is still sub5ect to bac%ground electrical noise. The transient
detector is relatively immune to white noise but can be fooled by colored noise and is not useful
for abrupt shut-down signals. If single-event load detection fails& the signal is sent to a multi-
event detector that searches a binary tree of possible combinations of temporally overlapping
loads. The step-change analy2er& the multi-event detector and the state estimator& discussed ne.t&
ma%e use of a consistent signal-analysis framewor% that considers load signals to be stochastic
events& each associated with a mean and variance.
State estimation
(vent detection is susceptible to occasional errors7 turn-on and turn-off events will be missed
entirely or mislabeled. )uch errors can accumulate and reduce the value of the NILM. +urther&
the NILM needs a starting point when it is first started& so that the status of individual loads can
be determined before turn-on or turn-off events are detected. To address these issues& a state
estimator was developed to determine the most li%ely set of loads associated with a given steady
power signal.
Estimation of power drawn by variable-speed drive (VSD) loads
)ince its inception& the NILM has successfully detected only constant-power loads. 8)#s are in
common use and have posed significant challenges. (ven if the power drawn by a 8)# fan were
nearly constant over the day& or could be appro.imated as such with acceptable loss of accuracy&
detection of the start-up of a 8)# is nearly impossible with either step-change or start-up
transient algorithms& due to the very long (order of /4 minutes) start-up period. This report
documents a fully functional approach to trac%ing 8)#s that is based on observing the higher
harmonics& particularly the fifth and seventh harmonics of real and reactive power that are
associated with 8)#s. This method cannot separate individual 8)# devices but can trac% their
aggregated power use.
estin! of software
NILM algorithms were developed in Matlab and tested off-line with NILM-collected data&
primarily from the test office building in )an +rancisco. The algorithms were converted to -99&
compiled& and embedded in the on-line NILM in the office building. The code was sub5ected to
limited testing& in two phases. +irst& the disaggregator for constant-power loads was tested by
in5ecting off-line data into the NILM. This procedure was necessary at the time to isolate
3
constant-power loads. The code successfully identified loads and maintained a log of their on-off
status& presenting this information via a :eb-based interface. )econd& the 8)# trac%ing
software was loaded and tested. It successfully separated the supply- and return-fan 8)# power
from the 18!- total.
"uture wor#
)everal important steps need to be ta%en7
- Trac%ing of variable-power& constant-speed loads. This last remaining load class includes
most chillers and those fans still using inlet vanes. ,elations between real and reactive power
for these loads will be e.plored. #irectly trac%ing a chiller is preferable to assigning to the
chiller that portion of 18!- power not due to constant-power and 8)# loads& because
chiller power 'by subtraction* is sub5ect to errors in following other loads. 1owever& the
subtraction approach is possible if the NILM is mounted at the 18!- service entrance.
- Integration of the disaggregator for constant-power loads and the 8)# trac%er. The code was
designed to be integrated. 1owever& most constant-power loads are for pumps and tower
fans associated with operation of a chiller. ;ntil the chiller can be detected or is assigned
remaining 18!- power& the integration has reduced practical value.
- Testing of the NILM and determination of its commercial value. This wor%& the most
important& reuires that NILM load-trac%ing output be compared with submeters and that a
potential commerciali2er determine whether NILM output& either load trac%ing or power
signatures associated with faults& has value to its customers.
- !utomated training of the NILM. The state estimator is a good start here& because it ma%es a
best guess of the loads in operation at the time the NILM is turned on. 1owever& the state
estimator and the load disaggregator rely on a %nowledge of the characteristics of individual
loads& including steady-state real and reactive power and start-up transients. To date& this
information has been gathered manually& by turning on and off individual loads several times
and analy2ing the power data. +or commercial applications& the training must be automated
to the e.tent possible and clear instructions must be prepared for remaining& manual steps.
- !utomated detection of faults. !s a starting point& detection of faults associated with
oscillatory power signals should be automated& based on freuency analysis. !utomatic fault
detection will aid those potential users who have limited time or ability to analy2e NILM
data.
Ta$le o' Contents
Introduction.......................................................................................................................... 0
/./ Motivation and bac%ground of research.....................................................................0
/.3 )ystem overview and e.ample power waveforms..................................................... <
/.= >rief history of non-intrusive power monitoring..................................................... /3
/.? @b5ectives and outline of thesis .............................................................................. /?
-urrent 1armonic "owers7 Mathematical Theory and "ractices....................................... /A
3./ Introduction.............................................................................................................. /A
3.3 +undamental powers................................................................................................ /<
3.= -urrent harmonic powers.........................................................................................34
3.? Method to compute current harmonic powers......................................................... 3=
3.B Implementation issues..............................................................................................3C
3.0 (.amples..................................................................................................................=4
3.A -hapter summary..................................................................................................... =?
)ingle Load (vent Identification using )teady-state Load -lassification and Transient
"attern ,ecognition............................................................................................................=0
=./ Introduction.............................................................................................................. =0
=.3 Load classification based on state space mapping................................................... =A
=.= (vent associator....................................................................................................... ?4
=.? )teady-state event classifier..................................................................................... ?=
=.B Test of steady-state load classifier........................................................................... ?C
=.0 Load -lassification based on transient pattern recognition..................................... ?<
=.A Transient pattern classifier....................................................................................... B3
=.C Test of transient pattern classifier............................................................................ BA
=.< -hapter summary..................................................................................................... 0/
Identification of Temporal @verlap of (lectric Load (vents via (vent )pace )earch and
Transient "attern -omposition.......................................................................................... 03
?./ Introduction.............................................................................................................. 03
?.3 @verview of multi-event handling........................................................................... 0=
?.= (vent space search................................................................................................... 0?
?.? Transient pattern composition..................................................................................0<
?.B Test of multi-event handler...................................................................................... AB
?.0 -hapter summary..................................................................................................... A<
(lectric Load )tate (stimation...........................................................................................C3
B./ Introduction.............................................................................................................. C3
B.3 @verview of load state estimation............................................................................C=
B.= )tate estimation........................................................................................................ C?
B.? Test of load state estimator...................................................................................... <3
B.B -hapter summary..................................................................................................... <0
(stimation of 8ariable )peed #rive "ower -onsumption from 1igher 1armonic "owers..
<C
0./ 8)# load model.......................................................................................................<C
0.3 8)# power estimation method.............................................................................. /40
0.= Test of 8)# power estimation method..................................................................///
0.? -hapter summary................................................................................................... //B
(stimation and (limination of 1armonic )ignals from "ower :aveforms.................... //0
?
A./ Introduction............................................................................................................ //0
A.3 #etection and estimation of harmonic signal from periodogram...........................//A
A.= #etection and pitch estimation of harmonic signal using autocorrelation............. /30
A.? +ractional pitch period estimation and least suares signal estimation................. /==
A.B !utoregressive modeling of harmonic signal and white filtering.......................... /?4
A.0 -hapter summary................................................................................................... /?A
Test ,esults& -onclusion and +uture :or%..................................................................... /B4
C./ Test results............................................................................................................. /B4
C.3 -onclusion............................................................................................................. /B0
C.= +uture wor%............................................................................................................ /B<
Deneral -hange-of-meanEvariance #etector.................................................................... /0=
!ir 1andling ;nit Modeling............................................................................................/A/
>./ 8!8 model........................................................................................................... /A/
>.3 )ustained oscillation- limit cycle.......................................................................... /AA
>.= -onditions to e.hibit limit cycle........................................................................... /C=
(stimation of +irst @rder "olynomial Mean of a ,andom "rocess................................. /CA
(LI) )oftware !rchitecture............................................................................................. /<4
>ibliography...................................................................................................................../<=
Chapter 1

,ntroduction
1.1 *oti#ation and $ackground o' research
(lectric power has been with us for more than a century and surrounds us in the form of
!- wall outlets& power supplies for mobile phones and laptops& power sources for the
onboard devices of automobiles& etc. It affects every aspect of our daily lives and our
daily activities depend on its implicit presence. It is undoubtedly one of the core
foundations of the modern industrial and information society. 1owever& this once well-
established technology recently faces three ma5or challenges7 deregulation& autonomous
power systems and power quality.
The purpose of the deregulation of the electric utility industry is to introduce competition
to the traditional monopoly F:ollenberg& /<<CG "hilipson and :illis& /<<CG Haccour&
/<<CI. >efore deregulation& the electric utility was vertically integrated& performing
generation& transmission& distribution and retail under one umbrella. It was given
e.clusive sales rights in specific regions& in return for the acceptance of regulatory
supervision and the guarantee of reliable service. !fter deregulation& those functional
units become independent entities with more hori2ontal relationships and participate in
open mar%ets FLee and Lin& /<<CI. The original intent of deregulation was to increase
efficiency& lower price and improve& or at least retain& the uality of service F>ar%ovich
and 1aw%& /<<0I. The deregulation of electric utilities is an inevitable trend& as it was the
case in the telecommunication and gas industry& but it can generate many adverse effects
along its course F#raper& /<<CG !bdoo& /<<<G Ilic et al.& 344/I. The most notable e.ample
is the recent -alifornia electricity crisis F>udhra5a& 3443G -row& 3443G McNamara& 3443G
)weeney& 3443I.
(lectricity is traded as a commodity in deregulated mar%ets. Denerators sell their capacity
on a wholesale mar%et. #istribution companies bid in a spot mar%et for best prices and
resell the electricity to their local customers. In an open mar%et& the ma.imum utili2ation
of e.isting resources and the acuisition and communication of information become
critical for the survival and profitability of the participants and for the prosperity of the
mar%et itself F1arris& 3444I. To this end& concerned parties have invested in and
developed various real-time transaction& trade and monitoring systems& fueled by
advances in computer and communication technologies FDill et al.& 344/G Mc-lanahan&
3443I. 1owever& one party is particularly unprepared for these changesJ the customers&
who are on top of the electricity food chain and the alleged beneficiary of all those
reforms FIlic et al.& 344/I.
-ustomers need to ma%e educated decisions in their energy purchases& if they have an
option to choose among different vendors and if the electricity price changes in real time.
The selection of a vendor will depend on the overall uality of service& including price&
reliability and customer relationship. !fter a service provider is selected& the customer
needs a monitoring tool to ensure that the desired level of service is being provided& and
0
to verify whether its energy consumption is optimal. ,eal-time pricing ma%es the
situation more complicated& because the customer must %now how much power is
consumed and what loads are being used& in real time& to ma%e prudent purchase
decisions. ;tilities are moving from monthly metering to interval metering& with metering
intervals as short as fifteen minutes FMc-lanahan& 3443I. It is rather unclear whether the
utilities will share metered data in real time& along with current prices and price forecasts&
with customers. (ven if metered data are available to end users& they will be the
aggregated sums of powers consumed by individual loads with no specific detail about
the nature of power consumption.
The manager of a commercial or production facility may opt to %now the state of electric
loads under his or her supervision to ma%e the best energy purchase decision sub5ect to
mar%et forces. 1owever& this information is typically unavailable unless a manual
inspection is performed or a separate instrument is installed at every single load. >oth
options are cost prohibitive and difficult to implement. In a deregulated mar%et&
customers need a ubiquitous electric meter& which can provide them with itemi2ed
information in real time to assist their energy purchase decisions.
8arious autonomous power systems are emerging on the hori2on& gaining more and more
significance. They refer to electric systems& independent or semi-independent from the
utility grid. )elf-sufficient power systems on-board automobiles or ships are good
e.amples. Isolated power systems based on alternative or renewable sources& such as fuel
cells and wind energy& are more widely used F:ol%& /<<<G -oo%& 3443G :al%er& /<<CG
>ull& 344/I. They are sometimes connected to the grid (utility-interactive) as a form of
distributed or dispersed generation FNai% et al.& /<<BG "rice et al.& /<<<I.
They pose difficult challenges such as how to generate and distribute power efficiently
within a self-contained system& and how to manage interactions if it is connected to a
grid. +or e.ample& the generation and distribution of electricity on automobiles becomes
more critical& as the auto manufacturers install more power-intensive components for
better fuel economy& safety and comfort F$assa%ian et al.& /<<0G $assa%ian& 3444G (madi
et al.& /<<<I. ;pgrade to a ?3 8 #- system is one of the answers to this inevitable trend
FNeubert& 3444I. !- distribution systems are also considered for automobiles FMiller&
/<<0G $han& /<<<I. !nother e.ample is how to connect a fuel cell generator to the utility
grid. It should able to be configured either to send power to or receive power from the
grid& while serving its local load F+arooue and Maru& 344/I. Those challenges can be
met only by the accurate %nowledge of the state of its own system& and the networ% when
connected.

The issue of power uality is getting more attention& partly because of the conseuences
of the first two challenges and partly because of the increased use of sensitive and
nonlinear electric loads F$ennedy& 3444G )tones and -ollinson& 344/I. The deregulation
of the electric utility seems to have degraded& at least temporarily& the reliability and
uality of power& leading to several blac%outs and brownouts and numerous disruptions
and disturbances F$endall& 344/G !rrillaga et al.& 3444I. !utonomous power systems
often e.perience wiring or ground faults& such as a short circuit between /? 8 and ?3 8
rails on automobiles FNeubert& 3444I& and can suffer severely if they are mission-critical
systems& e.g. modern naval combat vessels and spacecrafts pac%ed with computers and
electronics.
The increased use of semiconductor-based power electronic devices creates severe
harmonic distortions in the current and voltage waveforms. 8oltage distortions can lead
to the malfunctions of euipment and protective circuit elements. The widespread use of
information processing euipment e.acerbated the situation& but gave rise to the public
awareness of the importance of the power uality. (very concerned party desires a type of
affordable power uality monitoring tool& possibly operating in real time& to characteri2e
the distortions and disturbances and ta%e necessary actions. Modern power uality
monitoring tools can detect and classify disturbances and are networ%-capable
F:i5aya%ulasooriya et al.& 3443G >ucci et al.& 344=I. (nd users often use the results to
specify power-uality-tolerant euipment F@liver et al.& 3443I. There also are attempts to
add power uality monitoring capability to the smart meters in the deregulated mar%et
F)aari et al.& /<<0G $oponen et al.& /<<0I.
In author6s viewpoint& the prime motive behind these challenges faced by the electric
power systems is the fact that we are now entering a new era of information society. The
power system infrastructure and practices were established during the first half of the
twentieth century& when the electricity became the basis of the energy-centered industrial
society. The second half of the last century saw the electricity e.panding its role also as a
basis of the new information society. -onseuently& by the turn of the century& the
e.isting installations and established technologies were inadeuate to meet the
reuirements of the new era& when the electricity is e.pected to perform the dual role of
the medium of both energy and information.
Ironically& possible solutions of the problems caused by the new needs of the information
society will reuire the %nowledge of the electric systems they intend to improve. In other
words& the information of the power systems will be the %ey in meeting the challenges
posed by the fledgling information society.
1owever& obtaining high uality information economically and reliably is always a
difficult ob5ective to achieve. The Non-Intrusive "ower Monitor (NI"M) is a novel
attempt to collect such information with a minimal physical installation. ,aw voltage and
current are measured at a single location to yield harmonic power signals. They typically
carry the fingerprints of the electric loads present in a system& and their analysis can
produce such information as the operational and diagnostic status of the loads. The power
signals can also be used for the system identification& parameter estimation and energy
consumption optimi2ation study.
This research was mainly sponsored by the -alifornia (nergy -ommission and
inaugurated in the summer of 3444& at the height of the -alifornia energy crisis. Its main
purpose was to develop an intelligent yet affordable monitoring device so that the
consumers can ma%e better choices in a deregulated and volatile electricity mar%et. It
focuses on the development of mathematical models and signal processing algorithms to
e.tract information for the customers at the end of the !- distribution system. Its results
will directly benefit the developments of a ubiuitous electric meter in a deregulated
C
mar%et& a diagnostic or prognostic tool for mission-critical systems and an intelligent
power uality monitor. :ith minor modifications& the results will also be useful for other
sectors& such as #- electrical systems and electric utilities.
1.2 System o#er#ie? and e@ample po?er ?a#e'orms
The NI"M is typically installed at the main electric service entry of a system& and thus
can monitor the powers consumed by the individual loads in the system. The NI"M
samples raw voltage and current waveforms and computes harmonic powers from the
samples. The harmonic power signals are analy2ed by various detection& estimation and
classification algorithms to e.tract desired information& such as the @NE@++ status of the
loads& e.istence and diagnosis of faults& optimality of energy usage& etc. The information
will be displayed to users& so that they can %now the system state and ma%e appropriate
operational and energy purchase decisions. The information may also be forwarded to an
electric load control system to ad5ust the state of the loads following certain guidelines&
such as minimi2ing energy usage while not compromising comfort or productivity. In this
thesis& the (lectric Load Information )ystem ((LI)) denotes the overall information
e.traction and handling system. +igure /-/ shows the overall schematic of the system.

K/ K3
L
Kn
I
8
;tility
#istribution
Transformer
-ustomer
Main
"anel
Individual Loads
#ata
!cuisition
"ower
-oder
"ower )ignal
!naly2er
(lectric Load
-ontroller
;ser
voltage
signal
current
signal
harmonic
power signal
electric load
information

+igure /-/ )chematic of (LI) based on NI"M.
+igure /-3 shows a daily power consumption pattern of a commercial building logged by
an NI"M. The building has large variable speed drives (8)#s) for its ma5or air-handling
unit. The 8)#s turn on around A7/B and turn off at /A7=4. These events are clearly visible
in the graph. The building also has a large chiller. It turns on slightly after C744 and turns
off around /A744. !gain& these events are clearly visible in the picture. The chiller is
euipped with capacity control. In the morning when the cooling load is minor& its power
consumption is relatively flat. 1owever& after //744& as the cooling load increases& the
chiller starts to consume more power and its power consumption varies with the varying
cooling load. )mall spi%es in the morning and the late afternoon are turn-on and turn-off
events of a single-speed cooling tower fan. >ecause the fan has no capacity control& it
periodically turns on and off when the cooling load is minor. 1owever& once the cooling
load e.ceeds a certain threshold& it operates continuously.

7 8 9 10 11 12 13 14 15 16 17 18
0
50
100
150
200
250
300
time (hr)
P

(
k
W
)
10/15/01
+igure /-3 #aily power consumption profile of a commercial building.
+igure /-3 suggests that by analy2ing the non-intrusive power waveforms& we can identify
load events and thus recogni2e which loads are on at a given time. The analysis may be
performed either in situ or offline. The %nowledge of load events allows us to infer the
behavior of the monitored loads. +or e.ample& the %nowledge of the freuent on-off
events of the cooling tower fan lets us %now that the cooling load at that moment is
minor. !lso& if its on-off freuency is too e.cessive& we may decide to ta%e an action such
as increasing the dead-time 2one in its @NE@++ control or add a capacity control.

+igure /-= shows another power waveform ta%en from the same building. @n that day&
there was a lea% in the evaporator chamber of the chiller. !s a result& outside air
infiltrated the refrigeration cycle. >ecause the air is lighter and less compressible than the
refrigerant& the centrifugal compressor of the chiller sees a sudden decrease in its load
when the entrapped air enters its compression chamber& resulting in a rapid decrease in its
electric power consumption. The graph shows the onset of this phenomenon and because
the air is circulated along with the refrigerant& the power fluctuates almost periodically. It
too% two hours for the engineer of the building to get to %now the situation& by the
symptom of lost cooling& and stop the chiller to purge the evaporator manually. >ut
during those two hours& the chiller was sub5ect to an e.cessive amount of wear and
thermal comfort was lost.
This e.ample suggests that certain faults are visible in power waveforms and that the
power signals can be used for their detection and possibly for their identification. +or
e.ample& a very simple variance estimator may be designed and when the variance
estimation e.ceeds a certain threshold over a prolonged period& the (LI) may decide that
there e.ists a significant fault. Identifying the nature and cause of a fault is much more
/4
difficult. It will inevitably involve accumulating the %nowledge of the faults and decide
the character of a detected fault& using a classification or pattern recognition algorithm.

12.412 12.414 12.416 12.418 12.42 12.422 12.424 12.426 12.428
120
140
160
180
200
220
240
260
280
time (hr)
P

(
k
W
)
11/08/01
+igure /-= @nset of chiller evaporator vacuum lea%.


0 50 100 150 200 250 300 350 400 450 500
72
74
76
78
80
82
84
86
88
90
92
sample number (120 Hz)
P

(
k
W
)
11/21/02
+igure /-? 1armonic signal present in power waveform.
+igure /-? shows a harmonic signal present in a power waveform. Its periodogram
/
in
+igure /-B clearly shows the spectral pea%s in a harmonic relationship. In power
engineering& these harmonic signals are called interharmonics& because their freuencies
are not the multiples of the fundamental freuency of the voltage waveform& i.e. 04 12 in
the ;). -ommon causes of the interharmonics are ripples in a #- bus between a rectifier
and an inverter and periodically changing loads. It usually is necessary to eliminate the
/
"eriodogram is defined in -hapter A. It is a natural estimator of the power spectral density (")#) of signal.
interharmonics in analy2ing a power signal& because they can interfere with the event
detection and classification algorithms& which rely largely on the change of mean
detection and the uniueness of event patterns. @n the other hand& an interharmonic may
be actively estimated& if it can be associated with the physical status of a load. +or
e.ample& if the interharmonic is generated by a periodic pulsation of a rotating machine&
e.tracting its parameters& such as amplitude and freuency& can be e.tremely useful to
monitor the machine. The amplitude estimation can be used to uantify the severity of the
pulsation& and the freuency estimation can be used to monitor its rotation speed.

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-40
-30
-20
-10
0
10
20
30
40
Periodogram of real power
d
B
normalized frequency
+igure /-B "eriodogram of harmonic signal.
The e.ample graphs in this section illustrate that a rich amount of information is
embedded in non-intrusive power signals and that appropriate analysis techniues will be
necessary to e.tract those information. :e certainly want most analyses to be performed
automatically by a computer& in real time if possible& although certain higher-level
analyses may still have to be performed by human users. Thus& the goal of this thesis is to
identify those tas%s suitable for computers and to develop various signal processing and
information handling algorithms to accomplish them.
1.3 rie' history o' non!intrusi#e po?er monitoring
The non-intrusive power monitoring was originally called non-intrusive appliance load
monitoring (N!LM)& and later called non-intrusive load monitoring (NILM) to imply that
the monitored loads are not necessarily household appliances. The author prefers to call it
the NI"M& in the sense that power is the signal we observe& whereas load is the
information we have to e.tract from the observations. It also is in line with the
philosophy of this thesis& in which the functionality of lower level non-intrusive power
signal acuisition is separated from that of higher level power signal analysis to e.tract
the electric load information.
/3
The N!LM was first pioneered by Deorge 1art in the /<C4s F1art& /<<3I. The original
goal was to provide electric utilities with a cost-effective monitoring tool& which can
collect the energy consumption data of their residential customers with minimal intrusion
on their properties. 1e resolved this problem by monitoring the power at a metering
location. 1owever& because the metered power is a lumped sum& the individual power
consumed by a load was not readily obtained. 1e developed a novel solution for this
problem based on the fact that a constant load tends to have a uniue power consumption
level and that its @N and @++ events show up as distinctive step changes in time plots.
1e measured the real and reactive powers at a / 12 sampling rate and the step changes
were recorded and sent to a central processing location via a telephone modem. The
archived step change data were plotted on a 3-dimensional (real-reactive) power plane&
and classified based on a distance metric from cluster centers. 1e also developed a load
event seuence analysis program based on the 8iterbi algorithm to e.tract the time
history of electric loads from the seuence of classified load events. The (lectric "ower
,esearch Institute ((",I) developed a prototype residential load monitoring system based
on his wor% F-armichael& /<<4G #ren%er and $ader& /<<<I. @ther residential power
monitors are also reported in the literature F)ultanem& /<</G -ole and !lbic%i& /<<CI.
1owever& his algorithm is limited in a sense that it can detect only the isolated @NE@++
events of constant loads. +or e.ample& if two separate events happen at a same time& his
classifier either fails or misclassifies& ma%ing subseuent load event seuence analyses
more complicated. !lso& some loads have multiple power consumption levels& e.g. a
heater with a L@ and 1I switch& and those levels may follow a certain seuence& e.g. a
dishwasher or washing machine. !lthough he proposed a finite state machine (+)M)
model to handle those multi-stage loads& he did not devise any actual classification or
seuence analysis algorithm. !nother shortcoming is the fact there is a class of loads that
do not have fi.ed power consumption levels- variable loads. ! light dimmer is an
e.emplary residential variable load. This issue becomes more prevalent as one tries to
monitor commercial or industrial loads& where significant portions are variable loads&
such as variable speed drives. 1is approach also lac%s real time analysis capability.
)teven Leeb approached this problem from a slightly different angle& based on his
observation that electric loads tend to have uniue patterns when they are turned on
FLeeb& /<<=I. To observe the transient patterns& it was necessary to reduce the time scale.
1e developed an analog spectral envelope processor& to estimate the real& reactive and
harmonic powers at the fundamental freuency (e.g. 04 12) of the voltage waveform. 1e
also developed a rudimentary transient detection and pattern classification algorithm and
proposed to e.amine power waveforms at multiple time resolutions to search for desired
patterns FLeeb et al.& Muly /<<BI. 1e and Leslie Norford investigated the feasibility of
using the NILM for commercial buildings FNorford and Leeb& /<<0I.
)teven )haw first developed a digital spectral envelope processor by simulating the
functionality of the analog processor in the digital domain F)haw et al.& /<<CI. 1e later
developed an improved processor based on the discrete +ourier transform (#+T) of
waveform samples F)haw& 3444I. >ecause the spectral envelope estimation was
performed on a computer& it also provided an opportunity to do the analysis locally in real
time. 1e developed a transient pattern matching algorithm based on the method of least
suares. 1owever& he did not devise a load status analysis routine& which can be lin%ed to
his transient event classifier. 1e also performed motor parameter estimation and
diagnosis based on raw current waveform measurements F)haw and Leeb& /<<<I.

#ong Luo modified Deorge 1art6s algorithm to analy2e the building heating& ventilation
and air conditioning (18!-) loads F#. Luo& 344/I. 1e was able to e.tract most of their
usage history from the archived power measurements. 1e also used the power signal to
detect and diagnose certain faults present in the 18!- systems F)haw et al.& 3443I.
+or the last twenty years& there have been many advances in monitoring electric loads
nonintrusively to e.tract their status information. 1owever& many problems remain
unsolved& such as the event overlap& multi-stage load state estimation& variable load
trac%ing& real-time local processing& etc. !lso& no practical system has been offered in the
mar%et& whether it is residential& commercial or industrial. This thesis is an attempt to
address these longstanding issues and lay a foundation for developing a commercial-
grade electric load information system based on the non-intrusive power monitoring.

1.. 7$jecti#es and outline o' thesis
Traditionally& electricity has been considered as a medium of energy. It is now considered
also as a medium of information with the widespread infiltration of computers and
communication networ%s to modern civili2ation. 1owever& the electric utility networ% is
still used and mostly considered as a medium of energy transmission& although there is an
attempt to use the power line for communication F"avlidou et al.& 344=G Lin et al.& 3443G
-hen and -hiueh& 3443I.
In this thesis& we are interested in e.tracting information from the power flows. !s a
medium of energy& electricity interacts with the loads& and during the process& it
inadvertently carries information about their nature. :e view the electricity as a medium
of energy plus embedded information about the characteristics of the loads. The concept
of the (LI) based on the NI"M is to e.tract high uality and uantity information
embedded in the electricity with a minimum physical installation and intrusion.
The e.amples and discussions in the previous sections suggest that there basically e.ist
three types of information that can be offered by the NI"M7 load disaggregation& fault
diagnosis and system analysis. Load disaggregation is essentially the study of how to
divide the total power observed into the powers consumed by individual electric loads. It
reuires the identification of load events to %eep trac% of the @NE@++ status of each load&
so that the total power can be allocated to each load accordingly at a given instant or
interval. +ault diagnosis is to detect the presence of a fault& identify its nature and
possibly probe its cause. It can be a simple summary of an observation or a more
complete description possibly with a remedy. )ystem analysis is to use the power signal
to identify the models of the electric loads and estimate their parameters. The result of the
system analysis can be lin%ed to a proper electrical& mechanical or thermal model of the
/?
system& for energy usage optimi2ation studies& supervisions of electric load control loops&
power management& etc. It will also be useful in designing load disaggregation and fault
diagnosis algorithms& by providing system models and parameters.
The common element of these three types of information is the state of the system. The
e.act %nowledge of the status of the electric loads in a system ma%es the tas% of load
disaggregation almost trivial (when supported by the %nowledge of their nominal
behavior). !ny event or gradual degradation that can not be supported or e.plained by the
legitimate combination of load states can be considered as a fault. !lso& detecting and
identifying a fault is basically similar to load event detection and identification& and the
latter algorithm can be used for the former application with the support of a fault
database. In addition& any energy usage optimi2ation or control loop supervision
implicitly assumes the in situ %nowledge of the states of the loads being monitored.
This thesis concentrates on state identification& and load disaggregation as its immediate
application. The fault diagnosis is not specifically pursued here& but many models and
algorithms developed for the state identification and load disaggregation will be useful
for diagnosis with minor modification. )ystem analysis is performed only at the level to
provide necessary models and parameters for the load identification routines.
The main ob5ective of this thesis is to develop fundamental models and signal processing
and information e.traction algorithms to benefit many possible future applications based
on the NI"M& rather than to develop a complete application pac%age for a specific sector.
It is the author6s belief that commercial-grade pac%age developments should be
performed by business entities. This thesis is intended as a cornerstone in developing
numerous future applications& such as a ubiuitous electric meter in a deregulated and
volatile utility mar%et& a diagnostic tool for end users or autonomous power systems& a
power uality monitor for various entities& etc.
-hapter 3 e.plains the concept of current harmonic powers& their estimation method and
implementation issues. It also illustrates their advantage in the efficient conservation of
the information present in a raw current waveform. The following three chapters are
dedicated to developing algorithms to identify the status of constant loads. The power
signal generated by an active constant load is modeled as a Daussian random process& and
this model is the %ey in developing various load status identification algorithms.
-hapter = deals with the simplest case when load events are isolated. ! new classifier is
introduced which can ta%e into account both step change observation and transient
pattern. The former is analy2ed by a ma.imum-li%elihood (ML) classifier and the latter is
categori2ed by a least suares (L)) error criterion. -hapter ? handles the case when
multiple load events happen at a same time. It generali2es the approach in the previous
chapter by searching a possible event space for event candidates and by developing an L)
error criterion that can account for multiple patterns at different time indices. -hapter B is
an endeavor to identify load states without relying on the occurrence of load events. It
develops a state model derived from the combination of individual Daussian random
process models& and performs a search to e.tract state candidates based on the steady-
state power observations. Then& an ML classifier selects the best state candidate. The state
estimation without load events is useful for the initial state estimation when the program
is first started and for resetting the event classifier when it fails.

-hapter 0 develops an algorithm to %eep trac% of the power consumed by a set of 8)#s&
using the correlations between the fundamental powers and higher harmonic powers.
8)#s are the most visible variable loads in commercial or industrial settings& but no
method has been available to estimate their power consumption from the non-intrusive
power signals& unless an NI"M is solely dedicated to a 8)#. -hapter A e.plains how to
estimate or eliminate the harmonic signals present in power waveforms. It is a sister
chapter to the previous one& because the 8)#s tend to generate harmonic signals. It
introduces three different estimation methods based on7 periodogram& autocorrelation and
fractional pitch estimation. @ne elimination method is presented& based on an
autoregressive signal model and white filtering from its parameters. -hapter C shows test
results& comes up with conclusion and discusses future wor%.
/0
Chapter 2
Current 6armonic Po?ersA *athematical Theory and Practices
This chapter discusses the issue of how to define powers when a current waveform has
spectral components other than the fundamental freuency of its corresponding voltage
waveform. #efinitions of current harmonic powers are proposed& which characteri2e the
harmonic components present in a current waveform with respect to a reference
fundamental voltage waveform. The method to obtain them from the current waveform
samples via the discrete +ourier transform (#+T) is introduced and its implementation
issues are discussed. There has been no consensus on the definition of powers when
voltage or current waveforms contain harmonic elements. The proposed current harmonic
power is not a complete solution to this problem& rather it is a convenient tool chosen to
characteri2e and analy2e the electric loads in the most effective and efficient manner. This
chapter provides a basis for the discussions in the following chapters& because it is the
current harmonic power we compute from the voltage and current samples.
2.1 ,ntroduction
)ince the adoption of !- electric power system over #- system& mainly because of its
low transmission cost& harmonics have always been present in power systems. @ne of the
main reasons is the versatility of the #- motor& which reuires a type of rectifier to
convert !- to #- FKacamini& /<<?I. @nly in recent years& have !- motors achieved a
comparable level of controllability& ironically at the cost of increased harmonics.
1armonics refer to spectral components present in a voltage or current waveform& whose
freuencies are integer multiples of the fundamental freuency (i.e. 04 12 in the ;)) of
the voltage waveform. 1armonic currents are created by non-linear loads& such as
variable speed drives (8)#s)& electronic ballasts for fluorescent lighting& switching power
supplies& rectifiers& etc. These harmonic currents are in5ected to the utility grid and
interact with the impedance of the system to create harmonic voltages. 1armonics can
cause such damage to utilities and end users as overheating of transformers and motors&
inadvertent tripping of relays& incorrect reading of watt-hour meters& etc. F$ennedy& 3444G
#riesen et al.& /<<CI. #istribution companies try to minimi2e the voltage harmonics
present in grids and set certain limits on the harmonic currents that can be in5ected to the
grids by their customers FI((( )tandard B/<-/<<3I.
In a deregulated electricity mar%et& power uality becomes more important and the
harmonics represent a significant portion of the overall power uality problem
3
. !
historical account of the harmonics in power engineering can be found in (manuel
F3444I. -urrent status and future directions of the power system harmonic analysis are
reviewed in Nu F344=I.
3
@r& a product& if the power uality is considered as a service to be delivered to customers.
#efinitions of the fundamental powers& such as real& reactive and apparent powers& are
based on the assumption of a single-freuency sinusoidal voltage waveform and linear
loads& which in turn create a monochromatic current waveform with the same freuency
but with a possible phase shift. 1ow to define powers when harmonics are present has
been debated over the years& but no consensus has been reached yet. The practical
definitions of powers under the presence of nonsinusoidal waveforms and unbalanced
loads are recommended by the I((( :or%ing Droup on Nonsinusoidal )ituations F/<<0I.
1owever& the implication and usefulness of the various power terms proposed in the
definitions are rather dubious. The physical meanings of those powers are partially
e.plained in (manuel F/<<4I.
:e adopt the term harmonic powers to represent various types of possible powers in
harmonic situations. -ertainly& the harmonic powers need to be defined and obtained
from the harmonic analysis of both the voltage and the current waveform& with a certain
amount of cross-correlation between the two waveforms. 1owever& this approach can
lead to an e.plosion of the number of the harmonic power terms. +or e.ample& if there are
N harmonic terms each in the voltage or the current waveform& N
3
harmonic power terms
can possibly be defined. The meaning of those terms is rather unclear& and the practicality
and rational of computing those unwieldy components are uestionable.
In this chapter& we concentrate on the characteri2ation of the harmonic content of a
current waveform& by defining a power-li%e uantity at each harmonic freuency. :e
may name it a current harmonic power& implying that only the current waveform is used
in computation. This approach can be 5ustified by the fact that the voltage distortion (ratio
of the sum of harmonic voltage amplitudes to a fundamental voltage amplitude) is usually
less than B O& whereas the current distortion (similarly defined) can be as high as /B4 O
FI((( :or%ing Droup& /<<0I. Thus& end users will find that the current harmonic power
is a practical and effective solution in characteri2ing and monitoring their electric loads to
combat the power uality problem.
>ecause utilities are more concerned about the uality of the voltage waveform they
deliver to their customers& they will find these definitions unsatisfactory. 1owever& the
same techniue can be e.tended to characteri2e the harmonic content of a voltage
waveform. The distribution companies can fully characteri2e the uality of the electricity
they mar%et& by the estimation of the current harmonic powers& augmented by an
additional harmonic analysis of the voltage waveform.

This chapter starts with the review of the concept of the fundamental powers to obtain a
useful insight in deriving the current harmonic powers. Then& the current harmonic
powers are formally defined and the relationship with the recommended definitions of the
harmonic powers is discussed. The method to compute the current harmonic powers from
the samples of a current waveform via the #+T is also proposed and its implementation
issues are discussed. The chapter also shows a few e.amples of the current harmonic
power estimations from raw waveforms. Their usefulness in characteri2ing and
monitoring electric loads is also illustrated.
/C
2.2 (undamental po?ers
:hen an electric load of impedance ) sin (cos u u j Z jX R + = + = B is e.cited by a
sinusoidal voltage waveform with a period T (e.g. /E04 sec)& it draws a sinusoidal current
with the same period and with a phase shift determined by its impedance. ;sing phasors&
&
/
;
B
9; , = = (3./)
where 9 is the admittance of the load. Thus& when the voltage waveform is given by
t V Ve t v
T
t j
T t
t
3
cos P ,eQ ) (
3
= = & the current lags the voltage by u& i.e.

). cos( P ,eQ ) (
3
3
u
t
u
t
= =

t I e Ie t i
T
t j
j
T
(3.3)
Note that V and I are pea% amplitudes. They are related to the root-mean suared (,M))
values via&
. 3 & 3
rms rms
I I V V = = (3.=)
Two power terms are associated with the load. The first term R I
rms
3
= is called a real
or active power& and is the power consumed by the resistive component of the load. The
second term X I !
rms
3
= is called a reactive power& and is the power periodically stored
and released by the reactive component of the load. The two power terms can be
combined into a single comple. uantity& called an apparent power7
). sin (cos u u j I V j!
rms rms
+ = + = A (3.?)
:e observe that the angle of the apparent power is the same as the impedance angle of
the load. +igure 3-/ summari2es these relationships.

impedance triangle
power triangle
u
Z
R
X
Z I
I V "
rms
rms rms
3
=
=
u cos
3 3
Z I R I
rms rms
= =
u sin
3
3
Z I
X I !
rms
rms
=
=
u
+igure 3-/ Impedance and power triangles.
The real power can be obtained by ta%ing the product of the voltage and the current
waveform and computing its one cycle average7

( )
. cos cos
3
3
3 sin
sin
/
3
3 cos /
cos
/
sin sin cos cos cos
/
) ( ) (
/
3
3 3
3 3 3
R I I V
VI
d VI
T
d VI
T
d I V
T
d i v
T

rms rms rms

T t
t
T
T t
t
T
T t
t
T T T
T t
t
= = =
+
+
=
+ = =
} }
} }
+ +
+ +
u u
t
t
u t
t
u
t t u t u t t t t
t t
t t t
(3.B)
The real power is the time average of the product of the voltage waveform and the in-
phase component of the current waveform. This implies that the reactive power will be
the average product of the voltage waveform and the <4 out-of-phase& or uadrature&
component of the current waveform. Mathematically& it is obtained by shifting the voltage
waveform by a uarter cycle before integration7

( )
. sin sin
3
3
3 cos /
sin
/
3
3 sin
cos
/
sin sin cos cos sin
/
) ( ) (
/
3
3 3
3 3 3
?
X I I V
VI
d VI
T
d VI
T
d I V
T
d i v
T
!
rms rms rms
T t
t
T
T t
t
T
T t
t
T T T
T t
t
T
= = =

+ =
+ = =
} }
} }
+ +
+ +
u u
t
t
u t
t
u
t t u t u t t t t
t t
t t t
(3.0)
1ere& the shifted voltage waveform does not e.ist in reality. ,ather& it serves as a
fictitious reference waveform for mathematical convenience. The fundamental powers
can also be interpreted as the correlations between the current waveform and the shifted
voltage waveforms.
2.3 Current harmonic po?ers
The concept of the real& reactive and apparent powers& introduced in the previous section&
wor%s well when a load is linear& i.e. it produces a sinusoidal current with a single
freuency when e.cited by a sinusoidal voltage with the same freuency. 1owever& when
the load is not linear& the current waveform is not monochromatic even when it is e.cited
by a perfectly clean sinusoid. ,ather& the current contains the harmonic components of
the fundamental freuency of the voltage waveform. +igure 3-3 illustrates the concept of
linear and nonlinear loads.
34

linear
load
nonlinear
load
) & ( t v e
) & ( t i e
T
t i
I = 3 F
) & (
. e e e = C
C

+igure 3-3 Linear and nonlinear loads.
1ow to define powers in the presence of harmonic currents and how to associate them
with meaningful e.planations are still unresolved challenging issues. 1owever& the
definitions of the fundamental powers in the previous section suggest that we may use a
properly shifted harmonic voltage waveform as a reference to compute a power at a
harmonic freuency. :e define the #th current harmonic powers by the following
formulae7

. & 3 & / ) ( ) sin(
/
) ( ) (
/
) ( ) cos(
/
) ( ) (
/
3
?
3
= = =
= =
} }
} }
+ +
+ +
# d i # V
T
d i # v
T
!
d i # V
T
d i # v
T

T t
t
T
T t
t
T
#
T t
t
T
T t
t
#
t t t t t t
t t t t t t
t
t
(3.A)
Note that / R and !/ R ! and hence that the definition is consistent at the fundamental
freuency. Thus& the #th current harmonic powers are obtained by computing the
correlations between the current and the properly shifted #th reference waveforms. :e
may call # as the #th real power and !# as the #th reactive power. 1owever& they do not
have clear physical meanings e.cept when # R /. ,ather& they uantify the magnitude of
the #th harmonic freuency component present in the current waveform. The phase
information is hidden in the relative brea%up of # and !#.
If we want to uantify the #th harmonic component of the current waveform without the
phase information& we may use the #th apparent power "# defined as7
.
3 3
# # #
! " + = (3.C)
! closer e.amination of (3.A) reveals that we only need the fundamental voltage
waveform (period& magnitude and phase) to compute the current harmonic powers. Thus&
the current harmonic powers can still be defined and computed as long as the
fundamental freuency waveform can be e.tracted& even from a (harmonically) distorted
voltage waveform.
>ecause the amount of voltage harmonic distortion is typically very small at end users6
sites and because they are more concerned about the current harmonic distortion created
by their electric loads& the current harmonic power is a very effective and possibly cost
efficient tool to characteri2e the electric loads and monitor the power uality at their
premises. (manuel et al. F/<<=I conducted a survey of harmonic voltages and currents at
several representative customer sites.
;tilities will find the definition of the current harmonic powers less satisfactory& because
they are reuired to supply a clean sinusoidal voltage waveform and try to maintain the
voltage harmonic distortion below a certain limit. Thus& they need a separate tool to
characteri2e the harmonic component of a voltage waveform& in order to monitor
constantly the harmonic pollution and ta%e appropriate actions when necessary. To
characteri2e the harmonic content of a voltage waveform& we may replace i(t) with v(t) in
the definition of the #th current harmonic powers and call the resulting uantities the #th
voltage harmonic intensities. ! survey of the harmonic voltages and currents at several
distribution substations can be found at (manuel et al. F/<</I.
It will be a good idea to review the definitions of powers under harmonic situations& as
recommended by the I((( :or%ing Droup F/<<0I& and e.plore the relationships between
the accepted definitions of powers and the current harmonic powers proposed in this
chapter. The recommended definitions start by representing the voltage and current
waveforms as the sums of harmonic sinusoids.

. ) sin( 3 ) (
) sin( 3 ) (
/
4
/
4

=
+ + =
+ + =
h
h h
h
h h
t h I I t i
t h V V t v
| e
o e
(3.<)
Note that Vh and Ih are ,M) uantities. The ,M) voltage V and current I are defined as7

. &
&
/ & 4
3 3 3 3
/
4
3 3
/ & 4
3 3 3 3
/
4
3 3

= =

= =

=
= + = =
= + = =
h h
h $ $
h
h
h h
h $ $
h
h
I I I I I I
V V V V V V
(3./4)
The apparent power % is defined as7
. ) (
3 3
/
3 3
N
% % VI % + = = (3.//)

/ / /
I V % = is called the fundamental apparent power. The suare of the non-fundamental
apparent power %N has three components7
. ) ( ) ( ) (
3 3
/
3
/
3
$ $ $ $ N
I V I V I V % + + = (3./3)
33
The first term is called the current distortion power and is usually a dominant term. The
second term is called the voltage distortion power& while the third term is called the
harmonic apparent power.
>ecause a current harmonic power is obtained as a correlation between a reference
waveform obtained from the fundamental voltage waveform and the current waveform&
we recogni2e that the current harmonic power roughly belongs to the realm of the current
distortion power. The difference is that the former represents the spectral component of a
current waveform at each harmonic freuency& while the latter uantifies the aggregate
amount of the current harmonics in an ,M) sense.
2.. *ethod to compute current harmonic po?ers
@nce we define the current harmonic powers& the ne.t issue is how to compute them from
the measurements of raw voltage and current waveforms& possibly in real time. Their
definitions suggest that& once we e.tract a cycle of the fundamental voltage waveform
from the voltage measurement& we can create reference waveforms and thus the current
harmonic powers can be obtained by performing integrations on the products of the
references and the current.
Leeb developed an analog current harmonic power estimator& composed of a phase-
loc%ed-loop to e.tract the fundamental voltage waveform& waveform synthesi2ers to
create reference waveforms& multipliers and low-pass filters to simulate integration FLeeb&
/<<=G Leeb et al.& /<<BI. This process of reference waveform generation& multiplication
and integration can also be performed by software using sampled waveforms F)haw et al.&
/<<CI. 1owever& one can easily imagine that this is a computationally intensive process.
In this section& we develop a more efficient current harmonic power estimation routine
based on sampled waveforms& using the relationship between the continuous-time +ourier
series (-T+)) and the discrete +ourier transform (#+T) via the sampling theory. ;sing
the #+T to find the -T+) harmonic power terms is discussed in H. Luo et al. F3443I and
(l-1abrou% and #arwish F344/I. @ther techniues are also reported in the literature& such
as the one based on neural networ% FMoreno-(guila2 et al.& 3444G @sows%i& /<<3I and the
one based on singular value decomposition F@sows%i& /<<?I. ,ech%a et al. F3443I
conducted a comparative study of the harmonic detection algorithms.
-onsider a voltage signal
) ( ) ( ) (
/
t v t v t v
$
+ =
and a current signal i(t) observed for 4 s t S
T& where ) cos( ) (
3
/
t V t v
T
t
= is the fundamental voltage waveform and ) (t v
$
is the
additive harmonic voltage distortion. The definitions of the current harmonic powers for
the given interval are restated7
. & 3 & / ) ( ) sin(
/
& ) ( ) cos(
/
4
3
4
3
= = =
} }
# d i # V
T
! d i # V
T

T
T #
T
T #
t t t t t t
t t
(3./=)

Note that the powers are defined for a given interval (period)& and in general will have
different values at different time instances.
Ne.t& we consider a periodic e.tension of the current waveform defined as7
. 4 ) ( ) (
T
& & 3 & / & 4 ) (
T
) (
T
T t t i t i r rT t i t i s s = = + = . (3./?)
Then& ) (
T
t i can be represented in terms of the -T+)7
& ) sin( ) cos( ) (
T
/
3
/
3
4

=

=
+ + =
#
T #
#
T #
t # b t # a a t i
t t
(3./B)
where the +ourier series coefficients a# and b# are given by the following formulae7
. / ) sin( ) (
3
& ) cos( ) (
3
4
3
4
3
> = =
} }
# d # i
T
b d # i
T
a
T
T #
T
T #
t t t t t t
t t
(3./0)
:e note that the current harmonic powers are simply the scaled versions of the +ourier
series coefficients& i.e.
.
3
&
3
# # # #
b
V
! a
V
= = (3./A)
Thus& the current harmonic powers of a given interval can be obtained by finding the
+ourier series coefficients of the periodic e.tension of the current waveform with an
appropriate scaling determined by the amplitude of the voltage waveform.
In terms of the continuous-time comple. +ourier series F@ppenheim et al.& /<<CI7

. ) (
T /
U
&
U
) (
T
4
3 3
}

=
= =
T
t j#
#
#
t j#
#
dt e t i
T
i e i t i
T T
t t
(3./C)
>ecause the current waveform is real& a# and b# are also real& and it can be shown that7
. / )
U
3 Im( )&
U
3 ,e( > = = # i b i a
# # # #
(3./<)
There also e.ists a continuous-time +ourier transform (-T+T) of the current signal7
. ) ( ) (
}

O
= O dt e t i j I
t j
(3.34)
If the transform is evaluated at the harmonic freuencies& i.e.
T
#
t 3
= O & considering the
fact that i(t) is time-limited&
3?
.
U
) ( ) (
4
3
3
#
T
t j#
T
i T dt e t i j# I
T
= =
}

t
t
@r& . ) (
/
U
3
T #
j# I
T
i
t
= (3.3/)
Thus& the +ourier series coefficients are simply the scaled versions of the +ourier
transform evaluated at the harmonic freuencies.
. / ) (
3
Im )
U
3 Im( & ) (
3
,e )
U
3 ,e(
3 3
> |
.
|

\
|
= = |
.
|

\
|
= = # j# I
T
i b j# I
T
i a
T # # T # #
t t
(3.33)
The current harmonic powers can also be represented in terms of the +ourier transform7
( ) ( ). ) ( Im
3
& ) ( ,e
3
3 3
T # # T # #
j# I
T
V
b
V
! j# I
T
V
a
V

t t
= = = = (3.3=)
If the +ourier transform of a continuous current signal can be computed& its current
harmonic powers can be obtained simply by evaluating the transform at the harmonic
freuencies with the scale factors. 1owever& due to the difficulty in performing the -T+T&
the signal is typically sampled and the #+T is performed instead. Thus& we need to find a
relationship between the -T+T and the #+T& to obtain the current harmonic powers from
the #+T of the current signal samples.
-onsider a continuous signal &c(t) and its sampled version &FnI with a sampling period Ts
(sec)& i.e. &FnI R &c(nTs). The discrete-time +ourier transform (#T+T) of &FnI is defined as7

. I F ) (

=
n
n j
e n & j X
e
e
(3.3?)
Note that X(je) is periodic with 3t& i.e. X(je) R X(j(e93t)). It is well %nown from the
sampling theory that X(je) is a scaled and aliased version of the -T+T of &c(t)
F@ppenheim et al.& /<<<I7

.
3 /
) (
s
# s
c
s
T
T
#
j X
T
j X O =
|
|
.
|

\
|
=

=
e
t e
e
(3.3B)
It should be noted that e is in FradI& whereas O is in FradEsecI. If Xc(jO) is band-limited to
s
O
3
/
& where
s
T s
t 3
= O
(radEsec)& no aliasing occurs and X(je) is simply&

.
/
) (
|
|
.
|

\
|
=
s
c
s
T
j X
T
j X
e
e
(3.30)

If N points are sampled for a given interval T& with a sampling period Ts& T R NTs and &FnI
is an N-point seuence. The discrete +ourier series (#+)) is obtained from the N-periodic
version of &FnI and the #+T XF#I is defined as a freuency-limited version of the #+). !s
a result&

. otherwise 4
/ 4 I F
I F
/
4
3
s s
=

N # e n &
# X
N
n
#n j
N
t
(3.3A)
XF#I can also be interpreted as the samples of the #T+T. -ombining the relationship
between the #T+T and the -T+T&

.
3 / 3 / /
/ 4 ) ( I F
3
3
|
.
|

\
|
=
|
|
.
|

\
|
=
|
|
.
|

\
|
=
s s =
=
=
T
#
j X
T NT
#
j X
T T
j X
T
N # j X # X
c
s s
c
s
#
s
c
s
#
N
N
t t e
e
t
t
e
e
(3.3C)
Thus& the -T+T evaluated at a harmonic freuency is simply the scaled version of the
#+T7
. I F
3
# X T
T
#
j X
s c
= |
.
|

\
| t
(3.3<)
Now& the current harmonic powers can be represented in terms of the #+T IF#I of the
sampled current waveform7

( ) ( ) ( )
( ) ( ) ( ). I F Im I F Im ) ( Im
3
I F ,e I F ,e ) ( ,e
3
3
3
# I
N
V
# I T
T
V
j# I
T
V
b
V
!
# I
N
V
# I T
T
V
j# I
T
V
a
V

s T # #
s T # #
= = = =
= = = =
t
t
(3.=4)
:e note the similarity in the definitions of the current harmonic powers in terms of the
-T+T and the #+T& i.e. the real or imaginary part of the transform is multiplied by the
pea% voltage and divided by a corresponding interval.
!n interesting case occurs when the fundamental voltage waveform is not necessarily
loc%ed to
) cos(
3
t V
T
t
for 4 s t S T& i.e. when there is a time shift in v
/
(t). -onsider a
sampled waveform v/FnI for 4 s n s N J / and assume that shifting v/FnI by n4 results in a
loc%ed cosine waveform7
. / cos I F U I F
4 4
3
/ 4 /
+ s s = = N n n n n V n v n n v
N
t
(3.=/)
! time-shift corresponds to a phase-shift in the #+T7
30



. / / I F
4 I F
I F
U
3
) (
/
3 /
/
4
3
4
3
s s +
s s
=

N # e # V
# e # V
# V
N
n N # j
N
#n j
N
N
t
t
(3.=3)
If we evaluate the transforms at # R /&

. I / F
U
I / F
4
3
/ /
n j
N
e V V
t
=
(3.==)
>ecause I / F
U
/
V is a positive constant ( I F
U
/
# V is the #+T of a pure cosine)& the argument of
V
/
F/I corresponds to the additional phase angle introduced by the phase-shift of the
cosine waveform.
>ecause the fundamental voltage waveform ) ( cos I F
4
3
/
n n V n v
N
+ =
t
& we want to ta%e
into account the phase shift in computing the #+T of the current waveform7
& I F I F I F I F
/
4
/
4
) (
3
4
3
4
3
# I e e n i e e n i # I
# j
N
n
#n j #n j
N
n
n n # j
N N N
|
t t t
= = = '


=

=
+
(3.=?)
where IF#I is the regular #+T of the current waveform and I) / F arg(
V
/ 4
3
V n
N
= =
t
| (V
denotes comple. con5ugate). 1owever& I / F
/
V is eual to VF/I& due to the independence of
the harmonic components& when the voltage signal is band-limited to
s
T
t
(radEsec). Thus&
I F# I '
is a phase-shifted version of IF#I& with the phase shift obtained from the #+T of the
voltage waveform& and must be used in place of IF#I in computing the current harmonic
powers.
The above method of using the phase angle of the first harmonic component of the
voltage #+T is convenient because it does not reuire the seuence of samples to start at
the 2eroth angle position of the voltage cosine waveform (Vcos4). )ampling can start at
any point as long as it samples N eually spaced points per cycle and one of the samples
is ta%en at the 2eroth angle. !ny time shift will be ta%en care of by the phase of VF/I.
In the implementation point of view& it is an inefficient process because we have to
perform the N-point #+T on the voltage samples to obtain a single scalar output.
1owever& it is robust because VF/I is immune to other harmonic components that might
be present in the voltage waveform. If this approach is not ta%en& a certain type of phase-
loc%ed-loop algorithm needs to be applied to the voltage waveform samples. @ne can
easily imagine that such an algorithm will perform poorly when there e.ist other
harmonic components and noise. :here resources are limited& a single-bin #+T output
routine& such as the Doert2el algorithm& can be used to compute VF/I only& instead of
computing the whole N-point #+T FMacobsen and LyonsG @ppenheim et al.& /<<<I.
2.D ,mplementation issues
The previous section shows that the current harmonic powers for a given interval can be
computed efficiently by ta%ing N samples of a current waveform and ta%ing their N-point
#+T. 1owever& a few conditions& imposed in deriving the relationship between the -T+)
terms and the #+T bins& need to be observed by any current harmonic power estimator
based on the #+T.
The first condition is that both the voltage and current waveforms need to be band-limited
to the Nyuist freuency of
s
T 3
/
12. This can be achieved easily by low-pass filtering the
raw waveforms either before or after sampling. 1owever& because the filtering can not be
ideal& certain amounts of attenuation and aliasing are inevitable. The distortion introduced
by the low-pass filtering is usually negligible.
The second condition is that N eually spaced samples need to be ta%en for a period. If
the phase rotation method based on VF/I is used& sampling can start at any point as long
as the 2eroth angle position of the fundamental cosine waveform is included in the
sampled seuence. If not& sampling should start at the 2eroth angle position. !t first& this
seems to be a trivial reuirement& but in reality& it is the most difficult condition to meet
in implementing a real-time high-precision current harmonic power estimator.
The difficulty stems from the fact that the generator of a utility grid and the cloc% of a
data acuisition (!E#) card can not be synchroni2ed. !lthough the nominal period of the
!- voltage is /E04 sec in the ;nited )tates& there usually are slight variations over time.
In addition& it is common to observe interruptions or transients in the voltage waveform&
the discontinuations of the sinusoid voltage waveform for short durations. Thus& even if
we start sampling at the correct 2eroth angle position& as time goes by& the sampler may
obtain an inconsistent number of samples per cycle (e.g. N J / or N 9 / samples per
cycle) and the first sample of a cycle may not necessarily be at the correct 2eroth angle
position. Thus& the sampler somehow has to chec% the shape of the fundamental voltage
waveform continually to ad5ust its sampling period and sampling start point accordingly
to satisfy the condition.
This is an e.ample of a rather common transmitter-receiver synchroni2ation problem& and
is related to the issues of non-integer delay& sampling rate conversion and interpolation
FLaa%so& et al.& /<<0I. In digital communication& a receiver must be synchroni2ed to
incoming symbols. It can be accomplished either by an analog feedbac% or feedforward
control of the phase and interval of a sampling cloc%& or by an autonomous sampler
followed by a fractional-delay filter to estimate symbols at correct sampling intervals
FDardner& /<<=G (rup et al.& /<<=G >ingham& /<CCI.

)haw F3444I approached this problem by first sampling the raw voltage and current
waveforms at a higher sampling rate (e.g. /0 %12) and then filtering and resampling them
at a lower rate to obtain N (e.g. /3C) samples consistently per cycle. The period and phase
of the voltage waveform can be estimated by finding ad5acent 2ero-crossings of the
3C
filtered voltage samples. New samples are obtained by interpolating old samples at
estimated sampling locations.
1owever& one can easily imagine that this filtering and resampling process will not
perfectly reconstruct the original bandlimited waveforms and thus certain amounts of
errors (both in magnitude and location) will be introduced to the new samples. The
magnitude errors will eventually introduce certain biases or random fluctuations to the
current harmonic power estimations. :e may treat them as noise in5ected by a non-ideal
power estimation process.
The effect of the location errors is rather subtle. >ecause the process resamples the
waveform starting from an appro.imate 2eroth angle position& the 2ero-crossing
estimation determines the accuracy of the new sample locations. The uncertainty of the
e.act 2ero-crossing location will result in a shift or bias in the locations of the resampled
data points. The actual accuracy depends on the type of 2ero-crossing estimation
algorithm used and the nature of signal. 1owever& it is reasonable to assume that the
position uncertainty is on the order of one sample interval (after resampling). If we denote
the un%nown bias by
4
n and assume that n V n v n n v
N
t 3
4
cos I F U I F = = & the bias introduces
a phase-shift at each #+T bin& following (3.=3). Note that ) / ( T
4
' n and that
4
n is a
fraction in general.
Thus& the bias in the sampling position results in the bias in the phase of a current
harmonic power estimation. The phase bias increases with increasing harmonic number
and becomes a ma.imum when the harmonic number reaches
3
N
. The #th current
harmonic power will have a phase bias or uncertainty on the order of #
N
t 3
(rad). :ith N R
/3C& this corresponds to 3.C/ (deg) for the fundamental powers and /<.0< (deg) for the
seventh harmonic powers. The phase bias can become uite significant even at modest
harmonic numbers. !lso& if the position bias is not consistent& the resulting phase bias of
a current harmonic power can move from one e.treme to another& creating large
fluctuations in its magnitude. The phase bias can be reduced by increasing the sampling
rate (larger N) and by using a more accurate 2ero-crossing estimation algorithm.
,ecursive estimation techniues can be employed to better estimate voltage waveform
parameters and hence to reduce the sampling position bias. The $alman filter is one of
the most well %nown tools to estimate the magnitude& phase and period of the voltage
waveform in a power system F>eides and 1eydt& /<</G #ash et al.& 3444I. 1owever&
these techniues may not wor% properly when there is a significant amount of harmonic
distortion in the voltage waveform. ! more robust and efficient fundamental voltage
waveform estimation techniue needs to be investigated in future.
:here the phase bias is a concern& one may use the apparent powers to characteri2e the
harmonic content of a current waveform& instead of using the separate real and reactive
powers. The #th apparent power is the magnitude of the #th comple. power& and thus is
free from the precarious time-shift of the reference voltage waveform. 1owever& the
dimension of the current harmonic powers is reduced by half& which can reduce their
resolving capability in characteri2ing and analy2ing electric loads.
! high degree of parallelism is also reuired between the voltage and the current samples&
such that two waveforms are sampled at the same time instances. The !E# card should be
able to sample the voltage and current simultaneously using parallel channels. The
filtering and resampling& if used& need to be done with care to ensure a common sampling
location for both the voltage and the current sample at a given time.
In this manner& N samples are ta%en from a cycle of current and N current harmonic
power terms are computed using the #+T. It can be conceived as an N-point windowing
operation on a stream of discrete current waveform samples. The window is used to
e.tract a cycle of the signal& and the window is shifted to the ne.t cycle. If there is no
overlap between the windows& the current harmonic powers are computed at every cycle
and their update rate is /ET 12 (e.g. 04 12 in the ;)). 1owever& if we shift the window by
less than a cycle& creating an overlap between successive windows& the powers are
generated faster than /ET 12. +or e.ample& if we shift the window by NE3 points& the
powers are computed at every half cycle and their rate is 3ET 12. )hifting the window by
less than N points can be useful& as it provides a means to interpret between the otherwise
/ET 12 power samples. It can be conceived as an interpolation.
)o far& we have assumed a single-phase power system. >ut the ma5ority of industrial and
commercial customers use three-phase power systems. #efining the current harmonic
powers in the three-phase situation is a rather complicated problem& in part because there
is no established standard in defining three-phase powers FI((( :or%ing DroupJ survey&
/<<0I. This is an unresolved problem and will be investigated in future. ;ntil then& we
will monitor only a single phase out of a three-phase system and compute its current
harmonic powers. :hen total three-phase (fundamental) powers are necessary& we will
simply multiply the single-phase powers by three& assuming balanced loads.
2.E E@amples
Measurements were ta%en at the main 18!- room of a commercial building at /04
)ansome )t. )an +rancisco& -!. The 18!- loads are served by a ?C4 8 three-phase
power system. The voltage was measured across the ! phase and the neutral& and the
current samples were ta%en at the ! phase.
+igure 3-= shows the raw voltage and current waveforms over three cycles. !t the time of
measurement& no 18!- load was on. 1owever& the 18!- panel also serves a number of
small lighting and plug loads. :e may call them collectively the bac#ground load. :e
note that the voltage waveform is a fairly clean sinusoid& whereas the current waveform is
slightly distorted. Thus& we e.pect the presence of higher harmonics.
+igure 3-? shows the current harmonic powers (of single phase) computed from the
waveforms in +igure 3-=. Three cycle averages up to the si.teenth harmonic are reported
in the figure& although ) / (
3

N
harmonics are available when N samples are ta%en per
=4
cycle. :e observe that the fundamental powers are dominant terms. 1owever& there also
e.ist substantial higher harmonic powers& especially the fifth and seventh components.

0 50 100 150 200 250 300 350
-400
-200
0
200
400
raw voltage waveform, 128 pts per cycle
(
V
)
0 50 100 150 200 250 300 350
-20
-10
0
10
20
raw current waveform, 128 pts per cycle
(
A
)
sample number, (60*128 Hz)
+igure 3-= 8oltage and current waveforms of bac%ground load.

0 2 4 6 8 10 12 14 16
-1
0
1
2
3
current harmonic powers
P
k

(
k
W
)
0 2 4 6 8 10 12 14 16
-1
0
1
2
3
Q
k

(
k
V
A
R
)
k, harmonic number
+igure 3-? -urrent harmonic powers of bac%ground load.
+igure 3-B shows the waveforms when the /44 hp (air) supply fan is turned on in addition
to the bac%ground load. The capacity (speed) of the fan is controlled by a 8)# and thus
the resulting current waveform is highly nonlinear. :e note that the voltage waveform is
relatively clean. +igure 3-0 shows the resulting current harmonic powers. +undamental
powers are still significant terms& but a few higher harmonic powers& especially the fifth
and seventh& are on the order of same magnitude.

0 50 100 150 200 250 300 350
-400
-200
0
200
400
raw voltage waveform, 128 pts per cycle
(
V
)
0 50 100 150 200 250 300 350
-200
-100
0
100
200
raw current waveform, 128 pts per cycle
(
A
)
sample number, (60*128 Hz)
+igure 3-B 8oltage and current waveforms of supply fan plus bac%ground load.

0 2 4 6 8 10 12 14 16
-10
0
10
20
current harmonic powers
P
k

(
k
W
)
0 2 4 6 8 10 12 14 16
-10
-5
0
5
10
Q
k

(
k
V
A
R
)
k, harmonic number
+igure 3-0 -urrent harmonic powers of supply fan plus bac%ground load.
!n interesting insight can be obtained by loo%ing at the current harmonic power
estimation from the viewpoint of signal compression. :e remember that the current
harmonic powers are nothing more than a spectral characteri2ation of the current
waveform based on the #+T. Thus& the signal energy
=
should be conserved before and
after the transform following "arseval6s relation F@ppenheim et al.& /<<<I. If we conserve
=
+or a deterministic N-point seuence &FnI& its signal energy ( is defined as&
. I F
/
I F
/
4
3
/
4
3


=

=
= =
N
#
N
n
# X
N
n & (
The latter euation is commonly referred to as "arseval6s relation for the #+T.
=3
a few significant odd harmonic powers& both real and reactive& and get rid of the rest& we
can achieve a significant amount of signal compression while conserving most of the
information originally present in the current waveform.
+or e.ample& from +igure 3-0& if we retain odd harmonic powers up to the seventh& we
achieve the signal compression ratio of <=.AB O. Ket& we retain <<.3< O of the original
signal energy. If we retain up to the thirteenth odd harmonic powers& the signal
compression ratio drops to C<.40 O. 1owever& the signal energy conservation ratio
increases to <<.<3 O. There is a tradeoff in choosing the number of harmonic powers to
%eep. ,etaining more harmonic powers is obviously better to transmit more information&
but it will increase the data handling and computation cost in subseuent analyses.
Thus& we can conceive the computation of the current harmonic powers from raw current
waveform samples as a type of coding process. The information present in the signal is
converted from one form to another& and the format is chosen so that the efficiency in
transmission and storage can be ma.imi2ed. The current harmonic powers are the codes
chosen to represent the spectral components of the current signal& and the coding is very
efficient in the sense that only a few symbols are necessary to carry the essential
information present in the signal. Thus& we may call the overall process of estimating the
current harmonic powers from the raw waveform samples as power coding& and the
estimator as power coder or pocoder& in so much as voice coders are called vocoders in
speech processing.

|P
k
|
k
,

h
a
r
m
o
n
i
c

n
u
m
b
e
r
20 40 60 80 100 120
5
10
15
|Q
k
|
k
,

h
a
r
m
o
n
i
c

n
u
m
b
e
r
n, window number (120 Hz)
20 40 60 80 100 120
5
10
15
+igure 3-A Magnitudes of real and reactive harmonic powers over successive windows.

+igure 3-A shows the magnitudes of # and !# in gray scale. Larger magnitudes are
indicated by dar%er pi.els. :indow si2e N is /3C& and the window is shifted by NE3.
Thus& the power coder generates harmonic powers at /34 12. Initially& both the supply fan
and the return fan (AB hp) are on. !t n R =0& the return fan is turned off. :e note that the
magnitudes of the ma5or odd harmonic powers are reduced at the return fan turn-off
event. :e also observe that !
/=
is mostly generated by the return fan& because it becomes
negligible after n R =0. This phenomenon is more clearly visible in +igure 3-C& where the
intensities of # and !# are plotted in a logarithmic scale.

+igure 3-C Intensities of real and reactive harmonic powers over successive windows.
This e.ample illustrates the fact that an electric load generates a set of characteristic
harmonic powers when it is on& and that its characteristic may be visible even when other
loads are on& which add up their own characteristic harmonic powers to our observations.
+igure 3-C hints that by monitoring the magnitude of !
/=
& we can ma%e a relatively sound
5udgment whether or not the return fan is on& because the supply fan does not generate a
significant amount of !
/=
.
The supply fan also turns off at n R C4. This event is indicated by large reductions in the
higher harmonic powers. !fter then& only the bac%ground load is on. The figures in this
section illustrate that electric loads generate characteristic harmonic powers and the
observation at a given time is their superposition. Thus& by analy2ing the composition of
the harmonic powers and monitoring their changes& we should be able to e.tract the
information about the operational status of the loads.

2.5 Chapter summary
The current harmonic powers were defined in this chapter& and a #+T based method to
compute them was also e.plained. Their usefulness in characteri2ing electric loads and
monitoring their status was also illustrated through the e.amples.
1ow to define powers when harmonics present is a much debated yet still unresolved
issue. The current harmonic power is not a panacea to this annoying ailment& rather it is a
=?
convenient and computationally-efficient (partial) remedy based on the fact that the
ma5ority of the distortion is present in the current waveform. !s the estimation of the
spectral components of a current waveform& the current harmonic power can be used to
uantify current harmonic distortions and to e.tract the information about the attributes
and status of the electric loads. The voltage harmonic intensity was also defined& which
can be used to estimate voltage harmonic distortions in a complimentary manner. Its
computation can be performed similarly using the #+T.
The current harmonic power computation method based on the #+T of the current
waveform samples is a very efficient tool compared to other approaches based on the
reference waveform generation and integration. ! few conditions were identified to
estimate the current harmonic powers accurately via the #+T. 1owever& obtaining eually
spaced samples starting at the 2eroth angle position of a reference waveform is the most
difficult problem due to the generator-sampler synchroni2ation problem. The filtering and
resampling approach was discussed and the phase bias of a current harmonic power due
to the position bias was also discussed. The apparent powers may be used where the
phase bias is a ma5or concern.
!n intriguing insight was obtained by treating the estimation of the current harmonic
powers from raw current waveform samples as a coding process. The current harmonic
powers are the codes we chose to carry the essential information present in a waveform.
The coding is very effective (high signal energy conservation) and efficient (high signal
compression) in the sense that only a few symbols are necessary to characteri2e the
original signal. +or e.ample& we are able to reconstruct the most of the original current
waveform with only a few odd harmonic power terms. !lso& the e.amples show that
analy2ing those symbols retained after the coding can yield valuable information about
the character and status of the monitored loads.

Chapter 3
Single )oad E#ent ,denti'ication using Steady!state )oad Classi'ication and
Transient Pattern -ecognition
:hen a non-intrusive power monitor is installed at the main electric service entry of a
site& the most basic and important information rendered by the monitor is the operational
status of the loads and the subseuent disaggregation of the total power observations into
individual power consumptions. The traditional method of load status identification
analy2es the seuence of steady-state power changes at a central location. 1owever& the
method performs poorly when there are similar loads and does not provide the capability
of real-time load disaggregation.
In this chapter& a new approach is proposed7 load identification using both steady)state
powers and transient patterns. The transient pattern of a load is mostly uniue and
repeatable and typically provides more discernibility than its steady-state power
consumption. ;tili2ing two complementary features of a load results in a more accurate
and robust load classification algorithm. ! decision is made locally based on a single
observation& which provides real-time load disaggregation capability.
3.1 ,ntroduction
This chapter presents two load disaggregation methods& steady)state load classification
and transient pattern recognition& and their synthesis. The method of steady-state load
classification was developed more than a decade ago and has been used widely including
a prototype residential application F1art& /<<3G #ren%er and $ader& /<<<I. 1owever& the
method does not perform well when two loads are close to each other in the steady-state
mapping space. The method also suffers when two load events happen sufficiently close
in time so that they can appear as a single event.
The problem of load pro.imity is related to the nature of loads and will not be solved
completely& although improvements can be made. The problem of temporal overlap of
events can be solved partially by increasing the time resolution of power waveforms.
1owever& increasing the time resolution introduces another problem of how to deal with
the transient responses of the loads. Traditionally& the transient responses are regarded as
nuisances in measuring the steady-state powers and eliminated as much as possible by an
adeuate filtering and downsampling F$arl et al.& /<<3G #. Luo et al.& 3443I.
@n the other hand& the study shows that the transient response of a load is relatively
unique and repeatable& which implies that the transient patterns can be used to identify
the loads FNorford and Leeb& /<<0I. This approach nullifies the uestion of how to
eliminate the transient responses and provides an additional resolving power when similar
loads e.ist.
=0
>y utili2ing two distinctive features of a load& steady-state power and transient pattern& in
a complementary manner& the tas% of load identification becomes more viable. !lso& due
to the improvement in the identification accuracy& the load classification decision can be
made locally after a single observation. -ompared to the conventional method&
characteri2ed by the local collection& central processing and off-line disaggregation& the
proposed method collects and processes data locally and has real-time load
disaggregation capability.
This chapter first develops the method of steady-state load classification. The overall
structure of the method is first presented& followed by the development of its individual
components. Then& its performance is evaluated& which naturally proves the necessity to
utili2e the transient patterns to compensate its limitation.
The second half of the chapter is devoted to the development of a transient pattern
recognition method. It first introduces the problem of pattern recognition in con5unction
with the electric load identification. It then develops the overall structure and the details
of the method. It also discusses the issue of how to coordinate the outputs from two
inherently different load classifiers. The overall load disaggregator with two parallel
classifiers is tested with the data under various conditions.
3.2 )oad classi'ication $ased on state space mapping
The method of steady-state load classification relies on the fact that constant loads
consume relatively fi.ed amounts of the real and reactive powers and that they tend to
occupy distinctive locations (clusters) once mapped to the comple. power plane (real vs.
reactive power). The steady-state load classification method wor%s reasonably well at
residential sites& where the ma5ority of loads are constant and small in si2e& and thus
where measuring the step heights are relatively easy and the mappings can be conducted
with fidelity F-armichael& /<<4I.
Two essential attributes of the method are measuring the step heights& both real and
reactive& when a load is turned on or off& and identifying the event via mapping the
measurements on the comple. power plane. Traditionally& the mapping has been
performed on the comple. power plane whose a.es represent the real () and the reactive
(!) power. 1owever& higher harmonic powers (
=
& !
=
& etc.) can also be introduced when
certain conditions are met& as additional dimensions to identify the steady state of a load&
and conseuently the mapping can be performed in the comple. harmonic power space
whose a.es represent the fundamental and harmonic powers. The comple. power space is
defined as the state space of the electric loads.
The problem of step-height measurement is essentially how to define the steady state of a
load and detect its change. The steady state of a load can be defined as the state of the
load when it consumes a relatively constant amount of power (real and reactive) after it is
turned on. ! steady-state event detector should be able to handle various transients and
measure the step heights of an event by computing the differences in the steady-state
powers before and after the event.
The state-space mapping is composed of two parts. The first part is the construction of
load classes in the comple. power space. -onventionally& no prior information about the
load classes was given and the mapping algorithm grouped the step-height measurements
into load clusters. !lternatively& the information about the load classes can be collected a
priori from a set of test data and the mapping can be conducted against the pre-
constructed load classes.
The second part is to classify the step heights of a steady-state event among the load
clusters or classes. @bviously the design of a classifier depends on how much the prior
information is %nown& and to less e.tent how step heights are measured.
1owever& the biggest problem in designing a steady-state load classifier is the possible
spatial pro.imity of the load classes in the comple. power space. This creates an
ambiguity when a steady-state event is mapped to a point in the vicinity of those classes.
The influence of the cluster pro.imity can be reduced partially by applying the techniues
such as changing bases of observation vectors using the $arhunen-LoWve e.pansion
F+u%unaga& /<A3I.
+igure =-/ shows the scatter plot of the steady-state powers of four /4-hp water pumps in
an office building at /04 )ansome )t. )an +rancisco& -!. (ach point of the plot is an
eight-point average of the /34 12 data. The graph clearly shows the overlap between the
chilled water pump / (-1:"/) and the -1:"3 and& to a less e.tent& between the
-1:"/ and the cooling tower pump 3 (-T:"3). Thus it will be very difficult to identify
a steady-state event that happens to fall in the vicinity of these three clusters.

+igure =-/ )catter plot of the steady-state powers of four /4-hp water pumps.
The figure also reveals an interesting fact that a load cluster can be composed of several
distinctive agglomerates. The -T:"/ cluster& for e.ample& is composed of two clearly
=C
separated agglomerates. It is because& at each different @N event& the power consumption
of a load can settle on a slightly different mean and furthermore the mean value can
change over time.
In some cases& increasing the dimension of the state space provides an additional
resolving power to the steady-state load classifier. The mapping is conducted in a
harmonic hyperspace spanned by & !&
=
& !
=
& L (ach basis is orthogonal to each other.
This techniue is useful in resolving a conflict& when two loads have overlapping clusters
in the -! plane but they have different positions on a harmonic a.is FLaughman et al.&
344=I.
The method of steady-state load classification should be able to measure the step heights&
construct the load classes and have a classification rule to identify an event. The design
should be fle.ible enough to select an arbitrary dimension for the harmonic state space.
The power coder generates a stream of harmonic power vectors. -urrently& a power
vector has a length of eight with odd harmonic powers up to the seventh (F !
=
!
=

B
!
B

A
!
A
I). It is desirable to reduce the number of the vectors& as long as the power
vectors after the reduction still contain enough information to identify the steady-state
events. +igure =-3 shows the bloc% diagram of the steady-state load classification.

down-
sampler
change-of-mean
detector
event
associator
event
classifier
load
allocator
harmonic vectors
F" X "= X= L I
.
.
constant mean segments
Fn l m
"
m
X
L I
.
.
steady state event vector
Fn dir A" AX LI
.
.
load status vector
Fs
!
s
>
L I
.
.
load vector
F"
!
"
>
L I
.
.
+igure =-3 >loc% diagram of steady-state load classification.
The reduction of the harmonic vectors can be achieved easily by downsampling. The
downsampling can be followed (or preceded) by a filtering or it can perform its own
filtering. +or e.ample& if a d-point downsampling is achieved by ta%ing a static average of
d data points& the downsampler has a built-in d-point averaging filter. @n the other hand&
a separate filter can be used to smoothen the output from a downsampler. ! variety of
filters can be considered and their selection is important especially when the transients
need to be filtered out F$arl et al.& /<<3I.
The change-of-mean detector analy2es the stream of power vectors and groups them as a
series of constant-mean segments. The output is the stream of constant-mean segments&
each of which contains the inde. (or time)& length& mean estimations& variance
estimations& etc. It basically e.tracts sufficient information to recogni2e the steady-state
events& from the stream of power vectors in the form of the constant-mean segments.
The event associator analy2es the constant-mean segments and associates a series of them
with a steady-state event. !n @++ event is easily identified with a pair of segments& the
second of which has a smaller mean compared to the first segment& in the condition that
the difference is larger than a prescribed threshold. The identification of an @N event is
more involved. It first has to recogni2e the beginning of a transient and ma%e appropriate
decisions for the subseuent constant-mean segments.
If all the loads to be monitored have a similar time scale in their transient response& the
event associator and the change-of-mean detector can be combined into a single steady-
state event detector preceded by an adeuate downsampling and filtering to eliminate the
transients. 1owever& where the loads have heterogeneous time scales& an independent
event association is necessary& because each load has a different time span in its transient
response to be accounted for.
;pon the successful association of a steady-state event& the event associator generates a
steady-state event vector& which contains the inde.& direction (@NE@++)& step heights& etc.
The step heights are simply the differences of the mean estimations between the segment
before the event and the segment where a steady state is reached.
The event classifier identifies the steady-state event vector with a load event by mapping
the step-heights in the state space. ! metric is computed for each load class and the load
class with the most favorable metric is selected. If no class has a favorable metric& the
event is either unclassified or dismissed. The event classifier outputs a load status vector&
i.e. the @NE@++ statuses of all the loads monitored in a single vector form.
The load allocator distributes the total power observed to the individual loads based on
the load status vector. !ny amount corresponding to an unclassified event is added to an
unclassified account. The leftover is assigned to the bac%ground load.
The general structure of the steady-state load classifier was e.plained and its five ma5or
components were introduced in this section. +or the load classifier to be developed in this
chapter& a downsampler based on the d-point static averaging is selected because it is easy
to implement and analy2e& and performs a decent amount of filtering. ! general change-
of-meanEvariance detector is developed in !ppendi. !& which we will employ as the
change-of-mean detector for the steady-state load classification.
3.3 E#ent associator
The change-of-mean detector outputs a time-ordered list of constant meanEvariance
segments from the analysis of its input signal. The event associator analy2es the outputs
?4
from the change-of-mean detector& and associates a part of them with a steady-state event
when certain conditions are met. It will be helpful to e.amine typical outputs of the
change-of-mean detector when there are steady-state events.
+igure =-= shows the turn-on and turn-off events of a /4-hp single-speed water pump.
The original data were ta%en at /34 12 and are downsampled by ta%ing the C-point static
averages. The change-of-mean detector is applied to the stream with the ma.imum
number of mean change #ma& R =& minimum segment length p R = and number of transient
points q R /. The window-shift * is ?B and the nominal window-overlap is p 9 q.


0 50 100 150 200 250 300 350
10
20
30
40
load events, 8 pt static average, 15.00 Hz
P
[
n
]

(
k
W
)
0 50 100 150 200 250 300 350
0
20
40
60
Q
[
n
]

(
k
V
A
R
)
sample number, n
+igure =-= Turn-on and turn-off events of a /4-hp water pump.
+igure =-? shows the mean estimations at the start locations of the corresponding
segments. The graph shows the characteristic of the detector& i.e. it detects more segments
where there is more activity and less where there is less activity. It identified four distinct
segments where there is an @N event. It also identified two segments where there is an
@++ event. In between& the segments typically have a full length of ?C points& implying
the detector found a single segment only for a given window.

The @N event is composed of four sections7 pre)event (n R =A& length R ?)& plateau (n R
?3& length R =)& transition (n R ?0& length R A) and steady state (n R B?& length R =B).
There can be many association strategies to identify an @N event from this type of
assortments. The simplest is idling7 after identifying a sharp step increase& the associator
waits a certain amount of time until it declares a steady state. The idling time can be
different when monitored loads are of different si2e. The nominal idling time for a load
(or load group) can be recorded in a load database. >ecause a load with a longer turn-on
response time tends to have a higher plateau step-height& an appropriate idling time can be
selected from the database based on the step-height measurement.

0 50 100 150 200 250
10
15
20
25
30
35
Change of means
m
P
[
n
]

(
k
W
)
0 50 100 150 200 250
0
20
40
60
m
Q
[
n
]

(
k
V
A
R
)
first index of a const. mean segment, n
+igure =-? -hange-of-mean segments of the signal in +igure =-=.
1owever& the aforementioned idling strategy does not fully utili2e the information
provided by the change-of-mean detector. +or e.ample& the associator may actively
pursue a series of segments& starting with a large mean and moderate variance plateau
segment& followed by a reduced mean and increased variance transition segment& and
anticipating a steady-state segment with a substantially reduced variance. @n the other
hand& the associator may perform a crude pattern recognition using the information and
identify& at least& the type or group of a load.
The idling strategy will be used in this chapter& due to its simplicity. >ecause it ta%es
several segments to complete an @N event& an associator flag is set after an initial step
change is observed and until a steady state is reached. @nce the associator decides that a
steady state has been reached& it computes the step heights between the steady-state
segment and the pre-event segment and generates a steady state event. The flag is
released.
The @++ event which occurred at n R 3=4 can be easily associated with the sudden drop
of the average value (of ) compared to the previous segment. Thus an @++ event is
generated when the average of a current segment is reduced by more than a preset
threshold& compared to the average of a previous segment. The threshold can be selected
easily as a fraction of the minimum of the steady-state powers recorded in the load
database. !n @++ event is generated only when the associator flag is not set& because a
large negative step change can occur as a during a turn-on response.
The event associator analy2es the outputs from the change-of-meanEvariance detector&
locates steady-state events and e.tracts relevant information for the event classifier. It
generates messages with the inde.& direction (upEdown)& step heights& segment lengths
before and after the event and total " value after the event. Table =-/ shows the steady-
state events identified by the associator for the constant meanEvariance segments of
+igure =-?.
?3
Table =-/ )teady-state events generated from the segments of +igure =-?.
inde. direction
A (%:) A! (%8!,)
length- length9 (%:)
B? up 0.C03? B.4/== ? =B /A.3CA
3=4 down -0.<4?A -B.4/3C C =< /4.?/4
3.. Steady!state e#ent classi'ier
The steady-state events created by the event associator need to be e.amined by the event
classifier to determine whether they belong to certain load classes. The event
classification reuires first the construction of load classes and then the design of the
classification procedure or criterion itself.
In the case of a residential load monitoring& it usually is difficult to collect and analy2e
training data manually. Thus& it is a common practice to program the system to collect
observations for a certain period and group them as clusters. !fter the training period& an
event is classified against these load clusters. The clusters contain all the necessary
information to classify an event& but they are not associated with the physical loads. The
actual load identification occurs when the load clusters are matched against the load
library F#ren%er and $ader& /<<<I. In this approach& the load cluster acts as an
intermediate load class& with its name un%nown until it is connected to the load library.
+or commercial buildings& it is generally possible and ma%es sense to collect training data
first and compile a priori information to construct the load classes. +rom the training
data& pertinent information& such as the mean and variance of the steady-state power of a
load& can be e.tracted to constitute a load class along with its name and nominal rating.
In this chapter& we assume a priori %nowledge of the loads monitored and the e.istence of
the load classes& and concentrate on the development of a classification method.
#epending on the circumstances& the load classes will be created e.plicitly from a set of
separate training data or implicitly via the clustering approach.
)uppose A@ is a steady-state event vector. Its components are the step changes in the
harmonic powers and its length is eual to the dimension of the harmonic power space
used for mapping (A@ R FA A! L I
T
). -onsider the following multiple hypothesis test7

& load to belongs 7
3 load to belongs 7
/ load to belongs 7
3
/
m $
$
$
m
@
@
@
A
A
A
.
(=./)
where m is the number of loads being monitored. The direction of an event (@N or @++)
can be easily identified from the sign of A.
The minimum distance classifier is widely used due to its simplicity. It computes the
(uclidean distances from A@ to the class centers and selects a minimum. 1owever&
because each class has a different statistical distribution property& the distance is
normali2ed by its variances. +or e.ample& on a 3-dimensional plane& the following
normali2ed distance is typically used7
& ) & (
3
3 &
3 & 3
3
/ &
/ & /
|
|
.
|

\
| A
+
|
|
.
|

\
| A
= A
i
i
i
i
i
& &
d
o

o

F @ (=.3)
where i and oi are the mean and the covariance of class i& respectively. +urthermore& the
event vector needs to be rotated first by an angle which the ma5or a.is of the ellipsis of
the cluster i forms with the hori2ontal a.is F"alomera-!rias and Norford& 344/I. The
aforementioned procedure is essentially a whitening transform& which is composed of the
$arhunen-LoWve ($L) e.pansion and the normali2ation F+u%unaga& /<A3I.
1owever& the whitening transform is cumbersome in the sense that the eigenvalues and
eigenvectors of a covariance matri. have to be found to construct the transformation
matri.. Instead& the following simple conditional ma.imum-li%elihood (ML) rule is
proposed for (=./)7

). Y ( ma. arg
U
i
$
#
$ p $
i
@ A =
(=.=)
The ML classifier uses the conditional "#+ of A@ given $i as the test statistic. To
evaluate the conditional "#+& the probability distribution of A@ given $i should be
%nown. +rom the observations& a normal distribution is assumed& i.e.

& ) ( ) (
3
/
e.p
) ( det ) 3 (
/
) Y (
/
3
/
3
|
.
|

\
|
A A = A

i i
T
i
i
i N
$ p F @ G F @
G
@
t
(=.?)
where N is the length of A@ and i and Ai are the mean vector and the covariance matri.
of load class i& respectively. +or simplicity& it is assumed that i and Ai are deterministic
and %nown at the time of classification.
To simplify the test& a new test statistic i is defined as

( ). ) det( ln ) ( ) (
3 ln ) Y ( ln 3
/
i i i
T
i
i i
N $ p
G F @ G F @
@
+ A A =
A =

t
(=.B)
:ith the new test statistic& the hypothesis test simply loo%s for the minimum of i

& & min arg
U
< =
i i
$
#
i
$
(=.0)
??
where is the threshold of the test. If there is no i that satisfies the threshold
reuirement& the event is unclassified.
The new test (=.0) is easy to implement and utili2es the covariance matri. as it is& unli%e
the minimum distance classifier that needs a whitening transform. !lso& it can be used for
any steady-state event vector with an arbitrary dimension& because the test is not bounded
by the (planar) geometrical notion of a load class distribution.
@ne subtlety in implementing the test is the fact that the event vector A@ and the
statistical property of a class (i and Ai) should be compatible& i.e. they should be able to
form a probability space from which the samples of A@ can be drawn.
To maintain statistical compatibility& the power vector @FnI R F ! L. I
T
is considered as
a random process& which is inde.ed by the discrete-time n and is generated at a power
coding rate. The probability space of load class i is defined at this coding rate and its
random variable is the load vector siFnI& i.e. the power vector when only load i is @N. !s
e.plained before& the load vector is modeled by a Daussian (normal) distribution& i.e. siFnI
+ N(i& Ai). +urthermore& we assume that two different time samples of a load vector are
statistically independent& i.e. siFnI R i 9 ?iFnI& where ?iFnI is 2ero-mean white Daussian
noise (:DN). !lso& it is reasonable to assume that two load vectors from two different
classes are independent from each other.
If # loads are on at time n& the power vector @FnI is simply the sum of those individual
load vectors
. I F I F
/

=
=
#
i
i
n n s @ (=.A)
>ecause any linear combination of independent Daussian random vectors is also Daussian
F"apoulis and "illai& 3443I& @FnI is Daussian with
mean7 ( ) ( )

= =
= = =
#
i
i
#
i
i
n ( n (
/ /
I F I F F s @ m
@
(=.C)
covariance7 .
/

=
=
#
i
i
G G
@
!lso& two different time samples of @FnI are statistically independent.
)uppose @ is an * point average of @FnI under a load status& i.e.

=
=
*
n
*
n
/
/
I F @ @
. >ecause
different time samples of @FnI are statistically independent& it can be shown that its time
average is also Daussian& i.e. ) & ( T
@ @
G m @ N & where
@ @
m m = and
@ @
G G
*
/
= F"apoulis
and "illai& 3443I. Thus an *-point average of @FnI has the same mean as @FnI& but with a
covariance matri. reduced by /E*.
Now& suppose that a steady-state event vector A@ is computed from segment j and
segment j J /& as the difference of two segment-averages7

.
/
= A
j j
@ @ @
(=.<)
(ach average is computed with *j and *j-/ samples& respectively. If # loads are on at
segment j J /&
( ). & T I F
/ /
/
/ /

= =

= =

#
i
i
#
i
i j
j j
N n G G F m @
@ @
If load m (m e Q/& 3& L&
#P) is on& at segment j& ( ). & T I F
/ /

= =
+ = + =
#
i
i m
#
i
i m j
j j
N n G G G F F m @
@ @
Now it is
straightforward to show that A@ is Daussian with m
j j
F m m m
@ @ @
= =

A
/
and
/ / /
/ /

+ = + =
A
j j j j j j
* * @ @ @ @ @
G G G G G
.
:hen the power vectors of segments are d-point static averages of raw power vectors& the
mean of A@ is unchanged& but its covariance matri. is changed due to the additional
averaging. +or an @N event of load m&

( ). & T
/ /
/ /

+ A
j j j j
d* d* m
N
@ @
G G F @
(=./4)
+or an @++ event&

( ). & T
/ /
/ /

+ A
j j j j
d* d* m
N
@ @
G G F @
(=.//)
In each case& j is the event segment and j J / is the pre-event segment.
/ j
@
G
can be
computed by summing up the covariance matrices of the loads which are on.
j
@
G
is
obtained by adding Am (@N event) to or subtracting it (@++ event) from
/ j
@
G
. It is
obvious by now that it is the normal distribution of A@& not the distribution of a load& to
be used for the hypothesis test. Thus when (=.B) is computed for each load class& the load
covariance matri. needs to be replaced by the event covariance matri. (=./4 or =.//). The
advantage of the ML classifier against the minimum distance classifier is apparent here&
because the latter has to perform the whitening transform for every event hypothesis due
to the changing event covariance matri. determined by the nature of the signal.
The statistical properties of a load class are estimated from its samples. >ecause its true
mean and covariance are un%nown& they are usually appro.imated by the sample mean
and covariance. The uality of estimation increases with the increasing number of
samples& if the random process is ergodic F"apoulis and "illai& 3443I.
?0

+igure =-B "#+ map of four pump classes (in +igure =-/).
In its original form (=.=)& the hypothesis test simply loo%s for a load class that gives the
largest conditional "#+ assuming $i is right. @n a 3-dimensional power plane& with the
third a.is as the height of "#+s& the test can be conceived as a search for a hypothesis
which gives the highest elevation for a given event point. +igure =-B shows the "#+ map
of four pump classes whose mean and covariances are obtained from sample means and
covariances. 1owever& it should be remembered that an event covariance matri. replaces
the load covariance matri. at the time of test.
Two events generated in the previous section are the @N and @++ events of the -1:"3.
The event messages are fed to the event classifier and the classification results are shown
in Table =-3.
Table =-3 -lassification results for the steady-state events of Table =-/.
inde. direction
A (%:) A! (%8!,)
load name
i
B? on 0.C03? B.4/== -1:"3 00.4C
3=4 off -0.<4?A -B.4/3C -1:"3 ?4.3B
The classifier successfully associated the events with the correct load class& with
reasonable test statistic values. The classifier also updates the load status upon a
successful classification. The load status is simply a vector with >oolean entries (/ for
@N& 4 for @++). If there are # loads being monitored& the load status vector s is
represented by P / & 4 Q & I F
3 /
e =
i
T
#
s s s s . s .
In this section& a simple ML rule is developed as the event classifier and the statistical
distribution of an event vector is specified. The resulting event classifier is easy to use
and can be applied to any series of power vectors with an arbitrary dimension and time-
averaging. ;pon the successful classification of a steady-state event& the classifier updates
the load status.
3.D Test o' steady!state load classi'ier
The overall steady-state load classifier was assembled with the components developed in
the previous sections and tested with a few off-line data sets. +igure =-0 shows two sets
of the -T:"/ @N and @++ events used to test the steady-state load classifier. The figure
shows the real and reactive power after the C-point static averaging. #ue to the activities
of the bac%ground load& the waveforms are highly irregular& suggesting the difficulty of
the step-height estimation. The classifier succeeded in the second @N-@++ event set& but
misclassified the first set as the -T:"3 events. The misclassification of the first @++
event is inevitable& because only the -T:"3 is assumed @N at the time of the search.
Table =-= shows the classification results.

0 200 400 600 800 1000 1200
10
20
30
40
50
load events, 8 pt static average, 15.00 Hz
P
[
n
]

(
k
W
)
0 200 400 600 800 1000 1200
0
20
40
60
Q
[
n
]

(
k
V
A
R
)
sample number, n
+igure =-0 Two -T:"/ @NE@++ event sets.
Table =-= -lassification results for the steady-state events in +igure =-0.
inde. direction
A (%:) A! (%8!,)
load name
i
33A on <.34= ?.0BC -T:"3
A.<3C/4
=
B?? off -A.ACB -?.43B -T:"3
3.C4?/4
=
A=C on A.BA3 =.<?/ -T:"/ -/.B3C
/4/< off -A.BAB -=.</4 -T:"/ /./<<
! closer e.amination of the figure reveals that a small bac%ground load is turned on at n
R 330& before the -T:"/ reaches its steady state& whereas the second @N event is free of
the bac%ground activities for the most of its steady state. -onseuently& the first step-
height is overestimated by /.B %: (step-height of the bac%ground event)& because the
event associator is programmed to wait appro.imately two seconds after detecting a
pump)li#e initial step change.
@ne may argue that by allowing an ample waiting time for the event associator& the
problem of bac%ground activity may be solved. 1owever& even with an e.tended idling
?C
time& there is no guarantee that the point at which the associator assumes a steady state is
not sub5ect to a bac%ground activity. ! longer waiting time simply gives more chances to
the bac%ground activities and thus the step-height measurements are prone to more errors.
@n the other hand& if the waiting time is too short& a steady state is declared when the
turn-on response of a load is still in its transition portion& whose mean is different from
the one of the steady state and usually is not repeatable. (ither approach (shorter or longer
waiting time) does not solve the problem of measuring the step-heights under active
bac%ground load activities.
The difficulties in defining and detecting a steady state and in classifying it against
pro.imal load classes are already discussed. These difficulties are partially addressed by
designing the general change-of-meanEvariance detector& event associator and event
classifier. The additional difficulty is introduced by the bac%ground activity& which can
interfere with the estimation of the step-heights. The steady-state load classifier alone can
not solve the problem of bac%ground load activity. It needs help from its siblingJ a
transient pattern recogni,er.
3.E )oad Classi'ication $ased on transient pattern recognition
! load classifier based on the state space mapping was developed in the previous
sections. >y utili2ing the fact that constant loads occupy distinctive locations in the
harmonic power space& the classifier could classify @NE@++ events with a reasonable
accuracy. 1owever& the classifier performs poorly when the load classes are ad5acent to
each other in the power space and when there are many bac%ground load activities.
+rom this section& the transient response of a load is treated as a source of features and
classified& rather than a nuisance in defining and detecting a steady state. Two classifiers
wor% in parallel& utili2ing two different aspects of a load response- the transient and the
steady state- and thus ma%ing a better use of the available information that can be
e.tracted from the harmonic power signal. ! certain amount of coordination will be
necessary to place two different classifiers simultaneously at wor%.
! transient response is observed when a load moves to and from its steady-state. Most
loads& such as single-speed motors& incandescent bulbs and florescent lamps& show
characteristic transients FLeeb& /<<=G Leeb et al-& /<<BI. 1owever& transient patterns are
not genuinely uniue and some loads can have similar transient patterns& especially when
they have similar ratings. The patterns can also have their own statistical properties.
Thus& the transient pattern recogni2er is basically another classifier and is sub5ect to the
same limitations as those of the steady-state load classifier& i.e. the overlaps in time and
space and the bac%ground load activities. 1owever& it at least brings in a redundancy
because it utili2es another aspect of a load event overloo%ed by the steady-state load
classifier. @n the other hand& it is e.pected to perform better in detecting and classifying a
load event& because the transient patterns have a more discernibility than the steady state
powers (the steady state mapping is performed with only two or three points whereas the
pattern recognition is accomplished with tens to hundreds data points). !lso it will be less
vulnerable to the bac%ground activities& because the pattern width is typically less than the
waiting time of the event associator and their effects are lessened due to the higher power
levels during a transient (higher )N,).
+igure =-A shows the transient responses of four /4-hp water pumps installed at /04
)ansome )t. The @N transient shape of a single-speed motor typically consists of a
tranuil pre-event segment& plateau (a relatively flat portion after the initial step increase)&
transition and steady state. !lthough the whole transient response shape can be used for
pattern recognition& it will be wiser to utili2e only the plateau and the beginning part of
the transition both to minimi2e the influence of the bac%ground load activities and to
reduce the computational burden.

0 20 40 60 80
0
10
20
30
40
50
Transient response of CHWP1
P
[
n
]

(
k
W
)
0 20 40 60 80
0
10
20
30
40
50
Transient response of CTWP2
n, sample number
0 20 40 60 80
0
10
20
30
40
50
Transient response of CHWP2
0 20 40 60 80
0
10
20
30
40
50
Transient response of CTWP1
+igure =-A. Transient response (@N) shapes of four /4-hp pumps.
! close e.amination of +igure =-A shows that the response shapes are generally similar
but differ in detail. 1owever& the -T:"/ and the -T:"3 have largely the same response
shape& although the -T:"/ has a slightly higher plateau. The pattern recogni2er may
have a difficulty in differentiating them. 1owever their steady-state clusters were& at least&
separate (+igure =-/). Thus it demonstrates the advantage of having both the pattern
recogni2er and the steady-state load classifier to complement each other.
The pattern recognition is essentially composed of two parts7 detection of a pattern and its
classification. In order to recogni2e a pattern& it first needs to be detected. -onsidering the
fact that most transient responses of the loads are accompanied by the power level
changes& a change-of-mean detection scheme& such as the cumulative sum (-;);M)
algorithm& can be used to detect a transient pattern F>asseville and Ni%iforov& /<<=G Leeb
et al-& /<<BI. )ometimes& a transient response (or a section of it) shows changes in its
variance only& not accompanied by a power level change. ! type of change-of-variance
detector is necessary to capture this sort of transient. @ne approach is to compute a finite-
impulse-response (+I,) filtered version of a signal and threshold the difference between
the +I, filtered version and the original signal F)haw& 3444I. This approach wor%s well
B4
when there are sharp changes& but performs poorly when there are moderate changes and
oscillations with relatively small magnitudes.
1owever& a very versatile change-of-meanEvariance detector was already developed for
the steady-state load classification and thus it can also be used to detect the transient
patterns. This approach also ma%es sense in the viewpoint that because two classifiers
wor% in parallel& tapping the same power signal& they need to be triggered by a single
event generator. Thus the pattern classifier will receive its input (raw pattern) from a
modified event associator that will prepare the outputs for both the steady-state classifier
and the pattern classifier once an event is detected.
@nce a transient pattern is detected and e.tracted& it is the 5ob of the transient pattern
classifier to compare the pattern against the e.emplars of %nown transient responses.
(ach e.emplar may have a different signal representation and the observed pattern may
have to be decomposed accordingly for each e.emplar to be compared. (ach comparison
produces a matching gain and in the end the classifier selects a load class that offers the
most favorable gain. ! pattern or e.emplar can be composed of several sections. In this
case& matching gains are computed for each section and the overall single gain is
computed from the section gains.
1owever& when two classifiers wor% in parallel& there is no guarantee that two will
produce the same result. Thus an arbitrator is necessary to determine a load event when
they come up with contradicting results. >ased on the outcomes of two separate
classifiers& three situations are possible7
/) >oth classifiers recommend the same load class and at least one of them has an
acceptable classification gain7 This is the most wanted case and the event is easily
recogni2ed with a high level of confidence.
3) (ach classifier comes up with a different load class& but one classifier has a more
favorable gain to the other7 The load class with the more favorable gain is selected for the
event. @bviously& the level of confidence is low.
=) Neither classifier comes up with an acceptable gain7 This implies that both
classifications are unreliable. The event is unclassified.
+igure =-C shows the bloc% diagram of the overall load disaggregator with the steady-state
load classifier and the transient pattern classifier placed in parallel.
There is also the uestion of how many channels should be utili2ed for the pattern
recognition& i.e. whether to use only the real power or use the reactive and other harmonic
powers. Initial study suggests that the real power () channel alone is enough for most
pattern recognition purposes. If needs arise& other channels can be searched for useful
patterns. The patterns from other channels are considered as additional sections and the
overall classification gain is computed as usual.

change-
of-mean
detector
event
associatorE
pattern
e.tractor
state
classifier
load
allocator
pattern
classifier
arbitrator
graphic
display
steady-state
event
transient
pattern
classification
result
load-status
vector
load
vector
+igure =-C >loc% diagram of the load disaggregator with two classifiers.
There basically are two types of events for the purpose of the load classification- @N and
@++. !n @++ event& in most cases& is a simple step-li%e downward transition with no or a
few transition points between two different power levels. Thus the information
e.tractable from an @++ pattern is merely its step-height.

@n the other hand& the @N events are typically accompanied by the distinctive transient
response shapes and thus their identification can be greatly benefited from the transient
pattern recognition. In this chapter& the transient pattern recogni2er is employed to
identify the @N events only. The @++ events are solely identified by the steady-state load
classifier. If needs arise& the pattern recogni2er can be tapped later to help identifying the
@++ events.
3.5 Transient pattern classi'ier
The transient pattern recogni2er is essentially another type of classifier. It basically
classifies the observed transient pattern by matching it against the %nown set of transient
e.emplars (classes). !n e.emplar with a best match is selected and the observed pattern
is associated with its load.
To recogni2e a pattern& it first needs to be represented in a suitable format and then tested
according to a classification rule. :hen a transient pattern is very long& it will be more
economical to e.tract a few distinctive sections than to save the whole length of the
signal.
! section needs to be represented by a basis& which can emphasi2e distinctive and
repeatable features and discourage non-characteristic and non-repeatable features of the
signal. Many types of basis are possible& such as time& wavelet& +ourier& etc. Nevertheless&
B3
whatever basis is used& once the signal is decomposed& the result is another finite length
vector (or a set of vectors) in *
3
( ) space& spanned by the selected basis. Thus& the way a
pattern is represented does not affect the core design of a classifier& although provisions
have to be made to accommodate different weighting strategies and the possibility of
subsections depending on the basis choice.
@nce the observed transient pattern is represented in the same manner as an e.emplar to
be matched against& its membership to the e.emplar needs to be tested by a classification
rule& or classifier. There can be various types of classifiers. )imple classifiers use
(uclidean distances between the pattern and e.emplars& or normali2ed inner products of
the pattern and e.emplars FLeeb& /<<=I. ! general purpose pattern classifier based on the
method of least suares was developed and tested by )haw F3444I.
If the statistical properties of the transient responses are %nown completely& a >ayesian or
ML type classifier can be designed. 1owever& estimating the statistical property of a
transient pattern poses a problem. (ven if the probabilistic distribution of a pattern is
assumed Daussian& its mean and covariance need to be estimated to compute its "#+. The
mean of a normal distribution can be estimated reliably with tens of samples.
Nevertheless& the covariance estimation needs at least hundreds of samples. ! few
transient responses may be obtained during a training period. >ut obtaining one hundred
responses is practically impossible.
Thus& the classifier is li%ely to have to ma%e decisions in an environment where there
e.ist a very crude estimation of the mean& obtained from one or a few samples& and
virtually no information about the covariance of each class. In this situation& designing a
classifier based on the method of least suares (L)) appears to be a viable strategy&
because it does not reuire any probabilistic assumptions about the classes F$ay& /<<=I.
The advantage of the L) classifier is that it can be used in an environment where the
additive noise is non-Daussian or changes with time. The down side is that no optimality
can be claimed because of the lac% of the statistical structure.
)uppose that y represents a data vector (observed pattern) and @ an e.emplar vector. :e
want to uantify (or estimate) the amount of correlation between y and @& so that a
decision can be made whether y belongs to @. :e assume that y can be linearly
appro.imated by @& with an offset term to account for a possible difference in the #-
level. !ny difference in the appro.imation is attributed to an un%nown additive noise
vector ?.
y R o@ 9 |1 9 ? R Aa 9 ?& (=./3)
where A R F@ 1I and a R Fo |I
T
. o is the correlation (matching) gain between the data
and the e.emplar. If @ and y are identical& o R / (and ? R /). !s @ and y are uncorrelated&
o diverges from the unity. | is the offset term. :e want to minimi2e the second norm of
the error vector& by an optimal choice of the parameter vector a& i.e.
minimi2e7
3 3
Aa y e = . (=./=)
This is a standard least suares problem. The solution is given by the following matri.
formula7
. ) ( I
U
U F U
/
y A A A a
T T T
= = | o (=./?)
oU is the L) estimation of the correlation gain between y and @.
:e define the residue of estimation r as the suare of the minimum second norm of the
error vector. It often is necessary to normali2e the residue by the length of vectors. If both
@ and y are of length .& the normali2ed residue of estimation rn is defined by
. U
3
/ /
a A y = =
. . n
r r (=./B)
1owever& the above discussion does not include the fact that there can be a time offset
(shift) between y and @& in the case of a time-based signal representation. Thus the best
shift also has to be found as a part of the minimi2ation process. Let y is of length N and @
of length .. +or simplicity& we assume N > .. The original minimi2ation of a least
suares error criterion now becomes a double minimi2ation problem.

& min arg )
U
& U (
3
4 &
Aa y a =
s s
i
. N i a
i
(=./0)
where | |
T
i
. i y i y i y I / F I / F I F + + = . y . aU and
i
U
are the optimal estimations of the
parameter vector and the shift& respectively. In implementation& for each i the parameter
vector is estimated and the resulting residue is computed& then
i
U
and aU are determined by
finding the minimum of the residues. !lso& note that (A
T
A)
-/
needs to be computed only
once because @ is not shifted and thus at each i only A
T
yi needs to be obtained.
:hen there e.ist multiple sections for a pair of pattern and e.emplar& the L) estimation
of the correlation gain and the resulting normali2ed residue is computed for each section.
)uppose there are # sections. :e also introduce a weighting factor wj (4 J /) for section j.
Now we solve a least suares problem again to obtain the overall correlation gain a.

& ) (
/
$ 6 6 6
$ 6
T T
a
a

=
~
(=./A)
where
T
#
w w w I F
3 /
. = 6 and
T
# #
w w w I F
3 3 / /
o o o . = $ . oj is the correlation gain of
section j. The normali2ed residue of the overall estimation is also defined as7

&
/ /
/
/
/
/

=
=
=
=
= =
#
j
n j #
j
j
#
j
j #
j
j
j
r .
.
r
.

(=./C)
B?
where .j& rj and
j
n
r
are the length& residue and normali2ed residue of section j&
respectively.
In this manner& the correlation gain a and the normali2ed residue can be computed for
each e.emplar given an observed pattern vector y. Now it is time to determine whether y
represents a load event and if so which load (e.emplar). !gain& this is a multiple
hypothesis test problem.

& ) ( load to belongs 7
) ( 3 load to belongs 7
) ( / load to belongs 7
3 3
/ /
m m
m $
$
$
@ y
@ y
@ y
.
(=./<)
where m is the number of loads being monitored. There can be a few test criteria& but the
simplest is to choose the e.emplar that gives a minimum normali2ed residue& i.e.

& min arg
U
thres i i
$
#
i
$ < =
(=.34)
where
i

is the normali2ed residue of e.emplar @i and thres is the threshold of the test.
This rule is appealing in the sense that it is a natural e.tension of the L) error criterion.
The minimum normali,ed)residue rule appears to wor% best in most circumstances.
1owever& it was observed to perform poorly when the (Daussian) noise magnitude is
large. It can be e.plained by the fact that as the noise magnitude increases (lowers )N,)
the normali2ed residues also increase in general and are affected more by the random
fluctuations of the noise and less by the differences between the true signal of an
observation and e.emplars.
!nother measure which can be used for classification is the correlation gain a. >ecause
the gain can be any real number& it first needs to be normali2ed before comparison. The
gain a is normali2ed by the following normali2ation function& f(a)

& / if & /
/ 4 if &
4 if & 4
) (
>
s <
s
= =
a /a
a a
a
a a f
n
(=.3/)
where an is the normali2ed gain& which has a range of 4 to /.
:ith the normali2ed gain for each e.emplar& it is tempting to set up a classification rule
that selects a hypothesis that yields a ma.imum normali2ed gain. 1owever& the
normali2ed gain classification rule performs poorly& because the gain is simply the slope
of the first order appro.imation and lac%s any measure of the scatter of the data from the
appro.imation (e.g. y R F4 4 4 4 /4 /4I
T
and @ R F4 4 B B B BI
T
gives a R /& though two
vectors are radically different.).
Thus& it is necessary to ta%e into account both the normali2ed gain and the normali2ed
residue in selecting a load class. @ne solution is to rescale the normali2ed residue and
multiply it with the normali2ed gain to obtain a composite gain i7

&
/
/
|
|
.
|

\
|
=
i
thres
thres
n i
a
a
i

(=.33)
where
thres i n thres
i
a a < s s < 4 and /
. Now the ma&imum composite)gain rule simply
loo%s for an e.emplar (load) that gives a ma.imum composite gain7

. ma. arg
U
i
$
#
i
$ =
(=.3=)
The ma.imum composite-gain rule wor%s better than the minimum normali2ed-residue
rule in an environment where Daussian noises have large magnitudes.
:hen the transient pattern classifier and the steady state-load classifier produce different
results& an arbitration is necessary. The only case when a true arbitration is necessary is
when both classifiers have acceptable gains but differ in terms of the declared load
classes. Though there can be many approaches& we design a simple arbitration rule based
on the rescaling of the gain of the steady-state load classifier.

& 4
/
/
4 /
thres
thres
thres n
a

< <

s
= (=.3?)
where n is the normali2ed (rescaled) gain of and thres is the threshold of . athres is the
threshold for the normali2ed gain an of the transient pattern classifier. Note n is
normali2ed to the scale of an. -onsider the following binary hypothesis test and its
empirical test rule7
$47 The output of the steady-state classifier is true.
$/7 The output of the transient pattern classifier is true.
(=.3
B)
&
/
4
U
U
n
$
$
n
a c
<
>
+
where c is the weight of the test. If c Z 4& it favors the transient pattern classifier and&
otherwise& the steady-state classifier. >ecause the pattern matching in general produces
better results the weight is usually positive.
B0

15 20 25 30 35 40 45 50 55 60 65
-5
0
5
10
15
20
25
30
35
40
45
50
n, sample numbers
P

(
k
W
)
Four transient responses of CTWP2 ON event
response 1
response 2
response 3
response 4
average
+igure =-< +our transient responses of -T:"3 @N event and their average.
!s evident from the development of the classifier based on the least suares error
criterion& it performs better when the e.emplar of each class is close to the true mean of
the transient response. Thus& when there are multiple samples available& it is better to
average them& rather than to ta%e a single sample as an e.emplar. The e&emplar
averaging enhances the performance of the pattern classifier especially when there are
similar loads. +igure =-< shows the average of four responses of the -T:"3 @N event.
In this section& the issue of representing a signal for recognition was discussed and a
pattern classifier based on the least suared error criterion was designed. The classifier
can handle a pattern with multiple sections& and does not reuire the statistical
characteri2ation of load classes but a single e.emplar for each class. )election rules were
developed for the classifier and the arbitrator.
3.H Test o' transient pattern classi'ier
The transient pattern recogni2er is assembled and placed in parallel to the steady-state
load classifier. The overall load disaggregator with the two parallel classifiers was tested
in a variety conditions.
The two sets of -T:"/ @NE@++ events (+igure =-0) are used again to test the new
classification scheme. :e remember that the first @N event was misclassified as the
-T:"3. The minimum normali2ed-residue rule is used to select the best matching
e.emplar and the following test parameters are used7 thres R /.B/4
?
& athres R 4.0& thres R ?4
and c R 4.40. Table =-? shows the result of @N event classifications.
Table =-? Transient pattern classification results for the @N events in +igure =-0.
inde
.
load name no. of sections a

34A -T:"/ / 4.<AC0 4.C=/B
A4< -T:"/ / 4.<<3? 4.33=/
:e observe that the first @N event is successfully matched to the -T:"/ @N by the
pattern classifier and that the arbitrator will rule in favor of the pattern classifier. +igure
=-/4 shows the pattern matching result. In the picture& solid line denotes the observed
pattern and [6s represent the matched e.emplar. +or the second @N event& both
classifiers agreed that it belongs to the -T:"/. Thus the load disaggregator& with the
help of the transient pattern classifier& correctly identified both @N events. #uring the
test& for a variety of =? @N events& the load disaggregator successfully classified all the
events (4 O misclassification rate).
!s discussed before& one advantage of the L) estimator is that it can be used in an
uncharacteri2ed noise environment. :e add a :DN with the standard deviation o R =
(%:) to the -T:"/ waveforms and test how the pattern matcher performs. :e simulate
an un%nown change in noise characteristic and thus the load database for the steady-state
load classifier is not updated The ma.imum composite-gain rule is used to select the best
matching e.emplar and the following test parameters are used7 thres R /.B/4
?
& athres R 4.0&
thres R 04 and c R 4.40.

1620 1640 1660 1680 1700 1720 1740 1760 1780
10
15
20
25
30
35
40
45
50
55
n, sample number (120 Hz)
P

(
k
W
)
207: CTWP1 , alpha=9.786e-001, residue=8.315e-001, 1 sections
+igure =-/4 Matching result of -T:"/ @N event.
Table =-B shows the result of the steady-state load classification and Table =-0 shows the
result of the transient pattern classification. #ue to the increased noise& the steady-state
classifier misclassified both @N events. 1owever& the transient pattern classifier correctly
classified the events with favorable gains so that the arbitrator can decide that those
belong to the -T:"/.
BC
Table =-B )teady-state load classification results for the waveforms in +igure =-0 with a
:DN (o R = (%:)).
inde. direction
A (%:) A! (%8!,)
load name
i
33/ on <.B?3 ?.B/3 -T:"3
3.3??/4
?
B?? off -A.C3C -?.044 -T:"/
/.CAB/4
=
A/0 on C.<0B ?.B<3 -T:"3
/.4=C/4
?
/4/< off -0.0B? -=.CAA -T:"/
?.<B=/4
=
Table =-0 Transient pattern classification results for the @N events in +igure =-0 with a
:DN (o R = (%:)).
inde
.
load name no. of sections a

34A -T:"/ / 4.<0=/ A.<?4
A4C -T:"/ / /.4== //.?<
+igure =-// shows the matching result of the first @N event. Thus& we verify that the L)
transient pattern matcher performs very well even under a severe noise environment. #ue
to the pattern matcher& the load disaggregator becomes more robust to noise. +or a variety
of =? @N events corrupted with the :DN& the load disaggregator misclassified only one
event (3.<? O misclassification rate).

1620 1640 1660 1680 1700 1720 1740 1760 1780
0
10
20
30
40
50
60
n, sample number (120 Hz)
P

(
k
W
)
207: CTWP1 , alpha=9.631e-001, residue=7.940e+000, 1 sections
+igure =-// Matching result when the pattern is corrupted by a :DN (o R = %:).
It often is the case that the noise added to a power signal is not Daussian but has a certain
bandpass characteristic. !n unbalanced rotating machine can generate a harmonic noise.
The harmonic (or periodic) noise can usually be estimated and thus can be filtered.
1owever& in some cases& the estimation (or filtering) is not feasible& and the pattern
matcher has to wor% with corrupted signal.
+or the test& a set of -T:"/ events is corrupted by a harmonic noise& obtained from the
/04 )ansome )t. building& and an additional :DN with o R / %:. +or the test& the
minimum normali2ed-residue rule is used to select the best matching e.emplar and the
following test parameters are used7 thres R /.B/4
?
& athres R 4.0& thres R 04 and c R 4.40
Table =-A )teady-state load classification results for a -T:"/ event set corrupted with
harmonic and white Daussian (o R / %:) noise.
inde. direction
A (%:) A! (%8!,)
load name
i
C0 on C./B< ?.?/? -T:"/
/.CBB/4
=
343B off -A.<40 -=.<34 -T:"/
A.//A/4
/
Table =-C Transient pattern classification results for a -T:"/ event set corrupted with
harmonic and white Daussian (o R / %:) noise.
inde
.
load name no. of sections a

?= -T:"3 / /.34/ 0.?C?
Table =-A and =-C show the classification results. The steady-state load classifier correctly
identified the @N event& whereas the pattern matcher misclassified it to that of the
-T:"3. +igure =-/3 shows the matching result& which shows that one of the noise pea%s
coincided with the beginning of the transient response. 1owever& because the steady-state
classification had a more favorable gain& the arbitrator could correctly determine that it
was a -T:"/ event.

300 350 400 450
0
10
20
30
40
50
60
n, sample number (120 Hz)
P

(
k
W
)
43: CTWP2 , alpha=1.201e+000, residue=6.484e+000, 1 sections
+igure =-/3 Matching result for a -T:"/ @N event corrupted by harmonic and white
Daussian (o R / %:) noise.
This is an e.ample where the transient pattern classifier& which usually performs better&
ma%es a mista%e and is corrected by the steady-state classifier. ;npredicted noise& spi%es
and fluctuations can 5eopardi2e the performance of the pattern classifier. The steady-state
load classifier is& in general& less affected by the temporal variation of the signal and adds
a redundancy and confidence in such a situation. +or a variety of =? @N events corrupted
with the harmonic noise and :DN& the load disaggregator misclassified three events
(C.C3 O misclassification rate).
04
The pattern matcher wor%s with an almost 2ero misclassification rate in a designed
environment& and can also be used in a highly noisy environment (both Daussian and non-
Daussian) although misclassification rate increases slightly. The development of the
transient pattern classifier based on the least suares error criterion and its parallel
placement with the steady-state load classifier have been very successful.
3.0 Chapter summary
! rich amount of information can be e.tracted from the non-intrusive power signal. The
power signal carries the fingerprint of the operational status of a load in terms of its
steady state power consumption and its transient response. >y systemically e.ploiting
those two properties of a load& its @N or @++ event can be identified and thus its
operational status can be obtained. The %nowledge of the statuses of the loads ma%e it
possible to disaggregate the total power observed into individual powers& in real-time.
The method of steady-state load classification utili2es the fact that a constant load
consumes a fi.ed amount of power in its steady state and that its steady state power is
distinctive in the harmonic power space. In this chapter& a simple steady-state load
classifier based on the @NE@++ two-state load model was developed. The power
consumed by a constant load was modeled by a Daussian random process. ! simple ML
classification rule was introduced and the probability space of a load event was derived.
The steady-state classifier generally wor%s well& but can perform poorly when there e.ist
pro.imal load classes and when there are many bac%ground load activities.
! transient pattern classifier based on least-suares error criterion was introduced to
augment the tas% of load identification. The method of transient pattern recognition
utili2es the source of information ignored by the method of steady-state load
classification and is less prone to the errors caused by the bac%ground load activities. The
transient pattern recognition can also have the problem of pro.imal load classes when the
loads have similar ratings& but is less affected because the transient pattern of a load has a
more discernibility than its steady state. !n arbitration rule was developed for the case
when two classifiers produce contradicting results.
The fact that electric loads are monitored from a single location& not individually&
complicates the 5ob of identifying their operational status the non-intrusive power signal.
>y utili2ing the information offered by a load both at its steady state and transient state&
the tas% of %eeping trac% of its operational status becomes more amenable and realistic.
The steady-state load classifier and the transient pattern classifier complement each other
for the common purpose of load identification.
Chapter .
,denti'ication o' Temporal 7#erlap o' Electric )oad E#ents #ia E#ent Space Search
and Transient Pattern Composition
This chapter describes the general method of identifying constant electric loads when
their @N andEor @++ events overlap in time. The identification of a single constant-load
event was fully discussed in the previous chapter. The preceding chapter developed the
methods of steady-state load classification and transient pattern recognition& to better
utili2e two different aspects of a loadJ steady-state response and transient responseJ for
the purpose of its identification. This chapter further develops the idea to identify
overlapping load events by the construction and search of the event space as a
combination of individual events and the composition and match of transient patterns.
..1 ,ntroduction
>ecause only a single set of voltage and current sensors is used to monitor tens of
different individual electric loads& the non-intrusive power signal is the (linear)
superposition of individual power signals. It is undoubtedly a challenging tas% to identify
the @NE@++ status of the loads from the observation of non-intrusive power signal alone.
+ortunately& many constant electric loads have uniue power consumption levels at their
steady state and furthermore have distinctive transient responses. (.ploiting these
fingerprints of loads in a systematic manner& the @N or @++ event of a load could be
identified reliably when load events are separated in time& as shown in the previous
chapter.
1owever& the identification fails when multiple load events occur sufficiently close in
time. +or e.ample& if two loads are turned on almost at the same time& resulting steady-
state power step change will differ from any step change of a load and their (combined)
transient response shape will not match any of single-load e.emplar patterns. Thus& when
a multi-event is detected& an appropriate event space needs to be created so that the
measured step-heights can be classified& and suitable transient patterns need to be formed
so that the observation can be matched.
>ecause the probability space of an individual event is %nown& the probability space of a
combination of individual events can be created by a combination of individual
probability spaces. Thus the multi-event identification& based on step-height
measurements& is simply the matter of forming a set of hypotheses and testing to choose
the best hypothesis. The set of feasible hypotheses is denoted as event space. 1owever&
because the number of possible combinations of individual load events is large& a
systematic and efficient algorithm is necessary to create and search the event space.
>y analogy& a transient pattern can be created for a hypothesis& as a combination of the
individual e.emplars specified by the hypothesis& and the most probable hypothesis can
be chosen from the set& which gives the best match against an observed response.
03
-onceptually& the transient pattern composition to identify a multi-event is no different
from the event space search. 1owever& because relative time-indices of individual
e.emplars are un%nown for a given hypothesis& the indices have to be found as a part of
the hypothesis test process. This inde.-search problem adds an enormous computational
burden. Thus& in reality& the transient pattern composition can not be implemented for
every hypothesis& but for a handful of best candidates.
..2 7#er#ie? o' multi!e#ent handling
>ecause the ma5ority of load events are single events& the identification of an event
should be performed with the assumption of a single event whenever possible to reduce
computational cost. ! multi-event handler will be called when the uni-event handler fails
or when the event is detected as a multi-event in first place.

change-
of-mean
detector
event
associatorE
pattern
e.tractor
uni-event
handler
multi-event
handler
single
event
multi-
event
success \
success \
no
yes
no
event unclassified
yes
load status
updater event
classified
+igure ?-/ )chematic of the overall load event classifier.
+igure ?-/ shows the schematic of the overall event classification scheme. The harmonic
power vectors (i.e. F !
=
!
=
LI) are generated at a power coding rate. The number of
vectors (or samples) is first reduced by a proper downsampling. #ownsampled power
vectors are inspected by the change-of-mean detector& which produces a series of
constant-mean segments from the stream of power vectors. The event associator inspects
the constant-mean segments to identify possible load events and to e.tract relevant
transient responses for the pattern matching. The event associator& in general& will be able
to detect whether the event of interest is a uni-event or multi-event. +or e.ample& if a
second @N transient is detected before a first @N transient is settled& the associator will
be able to recogni2e two transients as a multi-event. 1owever& the associator can not be
perfect and it will sometimes mis5udge a multi-event as a uni-event when its transients
happen sufficiently close in time.
1owever& a multi-event& first recogni2ed as a single event& will generate a very poor
classification result and reenter the multi-event handler for a second chance. If the
classification result is poor even by the multi-event handler& the event is declared
unclassified. (ach handler utili2es both the steady-state load classifier and the transient
pattern recogni2er and analy2es both outputs to classify an observed event. ;pon a
successful classification& the load status (@N or @++ for each load) is updated.
!s discussed before& because the transient pattern composition search ta%es an enormous
time& it is desirable to conduct the search for a handful of hypothesis candidates. Thus& the
multi-event handler first performs the event space search. ! few hypotheses are selected
from the search& which best account for the step height change observations. Then& the
transient pattern composition and matching is conducted for each selected hypothesis. !
separate arbitrator determines a best hypothesis from the results of both searches& based
on an appropriate arbitration rule.
..3 E#ent space search
)uppose a load event A@ R FA A! L I
T
is detected. A& A!& L are the increments of the
steady-state power levels due to the event. Now& we want to classify it by forming and
searching the event space& i.e. the set of event hypotheses based on the steady-state
behavior of individual loads.
:e assume that there are * electric loads of interest& of which . are on and N are off
before the event (* R . 9 N). :e define the event direction vector u as
. I / / / F I / / / F I F
3 /
T
N .
T
*
u u u
(
(

= =
_
.
_
. . u (?./)
Note that ui R J/ indicates that load i has a potential to change its status from @N to @++
during the event. )imilarly ui R / indicates an @++-to-@N change. 1owever& we still do
not %now what loads are participating the event and change their status as a result. :e
define the event status vector # R Fv
/
v
3
L v*I
T
& whose entry can be either / for change or
4 for non-change. Thus& the product uivi completely characteri2es the behavior of load i at
the event (/ for @++-to-@N& 4 for non-change and J/ for @N-to-@++). >ecause& there are
* loads& there e.ist 3
*
J / possible selections of #. In terms of hypothesis test& for a given
u&

. I / / / F 7
I 4 / / F 7
I 4 / 4 F 7
I 4 4 / F 7
/ 3
=
3
/
T
T
T
T
*
$
$
$
$
.
.
.
.
.
=
=
=
=

#
#
#
#
(?.3)
! simple conditional ma.imum li%elihood (ML) rule can be conceived for the test7
0?

). Y ( ma. arg
U
i
$
#
$ p $
i
@ A =
(?.=)
(ach load i is assumed to e.hibit a Daussian (normal) distribution at its steady state with
mean i and covariance matri. Ai& i.e. siFnI + N(i& Ai) where siFnI is the load vector of
load i at discrete time n. Thus& the event vector A@ is also Daussian& because it is a linear
combination of independent load vectors. Its probability density function ("#+) under $i
is given by the following formula7

& ) ( ) (
3
/
e.p
) ( det ) 3 (
/
) Y (
Y
/
Y Y
Y
3
/
3
|
.
|

\
|
A A = A
A

A A
A
i i i
i
l $ $
T
$
$
i
$ p
@ @ @
@
m @ G m @
G
@
t
(?.?)
where l is the length of A@. The mean and covariance are given by the following
formulae7

. ) Y ( &
& ) Y (
/
/ /
Y
/
Y
/ / /

=
=
+ = + =
=

*
#
# i # # d* d* $
*
#
# i # # $
$ v u
$ v u
j j j j j j i
i
G G G G G G
F m
@ @ @ @ I@
I@
(?.B)
d*j and d*j)/ are the numbers of samples for the event segment and the pre-event segment&
respectively.
/ j
@
G
is the sum of load covariance matrices that are on at the pre-event
segment. The original ML rule is converted to a simpler minimi2ation problem by ta%ing
a negative logarithm of the "#+7

( ). ) det( ln ) ( ) (
3 ln ) Y ( ln 3
& min arg
U
Y Y
/
Y Y
i i i i
i
$ $ $
T
$
i i
i
$
#
l $ p
$
I@ I@ I@ I@
G m @ G m @
@
+ A A =
A =
=

(?.0)
Now the hypothesis test is simply the matter of evaluating i for each hypothesis and
selecting a minimum. !lthough simple conceptually& this hypothesis test poses two
problems. The first is that we are interested in finding a few (e.g. three or five) best
candidates& not a single minimum& for further evaluations using transient patterns. The
second is that due to its combinatorial nature& no polynomial-time algorithm e.ists even
to find a single minimum
?
.
! systematic and efficient method is reuired to list and search every possible #.
!lthough many techniues can be conceived& the fact the entries of # are either / or 4
(binary) suggests us a structure similar to binary tree F-ormen et al.& 344/I as shown in
+igure ?-3.
?
This is a nondeterministic polynomial-time (N") optimi2ation problem with '(3
*
) comple.ity.

load .9/
load .
load /
load 3
/ 4
/
/
4 4
4
L
4 4
4
/
/
/
/
/
4 4
/
a search
path
L
L
L
load .9N
4 /
/
4 /
4
4 /
/
L
L
4 4 / /
4
/
4 /
4
root node
leaf nodes
+igure ?-3 >inary search tree of the possible event status vector #.
1ere& the tree has the height of * and each height level corresponds to a load. (ach node
has a left child designated as / and a right child designated as 4. The tree has 3
*
leaf
nodes and they represent all possible combinations of # (e.cept the rightmost node). The
specific value of # of a leaf node can be found by bac%trac%ing its ancestors. +or e.ample&
the second leftmost leaf node gives # R F/ / L / 4I
T
. 1owever& if we implicitly assume #
R F4 4 L 4I
T
at the root node& every possible # can be found by evaluating # at each /
node& which is circled in the figure. Thus& the left child of the root node gives # R F/ 4 L
4I
T
and its left child gives # R F/ / 4 L 4I
T
. :e also note that the right child of a node
gives the same # as its parent and thus does not need to be evaluated.
Thus& starting from the root node& by finding # and evaluating i at each / node& the set of
hypotheses can be searched e.haustively and the best candidates can be selected by
sorting the output. 1owever& this scheme reuires an e.haustive search of the tree and
sorting& which are very time consuming. ! heuristic method is often added to the tree
search to e.pedite the process. 1euristics is problem-specific information to guide the
search path& or to help decide which node is the best one to e.pand ne.t FNilsson& /<<CI.
!lthough many different types of heuristics can be conceived& for the given problem& we
define the (uclidean distance between the mean (of node n) and the event vector in
selected dimensions as a heuristic (evaluation) function f(n) of node n or& eually&
hypothesis $n7
.
T T
) (
Y
@ m
I@
A =
n
$
n f (?.A)
T (tilde) indicates that the distance is evaluated for a few selected dimensions (e.g. and
! only)& for the reason to be e.plained later. The heuristic function indicates the distance
between a current node and the nominal target A@& and can be used to steer the search
direction. In selecting a search direction& we nominally choose to evaluate the children of
00
a node first rather than its peers& a strategy that can be termed depth-first search
(compared to breadth)first search). Thus& a typical search path goes all the way down to
the leftmost leaf node by evaluating the left children of nodes first& comes bac% up to
evaluate the right child of an immediate parent node& etc. ! typical search path is shown
in +igure ?-3.
:e note that in this depth-first search scheme& #
T
# increases monotonically along a search
path from an intermediate node to its leftmost leaf node
B
. :e select the dimension of the
mean and event vector for the heuristic function as the dimension of a subspace of the
harmonic power space& in which the elements of every mean load vector i are positive.
Thus& if every load vector has a positive mean in its (real power)& the dimension of is
selected. If every load vector also has a positive mean in its ! (reactive power)& the
dimension of ! is selected in addition to
0
. The elements of
n
$ Y
T
I@
m
decrease
monotonically from height / to height .& and increase monotonically from height .9/ to
height .9N.

"
X
A@
n
$ Y
T
@
m
A
tra5ectory
of
n
f(n)
+igure ?-= Typical search tra5ectory and resulting behavior of the heuristic function.
+igure ?-= shows an e.emplary search tra5ectory on the -! plane and the resulting
behavior of the heuristic function& when the event vector A@ is located at the third
uadrant of the plane (a net-@++ event). The e.act shape of f(n) depends on the relative
location of A@ with respect to the tra5ectory. 1owever& we note that after the tra5ectory
ma%es a [;-turn6 at height .& f(n) becomes a conve. function& i.e. it reaches a minimum
in the vicinity of A@ and increases monotonically afterwards. Thus& with a threshold
radius rthres& it is unnecessary to continue search& once the tra5ectory leaves a circle with
the radius of rthres after the ;-turn.
B
The leftmost leaf node of a node is a leaf node located farthest left& which can be reached from it without
going bac% to its ancestors.
0
This is the case where all loads have inductive components (a lagging power factor or a positive reactive
factor).
:e designate the following pruning rule to eliminate a whole branch from a search7
"rune the branch from node n& if (f(n) > rthres) ] (V f(n) Z 4) ] (height > .9/). (?.C)
The second (gradient) condition is necessary to ensure that f(n) increases monotonically
and the third condition is necessary to apply the pruning rule after the ;-turn. In terms of
search efficiency& pruning should happen as early as possible. The e.amination of the
pruning rule reveals that pruning occurs early when . is small (. SS N) and when A@ is
located far away from the tra5ectory after the ;-turn. 1owever& the above pruning rule is
inefficient when . is large (. ZZ N). In this case& we may swap the event direction
vector u& such that
. I / / / F I / / / F I F
3 /
T
. N
T
*
u u u
(
(

= =
_
.
_
. . u (?.<)
!lso& the binary search tree is constructed with N loads to be turned on closer to root node
and . loads to be turned off closer to leaf nodes (N-to-. direction). In this case& the
search tra5ectory first travels to the first uadrant& ma%es a ;-turn and heads toward the
third uadrant (when the -! plane is selected as the evaluation dimension of the
heuristic function). In this case& the pruning rule is the same& e.cept that the third
condition is replaced by (height > N9/).
!n ambiguity can occur when . and N are relatively the same (. T N). (ither search
direction can be used& but we may choose the .-to-N direction when A S 4& assuming a
net-@++ event and the N-to-. direction when A Z 4& assuming a net-@N event.
In this manner& every node of a tree is visited and its heuristic function is evaluated. If f(n)
S rthres& the node (or hypothesis) is moved to a separate bin. If the node satisfies the
pruning rule& the entire branch stemming from the node is eliminated from the search.
!fter the search is complete& i is evaluated for each hypothesis in the bin and the results
are sorted to select a number of best minima.
! practical consideration is whether to generate a static tree at program initiali2ation or to
generate trees dynamically when needs arise. ! static tree is convenient when the number
of nodes is small and when its presence e.pedites a search process. 1owever& in our case&
load states change constantly& reuiring a different tree at each load event. !lso& there
really is no advantage in search speed by %eeping a static tree. Thus& it ma%es more sense
to create a tree dynamically when a load event occurs.
!lso& considering the fact that a node is completely identified by its event status vector #
and its height& it is desirable to create nodes during a search& rather than to create a whole
tree before the search. :e create a stac% that holds nodes to be searched. Initially the
stac% holds the root node (or euivalently # R F4 4 L 4I
T
and height R 4) only. +irst& the
root node is removed from the stac% and its two children are placed in the stac% (with the
left child on top of the right child). +rom then on& the top element of the stac% is removed&
0C
its heuristic function is evaluated and its two children are placed bac% into the stac%
unless it is a leaf node or a pruning has occurred. In this manner& every node is created
and tested until the stac% is empty. :e note that the ma.imum depth of the stac% is at
most *9/& and hence reuires a very small amount of memory.
>ecause the number of nodes is proportional to 3
*
& it is advantageous to reduce the
number of loads whose status changes are searched for. +or e.ample& if certain loads are
%nown to have turned on& turned off or remain unchanged& these loads are separated from
the set of total loads and the search is conducted among the rest
A
.
... Transient pattern composition
Let6s assume that m event status vectors (#6s) are selected from the event space search&
along with a single event direction vector u. :e are interested in selecting the best
hypothesis (or #) by creating a composite e.emplar from the combination of individual
e.emplars for each hypothesis& and comparing each composite e.emplar against an
observed transient pattern y to find the best match. :e assume the transient vector y is of
length N& i.e. | |
T
N y y y I / F I / F I 4 F = . y .
+or a given hypothesis& we assume that there are #
ma.
loads that either turn on or off. Let
/ j
n
represent the time inde. of e.emplar j& i.e. load j is turned on at time
/ j
n
. Let @j
represent the length-N e.emplar vector of load j. @j is filled with 2eros between inde. 4
and
/
/

j
n
& and the rest is the e.emplar of load j whose transient is referenced from 2ero
#- level. Then the composite e.emplar vector @ is simply the sum of individual
e.emplar vectors7
.
ma.
/

=
=
#
j
j
@ @ (?./4)
:e also ma%e y referenced to 2ero #- level& by subtracting the average of prevent
segment |& i.e.
| = y y
T
. The computation of | can be done easily because the location
of the first event is available from the change-of-mean detector.
:e want to minimi2e the distortion 0 between y
T
and @& which is defined as the second
norm of the error vector&
minimi2e .
T
3
@ y = 0 (?.//)
@bviously& the smaller is the distortion& the higher is the similarity between the pattern
and the e.emplar. The distortion is computed for each hypothesis and the hypothesis with
a minimum distortion is selected as a match.
A
The evaluations of the heuristic function and i should still ta%e into account of those separated loads.
1owever& the time indices of individual e.emplars need to be fi.ed to evaluate 0. :ith no
prior %nowledge of the time indices& they have to be found as a part of the minimi2ation
process by enumeration. Thus& the minimi2ation amounts to ) (
ma.
#
N ' comple.ity to
evaluate 0 for a given hypothesis. In most cases& we can ma%e an educated guess of inde.
bounds from a careful analysis of the change-of-mean detector output. 1owever& because
two different events may happen e.actly at the same time so that they can appear as a
single event& the inde. bound assessment has to be performed under worst-case scenarios.
If average inde. bound length is reduced to n (n S N) after the assessment& the comple.ity
can be reduced to ) (
ma.
#
n ' . 1owever& the comple.ity is still e.ponential in terms of #ma..
!t this point& one may wonder if we can use 1ynamic rogramming (#") to find the
minimum of 0. If applicable& the #" can reduce the e.ponential comple.ity of a problem
to a polynomial comple.ity. +or a minimi2ation problem to be #"-applicable& its cost
function should be represented in terms of states& the overall cost needs to be the sum of
additive costs which are defined by the states& and the selection of a current state must be
independent from the selection of previous states& e.cept its immediate previous state
(history independence or Mar%ovianity). +or our current problem of minimi2ing the
distortion& the time indices can be chosen as states& but because their relative locations are
un%nown the additive costs can not be defined.
Thus to ma%e the problem #"-applicable& we have to impose a certain constraint to the
time indices. )uppose the time indices are ordered& i.e.
j n n
j j
s
/
. Then the #th
additive cost can be defined7

& & & / & 4 I) F I F
T
( I / & F
ma.
/
3
/
/
# # n & n y n n
#
#
n
n n
# # #
. = = A

(?./3)
with n-/ R 4 and
N n
#
=
ma.
. Note that the #th additive cost is the distortion between
indices n#-/ and n# J / and that the non-2ero portion of e.emplar # starts at n#-/. The overall
cost is simply the sum of additive costs7
. I / & F
ma.
4
/
=

A =
#
#
# # #
n n 0 (?./=)
To find the minimum of 0& we can conceive #th intermediate optimal cost at inde. *7

. I / & F min I F
4
/
/ & 4
& & &
/
/ / 4

=

+ = =
A =

#
j
j j j
* n n
n n n
#
n n * I
#
#
.
(?./?)
To solve the problem using the #"& we need to convert the intermediate optimal cost to a
recursive form& which reuires the Mar%ovianity of the states. In other words& the
minimum additive cost of current state (n#-/) should be independent from the selection of
previous states (n#-3& L& n4). 1owever& a close e.amination of the cost model reveals that
A4
this condition is not satisfied. +or e.ample& a current additive cost changes depending on
the start inde. of its immediately preceding e.emplar& unless the e.emplar has a constant
value after the start inde.. Thus& the #" is not applicable in a strict sense. 1owever& if we
assume that the optimal locations of preceding e.emplars give a minimum current
additive cost& the #" can be used in an appro.imate sense. The rationale is that previous
best selections conforming to an observed pattern will give rise to the best selection of a
time inde. to add a new e.emplar. Now& we can write the intermediate optimal cost in a
recursive form7

( ). I & F I / F min I F
/ / /
/
* n n I * I
# # # #
n
#
#

A + =

(?./B)
The overall optimal cost is found when # R #
ma.
and * R N J /.
!s e.plained before& it is advantageous to limit the inde. bounds& if possible. Let nj
-
be
the lower bound and nj
9
be the upper bound of nj& i.e. nj
-
s nj s nj
9
. The upper and lower
bounds should increase monotonically& i.e. nj-/
-
s nj
-
and nj-/
9
s nj
9
. !lso& / s n
4
-
and
/
/
ma.
s
+

N n
#
. Now the minimum distortion&
I / F
ma.
N I
#
& for a given inde. order can be
found by the following #" algorithm7
/. Let # R 4 and I
J/
FI R 4. +or * R n
4
-
/& n
4
-
& L& n
4
9
/& compute
. I) F
T
( I & 4 F I F
4
3
/ 4 4
=

= = A =
*
n
n y * n * I
3. Let # R /. +or * R n#
-
/& n#
-
& L& n#
9
/& compute

( ). I & F I / F min I F
/ / /
) / & ( min
/ / /
* n n I * I
# # # #
* n n n
#
# # #

+ s s
A + =
+

!lso& for each *& determine

/
U
#
n at which I#F*I becomes minimum and call it n#-/F*I.
=. Let # # 9 / and iterate until # R #ma..
?. Let # R #ma.. +or * R N J /& compute

( ). I & F I / F min I F
/ / /
) / & ( min
/ / /
* n n I * I
# # # #
* n n n
#
# # #

+ s s
A + =
+

!lso& find
I / F U
/ /
ma. ma.
=

N n n
# #
& i.e. the start inde. of last segment.
@ptimal time indices are found by the following bac%ward recursion7

. I / U F U
I / U F U
I / F U
/ 4 4
/ 3 3
/ /
ma. ma. ma.
ma. ma.
=
=
=


n n n
n n n
N n n
# # #
# #
.
(?./0)
:e note that the time indices also need to be found during the e.ecution of the #"
algorithm to compute additive costs& i.e. to construct @.
In this manner& the minimum distortion for a given inde. order can be found efficiently.
:e define the normali2ed residue (of estimation) as the minimum distortion per sample7
. I / F
/
ma.
= N I
N
#
(?./A)
:e also define the correlation gain or shape factor (of estimation) oU & as the one which
gives a minimum distortion between
y
T
and @U o & where @U is the optimum @ found from
the #" algorithm.

. U
T
min arg U
3
@ y o o
o
=
(?./C)
oU is the solution of the standard least-suares (L)) problem& i.e.
.
T
U ) U U ( U
/
y @ @ @
T T
= o (?./<)
The above optimal cost was found by assuming an order in #ma. time indices. >ecause the
e.act order is usually un%nown& the optimal cost has to be found for every possible order
and an order that gives a minimum is to be selected. Thus& we need to run the #"
algorithm #ma.^ times for a given hypothesis to obtain the minimum distortion. Thus&
finding a minimum distortion for a given hypothesis using the #" algorithm has a
comple.ity on the order of ) ^ (
3
ma. ma.
N # # ' . -ompared to the original enumeration
scheme& the #" algorithm is advantageous whenever
3
ma. ma.
ma.
^

<
#
N # # . >ecause #
ma.
is
usually less than /4 and N is on the order of /444& the #" algorithm is almost always
more efficient. This remains true even if N is replaced by n& the average length of inde.
bounds& which is typically on the order of /44. Thus we substantially reduced the
computational cost of the original e.ponential comple.ity problem.
1owever& readers will remember that permutation grows very uic%ly. Table ?-/ shows
the permutations of a few selected numbers. Thus& it may be impossible to run the #"
algorithm if #ma. is very large& although the e.act upper bound will be determined by the
computational speed and load of an (LI) computer. #epending on test sites& this may or
may not be a problem. +or e.ample& if a site has a very low activity rate and event
overlaps rarely occur& the worst case scenario may not happen or even if it happens the
computer may have enough time to try every possible seuence. @n the other hand& if a
A3
site has a very high activity rate and event overlaps are common& based on computational
budget& the computer may have to declare a failure if #
ma.
turns out to be very large.
Table ?-/ "ermutations of a few selected numbers.
=^ 0
?^ 3?
B^ /34
0^ A34
A^ B4?4
C^ ?4=34
<^ =03CC4
If a failure is declared& the classification of a multi-event is either aborted or can be done
based on the event space search only. In latter& classification result is not very reliable and
it may need to be double chec%ed by a status estimation techniue to be developed later.
!nother option is to launch a separate process to do the e.emplar composition search.
The parent process puts the classification on hold until the result from the child process is
available& while continuing its nominal tas%s.
It also is possible to reduce the number of permutations by an educated guess of possible
seuences. +or e.ample& if three similar pumps and a drastically different fan compose a
hypothesis& we may first want to fi. the location of the fan (four possibilities) and then fi.
the order of the pumps (additional si. possibilities)& instead of trying all twenty-four
possible seuences. This is another e.ample of using problem-specific information to
reduce the si2e of search space. This and other computation saving techniues need to be
investigated in future.
!fter all the possible permutations of time indices are evaluated& the best seuence has to
be selected for a given hypothesis. The simplest is to choose a seuence that gives a
minimum normali2ed residue. This is in line with the philosophy of minimi2ing the
distortion between an observation and composite e.emplars. 1owever& when there are
large amounts of noise& (lower )N,)& this minimum normali2ed residue rule often loses
its discernability. In this case& a ma.imum composite gain rule may be proposed by
normali2ing the shape factor and residue& compute their product (composite gain) and
finding the ma.imum composite gain.
! practical consideration is how to generate all #
ma.
^ permutations when #
ma.
is given.
There e.ist several methods to produce them in a systematic manner. Le.icographic
ordering is most natural to humans& but is computationally e.pensive. !n alternative is to
produce a ne.t permutation by changing only two elements of current permutation
(transposition) FNi5enhuis and :ilf& /<ABG ,eingold et al.& /<AAI. >ecause the number of
total permutations is %nown& all the permutations can be generated seuentially by the
transposition starting from an arbitrary seuence& until a counter reaches #
ma.
^.
)o far& we have treated each e.emplar as having a single section. >ut in reality& an
e.emplar can have multiple sections. In terms of finding the best e.emplar seuence to
account for an observation& these additional sections can be treated as individual
e.emplars& e.cept that there are seuence constraints. +or e.ample& the second section of
an e.emplar can not be placed ahead of its first section. )uppose that rj (rj > /) is the
number of sections of load j. Then the total number of permutations increases to7

&
^ ^ ^
^
ma.
3 / #
r r r
r

(?.34)
where

=
=
ma.
/
#
j
j
r r . In the implementation viewpoint& r^ permutations are generated as
usual by the transposition and only those seuences that satisfy the seuence constraint
have chances to initiate the #" algorithm.
In this manner& the best inde. seuence is selected for a given hypothesis& along with &
oU and @U & to best account for an observation y. 1owever& because there are m hypotheses&
the procedure is repeated for each hypothesis& and the best hypothesis is selected based on
a criterion. :e note that i6s are also available from the event space search. There can be
many possibilities in designing the classification criterion based on the three different
measuresJ the normali2ed residue& shape factor and event space search test statistic.
To streamline the discussion& we normali2e the measures. Let oU = a & then the normali2ed
shape factor an is defined by the following normali2ation function7

.
/ if & /
/ 4 if &
4 if & 4
) (
>
s <
s
= =
a /a
a a
a
a f a
n
(?.3/)
Two other measures are normali2ed to the scale of an.

&
/
/

thres
thres
n
a
=
(?.33)
where athres is the threshold of an (4 S athres S /) and thres is the threshold of (thres Z 4).

otherwise&
/
/
4 /

thres
thres n
a

s
= (?.3=)
where thres is the threshold of (thres Z 4). Now we define four classification rules. The
first two rules utili2e the transient pattern composition result only and the remaining two
utili2es both results. :e assume that there are m hypotheses to choose ($
/
& $
3
& L& $m).
,ule /7
. ma. arg
U
Y
i
i
$ n
$
#
$ =
A?
,ule 37
. ma. arg
U
Y Y
i i
i
$ n $ n
$
#
a $ =
,ule =7
. ma. arg
U
Y Y Y
i i i
i
$ n $ n $ n
$
#
a $ =
,ule ?7
. ma. arg
U
Y Y
i i
i
$ n $ n
$
#
$ =
,ule / uses the event space search simply to find candidates for the transient pattern
composition and the best hypothesis is the one that gives a minimum residue. This
criterion wor%s well where noise is small but the event space search is relatively
unreliable (e.g. due to high bac%ground load activity). If noise becomes an issue& ,ule 3
can perform better than ,ule / because the shape factor is relatively immune to noise.
,ule = and ? ta%e into account the results of the event space search& and thus can be used
where the search yields very reliable results. #epending on noise content& either ,ule =
(high noise) or ,ule ? (low noise) can be used.
..D Test o' multi!e#ent handler
In this section& we test the multi-event handler with a few multi-events created from
single event data. The data were obtained from a test building in /04 )ansome )t.& )an
+rancisco& -!. The building has si. constant loads. Table ?-3 shows their list with their
identification numbers.
Table ?-3 Load names and identification numbers.
load id. no. load name
/ -hilled water pump / (-1:"/)
3 -ooling tower pump / (-T:"/)
= -ooling tower fan / (-T+an/)
? -hilled water pump 3 (-1:"3)
B -ooling tower pump 3 (-T:"3)
0 -ooling tower fan 3 (-T+an3)
:e also define a load identification vector
T
*
w w w I F
3 /
. = ? & whose individual entry
holds a load identification number. The load identification vector is to be used in
con5unction with the event direction vector u and the event status vector #& to characteri2e
the load state changes of an event.
+igure ?-? shows an e.ample of multi-event where three loads turn off almost
simultaneously. The original /34 12 data are downsampled by the C-point static
averaging and then fed to the change-of-mean detector. The figure shows the waveforms
of and ! after downsampling. >ecause three @N events are distinct& they are
recogni2ed as single events and correctly identified by the uni-event handler. 1owever&
the last @++ event is the combination of three individual @++ events and it triggers the
multi-event handler.

+igure ?-? ! multi-event where three loads turn off simultaneously.

4080 4100 4120 4140 4160 4180 4200 4220 4240
5
10
15
20
25
30
35
40
45
n, sample number (120 Hz)
P

(
k
W
)
516: alpha=1.003e+000, residue=3.518e-001, CTWP1 off; CTWP2 off; CTFan2 off;
+igure ?-B Matching result of a composite e.emplar for a three-load-off event.
:ith ? R F3 B 0 / = ?I
T
and u R F-/ -/ -/ / / /I
T
& the event space search yields a single # R
F/ / / 4 4 4I
T
with R -C./0CA. The search result suggests that there was no other state
combination within the threshold radius from the event vector (A@). >ecause there is only
one state combination candidate& the transient pattern composition search simply finds the
best time indices of the individual e.emplars to account for the observation. +igure ?-B
shows the result of the matching along with the shape factor and residue. The observation
is shown in a solid line and the composed e.emplars are shown in dots. The graph
suggests that the multi-event classification has been very successful.
A0

+igure ?-0 ! multi-event where two pumps turn on almost at the same time.
+igure ?-0 shows another e.ample of multi-event where two pumps turn on. >ecause two
events happen sufficiently close in time& the event associator recogni2es it as a single
event and sends it to the uni-event handler. 1owever& because the uni-event handler can
not find any match& it declares a failure and sends the event to the multi-event handler.
:ith m R =& ? R F/ 3 = ? B 0I
T
and u R F/ / / / / /I
T
& the event space search yields three
event status vectors as shown in Table ?-=. The actual event& -1:"/ and -1:"3 @N&
is the first candidate with the smallest . The remaining two candidates are other two-
pump-on events& which are perfectly legitimate.
Table ?-= Three event status vector candidates for the multi-event in +igure ?-0.
#
T

F/ 4 4 / 4 4I 3.0=4/e9443
F4 4 4 / / 4I /.CA33e944=
F/ 4 4 4 / 4I 3.BBBAe944=
Table ?-? Measures of the e.emplar composition search for the candidates in Table ?-=.
#
T
a

F/ 4 4 / 4 4I <.<///e-44/ 3.BABCe-44/
F4 4 4 / / 4I <.C?B3e-44/ =.B==/e9444
F/ 4 4 4 / 4I <.?030e-44/ A./?4=e9444
The multi-event handler performs the e.emplar composition search for these three
candidates& which returns the measures shown in Table ?-?. :ith ,ule / as the
classification rule& the multi-event handler correctly selects the first hypothesis& which has
the smallest residue. +igure ?-A shows the matching result.

400 450 500 550 600 650 700 750 800
10
15
20
25
30
35
40
45
50
n, sample number (120 Hz)
P

(
k
W
)
61: alpha=9.911e-001, residue=2.576e-001, CHWP1 on; CHWP2 on;
+igure ?-A Matching result of a composite e.emplar for two-pump-on event.
!lthough rare& it is possible to observe a mi.ed multi-event& i.e. certain loads turn on
while other loads are turned off almost at the same time. +igure ?-C shows an e.ample of
mi.ed event. The -T:"/ first turns on& and later the -T:"3 and -T+an3 turn on
almost at the same time. The -T:"/ turns off slightly after the -T+an3 is turned on.
Thus& in this mi.ed multi-event& two loads are turned on while a third load is turned off.

+igure ?-C ! multi-event where one load turns off while two loads turn on.
:ith m R =& ? R F/ = ? B 0 3I
T
and u R F/ / / / / -/I
T
& the event space search yields three
event status vectors as shown in Table ?-B. ,eaders will recogni2e that the actual event is
selected as the second candidate. The first candidate is simply the turn-on event of the
-T+an3. This e.ample suggests that the event space search can be very unreliable by
itself. The reason is that& depending on the nature of individual loads& state combinations
AC
can be densely populated in certain regions and& as a conseuence& a certain combination
may loo% more probable than others. !lso& the event space search is very vulnerable to a
bac%ground load activity. The e.emplar composition search is indispensable& although its
cost is e.tremely high.
Table ?-B Three event status vector candidates for the multi-event in +igure ?-C.
#
T

F4 4 4 4 / 4I A.00=/e9443
F4 4 4 / / /I /.=/BBe944?
F4 4 / 4 / /I /.A?<Ae944?
Table ?-0 shows the measures of the e.emplar composition search. :ith ,ule /& the
multi-event handler correctly selects the second hypothesis. +igure ?-< shows the
matching result. It is interesting to observe that the -T:"/ turns off in the middle of the
plateau of the -T+an3 @N event. This e.ample also shows the versatility and fle.ibility
of the e.emplar composition search based on the #" algorithm and the permutation& even
though the #" algorithm was applied in an appro.imate sense.
Table ?-0 Measures of the e.emplar composition search for the candidates in Table ?-B.
#
T
a

F4 4 4 4 / 4I /.4/C?e9444 =.=3<=e944/
F4 4 4 / / /I <.C3A3e-44/ =.=B/Ae9444
F4 4 / 4 / /I /.4//<e9444 0.AA3Ae9444

1500 1600 1700 1800 1900 2000 2100 2200 2300 2400
10
20
30
40
50
60
70
80
90
n, sample number (120 Hz)
P

(
k
W
)
204: alpha=9.827e-001, residue=3.352e+000, CTWP2 on; CTFan2 on; CTWP1 off;
+igure ?-< Matching result for two-load-on ] one-load-off event.
..E Chapter summary
! systematic method has been developed in this chapter to identify the @NE@++ status of
constant electric loads when their events overlap in time. ! set of hypotheses& i.e. which
loads turn on or off& is first proposed for a given event observation. The probability space
of a hypothesis is created as a linear combination of the probability spaces of individual
loads& and a few hypotheses that give largest probabilities are selected as candidates. !
binary-tree-li%e structure with heuristics was developed to generate and search the event
space efficiently.
The e.emplar composition search is conducted for the hypothesis candidates selected
from the event space search. ! composite e.emplar is formed as the sum of individual
e.emplars based on a hypothesis to best account for the transient response observation.
This is a non-deterministic polynomial-time optimi2ation problem& but its cost is
substantially reduced by the permutation of time-inde. seuences and the application of
dynamic programming in an appro.imate sense. The cost is further reduced by an
educated guess of inde. bounds. 1owever& the overall cost is still high due to the
permutation. The best hypothesis is selected according to a classification rule& based on
the measures from both searches.
Test results prove that the multi-event handler correctly classifies overlapping events& and
also demonstrate the need to conduct the two searches to identify a multi-event. The two
searches are connected in series for two reasons. +irst& because the e.emplar composition
search has an enormous comple.ity& it is imperative to conduct the event space search
first to confine the search space of an e.emplar composition. )econd& certain regions of
the state space can be so densely populated with possible state combinations that the
event space search may lose its discernability when an event vector is located in one of
those regions. !lso& the event space search is vulnerable to bac%ground load activity.
Thus& it is essential to double-chec% an event space search result with the e.emplar
composition search.
@ne practical concern is the time cost involved in the multi-event handling& especially
that of the e.emplar composition search. !lthough the e.act upper bound has to be
determined after the multi-event handler is implemented in real time& it is obvious that it
will not be able to function when tens of events overlap. !fter an upper bound is
determined& there are two possibilities depending on the nature of test sites. The first case
is when the ma.imum number of overlapping events rarely e.ceeds the upper bound. In
this case& the multi-event handler developed in this chapter can be employed without any
modification. The second& more general& case is when the ma.imum number of
overlapping events indeed e.ceeds the upper bound. In this case& further refinements of
the multi-event handling algorithm may be necessary& such as reducing the si2e of event
space by eliminating certain loads whose statuses are clearly %nown and reducing the
number of permutations in the e.emplar composition using heuristic information.
1owever& we have to admit that the multi-event handler sometimes can declare a failure
when the comple.ity of a given multi-event e.ceeds its capability.
The possibility of multi-event handler failure suggests that we can not rely solely on load
events to identify the status of loads. :e may have to utili2e the steady-state outputs of
loads. )tate estimation (based on steady-state outputs only) is necessary to identify the
initial status of loads when the program first starts and to ma%e a better decision when the
C4
edge estimation (based on event outputs) either fails or produces an unreliable result. :e
will ne.t study the state estimation problem.
Chapter D
Electric )oad State Estimation
This chapter describes the general method of identifying the @NE@++ status of constant
electric loads from the observation of steady-state power waveforms. ! set of hypotheses&
concerning the status of electric loads& is first proposed and& when conditions are met& the
observed segments of steady-state powers are used to determine the best hypothesis. The
state estimation& combined with the identification techniues based on load events&
provides a complete solution to identify the status of multiple constant electric loads from
the harmonic power waveforms generated at a single service entry.
D.1 ,ntroduction
It usually is a difficult tas% to identify the @NE@++ status of electric loads from the
observation of the non-intrusive power signal alone. +ortunately& many constant electric
loads have uniue power consumption levels at their steady states and furthermore have
distinctive transient responses. (.ploiting these fingerprints in a systematic manner& the
@N or @++ event of a load can be identified reliably. :hen multiple loads turn on or off
at the same time& the resulting event can still be identified by combining individual load
events to best emulate the observation& as shown in the previous chapter.
1owever& these approaches rely on the changes of load status& i.e. load events& to estimate
the status. In other words& if there is no load event& the load status can not be determined.
!n immediate problem is the initial status estimation& i.e. when an identification program
is first initiali2ed it does not %now what loads are on at that time. (ven after a few load
events are identified& the initial un#nowns remain un%nown. This may not be a problem&
when the load status is identified off-line by analy2ing a seuence of load events. >ut in a
real time situation& the initial load status needs to be %nown to ma%e timely decisions. @n
the other hand& when a load event is misclassified& there is no mechanism to correct the
error& and the program is prone to further misclassifications due to the erroneous status
information generated by the prior misclassification. Inter-event power signal has been
under-utili2ed& which may hold potential information to determine the load status.
:e introduce the concept of state as the combination of the @NE@++ status of individual
loads at a given instant& and edge as the boundary at which a state changes. The previous
approaches rely on edge outputs to determine states. In this chapter& we develop a state
estimation techniue based on state outputs. The state estimator provides a means to
identify the initial load status. It also can be used to augment the edge classifier by
correcting any misclassification error. !lso& depending on the circumstances& a separate
decision ma%ing authority may be necessary to ma%e a decision concerning the load status
based on both the state and the edge output.
C3
D.2 7#er#ie? o' load state estimation
The state estimator utili2es the harmonic power signal& generated when certain loads are
at their steady @N state while the others are completely @++& to determine their collective
@NE@++ status. Thus& the state estimator
C
compliments the edge classifier that relies on
edge outputs& and hence ma%es a better use of two different characteristics of loadsJ state
and edgeJ for the ob5ective of reliable real-time load status identification.

change-
of-mean
detector
event
associator
load status
updater
state
associator
event
classifier
state
estimator
arbitrator
+igure B-/ )chematic of the overall load status identifier.
+igure B-/ shows the overall load identification scheme& where two estimators are placed
in parallel. ! stream of harmonic power vectors is first e.amined by the change-of-mean
detector and is converted to multiple constant-mean segments. The event associator
e.amines the constant-mean segments to find a load event. ! load event typically
involves a transient that can span over several constant-mean segments. The event
associator assembles a load event from the analysis of those transient segments& and sends
it to the event classifier. !lso& the event associator mar%s those transient segments and
forward the information to the state associator.
The state associator assumes a distinct load state for a series of constant-mean segments
unmar%ed by the event associator. The state associator generates several state hypotheses
from the first segment of the series and the test metric for each hypothesis is updated with
the accumulation of the constant-mean segments. The series is terminated either when a
segment is mar%ed by the event associator or when a decision has to be made by the
arbitrator. ! new series is created immediately after a new unmar%ed segment appears.
Thus& a homogeneous state is assumed throughout a series and& at the end of the series&
state hypothesis candidates are available with their test metrics.
C
)trictly spea%ing& it should be called a state classifier& because it involves coming up with a finite number
of state hypotheses and choosing the best. (stimator is usually reserved for a decision ma%ing process&
where the number of possibilities is infinite or the variable of interest can ta%e a continuous value. 1owever&
in this thesis& we slightly abuse the terminology and a classifier is also called an estimator.
The arbitrator ma%es an intelligent decision about the load status& based on the outputs of
both the state and the edge estimator. The decision strategy may depend on the nature of
the monitored loads. +or e.ample& if the state estimation is fairly reliable after a certain
observation interval& it may ma%e a decision based on the state estimation output only. @n
the other hand& if the state estimation is not very reliable& it may not ma%e a decision until
a load event is observed. In this case& the event estimator performs event classification
repeatedly based on each state hypothesis and the hypothesis that best accounts for the
observed event is finally selected. Thus& the state estimation is coupled with the event
classification under this scenario.
The state estimator can also be used to chec% the event classification result. +or e.ample&
if the load status determined from an event classification is not found among the state
hypotheses generated by the state estimator after the event& the arbitrator can determine
that there was an error in the previous event classification. The arbitrator may choose
another possibility proposed by the prior event classification or simply designate the
current status as un#nown and enter the initiali2ation mode.
D.3 State estimation
:e assume that there are m constant electric loads to be monitored by the (LI). In
addition to those m loads& we assume the presence of a bac%ground load. The bac%ground
load can be either a collection of small physical loads& such as lighting or plug loads&
which tap currents from the main trun% where the current sensor is located& or simply the
#- offsets present in harmonic power vectors even when those m loads are all off. :e
define a load status vector& # R Fv4 v/ v3 L vmI
T
& whose ith entry is either / for the @N
status or 4 for the @++ status of load i. The 2eroth entry v4 represents the status of the
bac%ground load& which is always /.
+or a constant mean segment of length N& there e.ist N harmonic power vectors& each of
length l7
| | . / 4 I F I F I F s s = N n n ! n n
T
. @ (B./)
The dimension of the power vector is determined by the nature of the constant loads.
Typically& only the real and reactive powers are used for the status identification and
hence l R 3. 1owever& other harmonic powers& such as the third& can be used if all the
loads generate constant amounts of those harmonic powers when they are on.
(ach load i is assumed to e.hibit a Daussian (normal) distribution at its steady state with
mean i and covariance matri. Ai& i.e. siFnI + N(i& Ai)& where siFnI is the load vector of
load i at discrete time n. :e assume that the load status remains unchanged for the
duration of the segment. Thus& the power vector can be represented as a linear
combination of the load vectors that are on.
C?
. I F I F
4

=
=
m
i
i i
n v n s @ (B.3)
>ecause each load is statistically independent& the power vector also follows a Daussian
distribution7
. & T I F
4 4
|
.
|

\
|
= =

= =
m
i
i i
m
i
i i
v v N n G G F F @
@ @
(B.=)
:e organi2e the N power vectors into a single large lN/ data vector y7

| | . I / F I / F I 4 F
T
T T T
N = @ @ @ y .
(B.?)
This type of arrangement is called temporal ordering or column rollout in array signal
processing F$ay& /<<CI. >ecause each power vector is Daussian and statistically
independent& the data vector y is also Daussian with&
mean7 | |
T
T T T
@ @ @ y
F F F m . = (B.B)
covariance7
(
(
(
(

=
@
@
@
y
G
G
G
G

. (B.0)
Note that the covariance matri. is a bloc% diagonal matri. with 2ero off-diagonal bloc%
entries.
Now the load status estimation& given the length-N constant-mean segment& is simply the
matter of determining # among 3
m
possible alternatives. In the form of a hypothesis test&

. I / / / / F 7
I 4 4 / / / F 7
I 4 4 / 4 / F 7
I 4 4 / / F 7
I 4 4 4 / F 7
/ 3
=
3
/
4
T
T
T
T
T
m
$
$
$
$
$
.
.
.
.
.
.
=
=
=
=
=

#
#
#
#
#
(B.A)
! simple conditional ma.imum li%elihood (ML) rule can be conceived for the test7

). Y ( ma. arg
U
i
$
#
$ p $
i
y =
(B.C)
The probability density function ("#+) under $i is given by the following formula7

. ) ( ) (
3
/
e.p
) ( det ) 3 (
/
) Y (
Y
/
Y Y
Y
3
/
3
|
.
|

\
|
=

i i i
i
lN $ $
T
$
$
i
$ p
y y y
y
m y G m y
G
y
t
(B.<)
>ecause A
y
is bloc% diagonal& the conditional "#+ can be simplified by using the
following relationships7

N
N
) (det det det
/
4
@ @ y
G G G = H =

and
(
(
(
(
(

/
/
/
/
@
@
@
y
G
G
G
G

. (B./4)
Now& the "#+ becomes7

. ) I F ( ) I F (
3
/
e.p
) (det ) 3 (
/
) Y (
/
4
Y
/
Y Y
Y
3 3
|
.
|

\
|
=

N
n
$ $
T
$
$
i
i i i
N
i
lN
n n $ p
@ @ @
@
F @ G F @
G
y
t
(B.//)
The original ML rule is converted to a simpler minimi2ation problem by ta%ing a negative
logarithm of the "#+7

( ) . ) I F ( ) I F ( det ln 3 ln
) Y ( ln 3
& min arg
U
/
4
Y
/
Y Y Y

+ + =
=
=
N
n
$ $
T
$ $
i i
i
$
#
i i i i
i
n n N lN
$ p
$
@ @ @ @
F @ G F @ G
y
t

(B./3)
There can be two scenarios of the hypothesis test& depending on the nature of the
bac%ground load. If the bac%ground load has a fi.ed and %nown mean& the above ML rule
can be used without any modification. 1owever& in a more general case& the bac%ground
load can have a variable mean. Ket& it is assumed to be constant for a given segment. In
this case& the bac%ground mean has to be estimated as a part of the hypothesis test process
and thus the above ML rule needs to be modified.
:e first e.amine the fi.ed bac%ground mean case. To conduct an effective search among
3
m
possible hypotheses& we use the depth-first search binary tree introduced in the
previous chapter. +igure B-3 shows the schematic of the binary tree. The load status
vector # is assumed to be F/ 4 4 L 4I
T
at the root node. !t height i& the ith entry of #
C0
changes its status at / nodes and remains unchanged at 4 nodes. In this manner& every
possible # is generated from the binary tree.

load /
load 3
/ 4
/ / 4 4
4
L
4 4
4
/
/
/
/
/ 4 4 /
a search
path
L
L
L
load m
4 /
/
4 /
4
4 /
/
L
L
4 4 / /
4 /
4 /
4
root node
leaf nodes
+igure B-3 >inary search tree for the possible load status vector #.
The (uclidean distance between the mean and the segment average is selected as a
heuristic function to prune unnecessary branches. +or a node n&

avg $
n
n f @ F
@
T T
) (
Y
= & (B./=)
where @avg is the segment average of the N-point power vectors& i.e.

=
=
/
4
/
I F
N
i
N avg
i @ @ .
Tilde indicates that the dimension used for the evaluation of the heuristic function is the
dimension of the subspace of the harmonic power space& in which the elements of every
mean load vector i are positive. Now& it is easy to show that f(n) is a conve. function and
we can conceive the following pruning rule7
"rune the entire branch from node n& if (f(n) > rthres) ] (V f(n) Z 4)& (B./?)
where rthres is the threshold radius. "runing occurs early when the minimum of the
heuristic function is located close to the root node. Thus the above search direction is
efficient when
n
n
$
$
avg
a
Y
T
ma.
T
@
F @ s
& where a is a fraction. If
n
n
$
$
avg
a
Y
T
ma.
T
@
F @ >
& we
may swap the search direction by assigning # R F/ / / L /I
T
at the root node. The
heuristic function and the pruning rule remain unchanged in this case.
In this manner& each node of the tree is visited and its heuristic function is evaluated. If f
(n) S rthres& a copy of the node (or hypothesis) is moved to a separate candidate bin. If a
node satisfies the pruning rule& the entire branch stemming from the node is eliminated
from the search. !fter the search is complete& the test statistic i is evaluated for each
hypothesis in the bin.
The tree search is conducted once for the first segment of a common load status series.
+or its subseuent constant-mean segments& 5ust the test statistic is evaluated and
recorded for each hypothesis in the candidate bin. If there are . segments for a given
series& the overall test statistic i for $i is simply the weighted sum of individual segment
test statistics7
&
/
/
&

=
=
=
.
#
#
.
#
# i #
i
N
N
(B./B)
where N# is the length and i2# is the test statistic of the #th segment.
In the fi.ed bac%ground load mean case& the state estimation is fairly reliable if an
observed series is long enough. Thus& the arbitrator may ma%e an appropriate decision
based on the overall test statistics in the bin and the accumulated segment length& i.e. it
selects a hypothesis with a minimum test statistic when the accumulated length e.ceeds a
certain threshold.
:e ne.t e.amine the case when the bac%ground load has a varying mean& which
commonly occurs at real buildings where small au.iliary loads are connected to the panel
designed for the ma5or (18!-) loads. +igure B-= shows an e.ample of the bac%ground
load activity when all ma5or loads are off. The data were ta%en at a test building in )an
+rancisco& -!. There are many small @NE@++ activities in the bac%ground. 1owever
their step heights are too small (T/ %:)& compared to other ma5or loads (T/4 %:)& and
their collective @NE@++ activities are too freuent to be %ept trac% of reliably.

80 100 120 140 160 180 200 220 240
8
9
10
11
12
13
14
15
16
17
18
time (sec)
P

(
k
W
)
Background load activity
+igure B-= !n e.ample of bac%ground load activity.
CC
1owever& it is reasonable to assume that the bac%ground load has a constant mean for a
short duration of time. The changes of the bac%ground mean can normally be captured by
the change-of-mean detector& e.cept when the changes are mas%ed by the larger transients
of the ma5or loads. :e also assume that there are upper and lower bounds in the value the
bac%ground mean can ta%e and that those bounds are %nown. :e further assume that the
bac%ground load has a constant covariance matri.& even though its mean changes with
time.
>ecause the bac%ground mean
4
is un%nown& it has to be estimated as a part of the
hypothesis test procedure for a given constant-mean segment. :e assume that
4
is a
deterministic but un%nown uantity with a certain set of %nown bounds. The conditional
"#+ is rewritten with
4
as an un%nown parameter7
. ) I F ( ) I F (
3
/
e.p
) (det ) 3 (
/
) Y G (
/
4
/
Y 4
/
Y
/
Y 4
Y
4
3 3
|
.
|

\
|
=

=
=

=
N
n
m
#
# $ # $
T
m
#
# $ #
$
i
i i i
N
i
lN
v n v n $ p F F @ G F F @
G
F y
@
@
t
(B./0)
:e use the generali2ed ML rule to account for the additional uncertainty involved in
estimating 4 F$ay& /<<CI7

( )& ) U ( det ln ) Y U G ( ln 3
& min arg
U
& 4 & 4 i i i i
i
$
#
$ p
$
i
F , F y
y
+ =
=

(B./A)
where i & 4
U F
is the ML estimation of
4
assuming $i is true. ,
y
(
4
) is the +isher information
matri. F"apoulis and "illai& 3443I assuming $i is true. It is straightforward to show that
. I F
/
U
/
Y
/
4
& 4
=

=
=
m
#
# $ #
N
n
i
i
v n
N
F @ F (B./C)
The evaluation of the +isher information matri. is rather involved. :e simply state the
result7
.
) G ( ln ) G ( ln
) (
/
Y
& 4
& 4
& 4
& 4
& 4

=
|
|
.
|

\
|
|
|
.
|

\
|
c
c
|
|
.
|

\
|
c
c
=
i
$
i
i
T
i
i
i
N
p p
(
@ y
G
F
F y
F
F y
F , (B./<)
If we substitute these two relationships into the generali2ed ML rule& after a few algebraic
steps& the test statistic i becomes7
. det ln ln ) I F ( ) I F ( det ln 3 ln
Y
/
4
/
Y Y
i i i
$
N
n
avg $
T
avg $ i
N l n n N lN
@ @ @
G @ @ G @ @ G + + + =

t (B.34)
-ompared to the test statistic of the constant bac%ground mean case& the sum of the load
means under a hypothesis is replaced by the segment average. The last two terms are
penalty terms introduced by the +isher information matri.. >ecause the fourth term is
common among the hypotheses given a segment& only the fifth term is a real penalty
factor in evaluating different hypotheses. :e also note that the influence of the penalty
term diminishes with increasing N.
The above test statistic has an interesting feature. >ecause
i
$ Y @
F
is replaced by @avg& the
sum of the load means under $i plays no role in determining the test statistic. !ny
difference between the segment average and the sum of the individual load means under a
hypothesis is attributed to the bac%ground load. @f course& because the bounds for the
bac%ground mean are %nown& certain hypotheses that give out-of-bound bac%ground
mean estimations can be eliminated from the consideration. Then& it is the covariance
matri. that allows distinguishing different hypotheses. The above generali2ed ML rule
has more resolving power when loads have more distinctive covariance matrices.
1owever& when loads show similar noise patterns& the test may not be reliable.
The hypothesis candidates can be found by using the binary tree shown in +igure B-3.
Let6s assume the bounds for the bac%ground mean are %nown for the real power ()7
&
4 4 4
+
s s

(B.3/)
where
4

is the real power component of the bac%ground mean vector 4. +or the
heuristic function& we select7
& ) ( U ) (
4
n n f

= (B.33)
i.e. the real power component of the ML estimation of the bac%ground mean at node n. If

=
s
m
i
i avg
a
/
I / F @ & where @avgF/I is the real power component of the segment average&
i
is the real power component of the mean vector of load i and a is a fraction& the load
status vector # is set to F/ 4 4 L 4I
T
at the root node. :ith this search direction& it is easy
to show that the heuristic function decreases monotonically along a branch. :e designate
the following pruning rule7
"rune the entire branch from node n& if

n f
4
) ( . (B.3=)
If

=
>
m
i
i avg
a
/
I / F @ & we set # R F/ / / L /I
T
at the root node. In this case& the heuristic
function increases monotonically. The pruning rule changes to7
"rune the entire branch from node n& if
+
>

n f
4
) ( . (B.3?)
<4
@ther harmonic powers& such as the reactive power& can be 5ointly used for pruning& if the
bac%ground mean bounds are %nown and all loads have positive means for the harmonic
powers of interest.
In this manner& each node of the tree is visited and its heuristic function is evaluated. If
the heuristic function& the ML estimation of the bac%ground mean& satisfies the bounds& a
copy of the node is moved to a separate candidate bin. !s before& the tree search is
conducted once for the first segment of a common load status series& and 5ust the test
statistic is evaluated for each hypothesis in the bin in subseuent constant-mean
segments. :hen the series is terminated& the overall test statistic for each hypothesis is
computed as a weighted sum using (B./B).
!s discussed before& because the sum of the load means under a hypothesis does not play
any role in determining the best hypothesis& the state estimation in the varying
bac%ground mean case is not reliable in most cases& e.cept when the loads have distinct
covariance matrices. Thus& it is ris%y for the arbitrator to ma%e a decision based on the
state estimator output only. @ne of the viable strategies in this case is to couple the state
and the edge estimator to ma%e a more intelligent decision. +or e.ample& the edge
estimator may perform its estimation based on the state hypotheses provided by the state
estimator. @n the other hand& the result of the edge estimation may be chec%ed against the
state hypotheses generated by the state estimator. :e propose two arbitration strategiesJ
when the load status is un%nown and when it is %nown.
:hen the load identification program is first started& the load status is un%nown and the
program is in an initiali2ation mode. The state estimator first conducts the binary tree
search to find state hypothesis candidates. Their test statistics are updated with the
accumulation of the constant-mean segments& and the process is terminated when a load
event is observed. :e designate a threshold width& A > 4& for the test statistics of the
state hypotheses. The hypotheses& which have their test statistics within A from the
minimum test statistic& are selected as valid ones to be forwarded to the edge estimator.
The edge estimator conducts its test to identify the observed load event& assuming a pre-
event load status for each forwarded state hypothesis. !fter the test& the hypothesis which
gives rise to the best event classification result is selected. !t this point& the load event is
identified. The arbitrator further e.amines if other state hypotheses yielded the same load
event classification result. If so& the state hypothesis with a minimum test statistic among
those hypotheses is selected as the state of the pre-event. @therwise& the original state
hypothesis is retained. Thus& the load status prior to the load event is also resolved after
the load event identification. The initiali2ation mode is released after the successful edge
event classification.
1owever& the edge estimation sometimes fails. +or e.ample& no valid state hypothesis
may e.ist at the time of edge estimation& no state hypothesis may give rise to a reliable
edge classification result& or if the computational comple.ity of particular event
identification is too high the edge estimator may voluntarily declare a failure. In this case&
the arbitrator decides not to ma%e a decision and simply enters the initiali2ation mode&
starting to generate new state hypotheses and waiting for a ne.t decisive edge event.
@nce the initiali2ation mode is released& the load status is %nown and the edge estimator
can perform its duty without the assistance of the state estimator. Ideally& subseuent load
status identifications can be performed solely by the edge estimator. 1owever& in a real
life situation& the edge estimator may fail or misclassify a load event. If the edge estimator
fails& the arbitrator simply enters the initiali2ation mode. The conseuence of an edge
event misclassification can be subtle. #epending on the circumstances& a misclassification
happens once and it may not greatly affect the outcomes of future edge classifications. @n
the other hand& it can lead to a series of future misclassifications.
Thus& it is necessary for the arbitrator to have a list of legal state hypotheses& even when
the current load status is %nown after a seemingly successful edge classification. The legal
hypotheses are those who have the test statistics within the threshold width from the
minimum test statistic& among the state hypothesis candidates generated by the state
estimator. @nce the accumulated constant mean segment length e.ceeds a certain
threshold& the arbitrator performs a test to verify if the current #nown status is among the
legal hypotheses. If so& the previous edge estimation result is confirmed and no further
action is necessary. If not& the arbitrator searches the list of edge event classification
candidates& archived by the edge estimator after the previous edge estimation. The
arbitrator searches from the most favorable to the least favorable candidate& and e.amines
if the resulting load status& assuming the candidate were true& is among the legal state
hypothesis. If so& the candidate is selected& the previous edge estimation is nullified& the
current load status is updated and the search is terminated. If the search terminates
without producing a result& the arbitrator sets the initiali2ation mode.
D.. Test o' load state estimator
In this section& we test the state estimator. The test data are obtained from a test building
at /04 )ansome )t.& )an +rancisco& -!. The building has si. constant loads. Table B-/
shows the list of loads with the load identification numbers. The building has a varying
bac%ground load mean& and the test program is written accordingly following the strategy
outlined in the previous section. In that building& the bac%ground mean has the following
bounds7

4
R C %: and
+
4
R /?.B %:.
Table B-/ Load names and identification numbers.
load id. no. load name
/ -hilled water pump / (-1:"/)
3 -ooling tower pump / (-T:"/)
= -ooling tower fan / (-T+an/)
? -hilled water pump 3 (-1:"3)
B -ooling tower pump 3 (-T:"3)
0 -ooling tower fan 3 (-T+an3)
+igure B-? shows an e.ample of load events& where the -T:"3 is turned on and then
turned off in about twenty seconds later. The original /34 12 data are downsampled by
<3
eight by ta%ing static averages before the analysis. >ac%ground load activities in the
beginning and around the -T:"3 @++ event are notable. !ll other ma5or loads are off
before the events. 1owever& because the program has no %nowledge about the load status
in the beginning& it starts to generate state hypotheses and the hypotheses are evaluated at
the first event.

+igure B-? -T:"3 on and off events with un%nown initial status.
Table B-3 shows the state hypotheses forwarded to the edge estimator when the first event
is evaluated. The table shows that only a single state hypothesis (all loads off& e.cept the
bac%ground load) was generated by the state estimator before the event.
Table B-3 )tate hypotheses before the first event in +igure B-?.
#
T

F/ 4 4 4 4 4 4I -/.4?3Ae944=
Table B-= List of post-event load status generated by the edge estimator after the first
event in +igure B-?.
#
T
F/ 4 4 4 4 / 4I
F/ 4 / 4 4 4 4I
F/ / 4 4 4 4 4I
F/ 4 4 4 / 4 4I
Table B-= shows the list of possible post-event load statuses generated by the edge
estimator after the successful identification of the first event. The most probable status is
listed on the top and the arbitrator accepts the recommendation. Now the pre-event load
status is identified (no load @N) and the current load status is identified (only -T:"3
@N).
1owever& the arbitrator still has to chec% if the event classification result gives one of
legal state hypotheses. !fter <3C points (at /34 12) are observed& the arbitrator runs a
state chec%. The legal state hypotheses available at the time provided by the state
estimator are shown in Table B-?. The status resulted from the previous edge estimation is
one of those legal hypotheses& although it is not the one with the minimum test statistic.
Thus& the arbitrator confirms the previous edge estimation result and the current load
status is completely determined.
Table B-? Legal state hypotheses after the evaluation of the first event in +igure B-?.
#
T

F/ / 4 4 4 4 4I -<.=<<3e9443
F/ 4 4 4 / 4 4I -<.=<BAe9443
F/ 4 4 4 4 / 4I -<.30/?e9443
Table B-B shows the load status history generated in the end by the program. It specifies
the start and end inde. (at /B 12) and the @NE@++ status of loads at each constant load
status series. >y default& the load status is not assigned during transients.
Table B-B Load status history for the events in +igure B-?.
start inde. end inde. -1:"/ -T:"/ -T+an/ -1:"3 -T:"3 -T+an3
=B /=B @++ @++ @++ @++ @++ @++
/?A ?BB @++ @++ @++ @++ @N @++
?BA @++ @++ @++ @++ @++ @++

+igure B-B -T:"3 events with -T:"3 @N in the beginning.
+igure B-B shows another -T:"3 @NE@++ event set. 1owever& in this case& the program
is started when the -T:"3 is already on. The first event is the -T:"3 @++ event. Table
B-0 shows the state hypotheses available at the time of the first event classification.
<?
>ecause four pumps have similar mean levels and covariance matrices& each one of them
shows up as the only load in each valid state hypothesis. !lthough the true hypothesis
(fourth one) has the minimum test statistic& other hypotheses in the table also have test
statistics close to the minimum. This e.ample illustrates the difficulty in performing a
state estimation based on state outputs only and the necessity to ma%e a decision in
con5unction with an edge event and its classification.
Table B-0 )tate hypotheses before the first event in +igure B-B.
#
T

F/ / 4 4 4 4 4I -/.4CC4e944=
F/ 4 / 4 4 4 4I -/.40CAe944=
F/ 4 4 4 / 4 4I -/.4CC3e944=
F/ 4 4 4 4 / 4I -/.4CCBe944=
Table B-A Load status history for the events in +igure B-B.
start inde. (nd inde. -1:"/ -T:"/ -T+an/ -1:"3 -T:"3 -T+an3
?/ 330 @++ @++ @++ @++ @N @++
33C ?=B @++ @++ @++ @++ @++ @++
?0C ABB @++ @++ @++ @++ @N @++
ABA @++ @++ @++ @++ @++ @++
The edge estimator correctly identifies the -T:"3 @++ event based on the fourth
hypothesis in the table and the arbitrator accepts the result. )ubseuent edge event
identifications can be performed with %nown load statuses. Table B-A shows the resulting
load status history.

+igure B-0 -T+an/ events with -T:"/ @N in the beginning.
+igure B-0 shows another e.ample of load events& where the -T:"/ is already on in the
beginning and the -T+an/ is turned on and off on top of it. Table B-C shows the state
hypothesis candidates at the time of the first event evaluation. The situation is similar to
Table B-0& where the true hypothesis (second one) has the smallest test statistic but with a
very narrow margin.
Table B-C )tate hypotheses before the first event in +igure B-0.
#
T

F/ / 4 4 4 4 4I -<.A03Ae9443
F/ 4 / 4 4 4 4I -<.</<=e9443
F/ 4 4 4 / 4 4I -<.0BC=e9443
F/ 4 4 4 4 / 4I -<.C?33e9443
Table B-< Load status history for the events in +igure B-0.
start inde. (nd inde. -1:"/ -T:"/ -T+an/ -1:"3 -T:"3 -T+an3
?/ 0/C @++ @N @++ @++ @++ @++
0A? /BBC @++ @N @N @++ @++ @++
/B04 /03= @++ @N @++ @++ @++ @++
/03B @++ @++ @++ @++ @++ @++
In this case& every state hypothesis gives rise to the same load event of -T+an/ @N and&
as a result& the state hypothesis with the minimum test statistic is selected as a valid pre-
event load status. Table B-< shows the resulting load status history.
D.D Chapter summary
! systematic method has been developed in this chapter to estimate the @NE@++ status of
constant electric loads based on the observation of the harmonic power waveforms under
a uniform load status. ! list of state hypotheses is generated via the binary tree search in
the beginning of a uniform load status series. Their ML test statistics are evaluated and
updated with the accumulation of subseuent constant-mean segments. If the bac%ground
load has a fi.ed mean& the hypotheses can be evaluated to identify the current load status&
once the accumulated observation length e.ceeds a certain threshold.
1owever& in a more general situation& the bac%ground load can have a varying mean&
assumed to be constant over a constant-mean segment. In this case& the arbitrator does not
ma%e a decision until a load event occurs. The edge estimator classifies the observed load
event repeatedly based on each state hypothesis forwarded by the state estimator. The
arbitrator selects the state hypothesis which gives rise to the best event classification
result. The state is estimated in con5unction with a following load event.
@nce a load event is successfully identified& the edge estimator performs the status
estimation. 1owever& the arbitrator occasionally runs a state chec% to verify that the result
of a previous edge estimation gives one of current legal state hypotheses generated by the
state estimator. If so& the previous edge estimation is confirmed. If not& the result of the
<0
previous edge estimation is searched to find a viable candidate. If a candidate is found&
the current load status is updated. If not& the arbitrator sets the initiali2ation mode.
The overall load identification program is written and tested with the data collected in a
real building& which has a variable bac%ground mean. The test result is uite satisfactory
and shows that the overall algorithm is accurate and very robust.
"revious load identification approaches have relied on the limited aspects of edge
estimation only. They were prone to freuent misclassifications& were vulnerable to noise
spi%es and disturbances& and lac%ed the capability to perform the identification in real
time. The approach presented in this chapter utili2es both aspects of the power
waveformsJ state and edgeJ to better identify load status. The resulting algorithm is more
accurate& reliable and robust& because it can tolerate the mista%es and failures by the edge
estimator. It also does not reuire any prior load status information in beginning. It
provides a complete solution to identify the status of constant electric loads in real time.
Chapter E
Estimation o' ;aria$le Speed Dri#e Po?er Consumption 'rom 6igher 6armonic
Po?ers
It is a difficult problem to estimate the real and reactive powers consumed by variable
speed drives
<
(8)#s) from the non-intrusive power signal& especially in the presence of
other constant electric loads. 8ariable speed drives continuously change their power
consumption and thus do not have clearly defined steady states. +urthermore& their time-
varying nature can interfere with the load status identifier& designed to %eep trac% of the
@NE@++ status of constant loads based on their steady-state power consumption levels.
1ence& it is necessary to estimate the power consumed by 8)#s& not only to trace their
power usage but also to guarantee the functionality of the constant load status identifier.
Li%e any good detection and estimation routine in signal processing& the 8)# power
estimation has to rely on a certain signal model that best represents the characteristic of
8)#s. In this chapter& a 8)# power estimation method is proposed& which relies on the
measurement of higher harmonic powers generated by a 8)# and its distinctive
relationship between the harmonic and fundamental powers. +or simplicity& it is assumed
that a single set of 8)#s is present along with other constant loads.
E.1 ;SD load model
8ariable speed drives are widely used !- motor drives& especially in industrial and
commercial applications& where single-speed operations can not meet the demands of
various physical loads. Machine tools& fans& pumps and chillers are typical applications of
8)#s. ! 8)# is an efficient way to control a motor when the connected load reuires a
wide range of speed or torue ad5ustment. ! 8)# typically consists of a rectifier& a direct
current bus& and an inverter F>ose& /<<=I. +igure 0-/ shows the topology of a 8)#& along
with a typical circuit model. Three-phase !- currents are converted to a #- via a
rectifier and the #- current is inverted to !- currents (though may not be sinusoidal) by
an inverter. The output freuency is ad5usted by controlling the inverter timing.
The three-phase rectifier forms a bridge with si. diodes as shown in +igure 0-/. The load
is fed via a three-phase half-wave connection& and the return current path is another half
wave connection to one of the three supply lines FLander& /<CAI. The output voltage is
further rectified typically by an *3 filter& before being fed to an inverter. -ho%es
(inductors) can be placed between the supply lines and the diodes to suppress the
propagation of harmonics bac% to transmission lines.
!t any moment& two diodes are onJ one in the supply path and the other in the return
path. +or e.ample& when the ! phase supply diode is on& either the > or the - phase
return diode is on. )imilarly& the ! phase return diode is on when either the > or the -
<
8ariable speed drives are also called variable freuency drives (8+#s) and ad5ustable speed drives
(!)#s). :e will use them interchangeably.
<C
phase supply diode is on. !s a result& the ! (> or -) phase current typically shows four
pea%sJ two in a positive swing and two in a negative swingJ in a single cycle FNu et al.&
/<<<I.

,ectifier
#- bus
Inverter
M
Motor
8ariable )peed #rive
04 12
Line voltage
= J 04 12
$
%
&
'nverter
()otor
(*oad
$
%
&
'nverter
()otor
(*oad
Three "hase ,ectifier
+igure 0-/ Topology and typical circuit model of variable speed drive.
:e have already observed a representative 8)# current waveform in -hapter 3 (+igure
3-B)& which clearly shows four pea%s per cycle. The harmonic powers computed from the
waveform showed significant amounts of / (or )& A& !/ (or !)& !B and !A (+igure 3-0).
Thus& we observe that 8)#s generate significant amounts of higher harmonics& especially
the fifth and seventh.
>ecause constant linear loads do not generate harmonics
/4
& e.cept real and reactive
powers at the fundamental freuency (04 12)& higher harmonic powers are free from the
influences of the activities of constant linear loads. Thus& if a site mostly has constant
linear loads in addition to a 8)#& e.g. an 18!- room& higher harmonics may be used to
%eep trac% of the fundamental powers consumed by a 8)#. This approach reuires the
%nowledge of a certain relationship between the higher harmonic powers and the
fundamental powers of a 8)#.
+igure 0-3 shows 8)# harmonic powers as functions of time. The data were collected at
a test building at /04 )ansome )t. )an +rancisco& -!. The building has two 8)#s to
drive a supply fan (/44 hp) and a return fan (AB hp) for its main air circulation system.
These fans are turned on and off at the same time and connected via the air duct system.
Thus& these two 8)#s can be conceived as a single unit for 8)# load trac%ing purpose&
because they are operated in tandem and serve the same physical system.
/4
1owever a constant linear load can change the relative phase relationship between its current and voltage
waveforms. If there is no phase difference between the two waveforms& the load is pure resistive (X R 4). If
current waveform lags voltage waveform& the load is inductive (positive X). If current leads voltage& the
load is capacitive (negative X).

"
"=
"B
"A
X
X=
XB
XA
+igure 0-3 1armonic powers consumed by a pair of 8)#s.
The graph shows an early morning startup of the 8)#s. To avoid a sudden pressure
buildup in the air duct& 8)# speeds are slowly increased over si. minutes. ! pressure
loop controls fan capacities (i.e. speeds) afterwards. The graph clearly shows that there
e.ist positive relationships between the fundamental powers and the higher harmonic
powers of a 8)# (or a set of 8)#s). The third harmonic powers are essentially 2ero&
although their noise variances increase with the magnitudes of the fundamental powers.
There are almost constant biases in the fifth and seventh harmonic powers before the
8)# set is turned on. These biases need to be subtracted from observations to measure
the true harmonic powers generated by the 8)#s. The biases of the fundamental powers
are not constant because of the occasional activities of the bac%ground load that consists
of light& plug and other small loads. 1owever& their variations are relatively small& and
thus can be subtracted by ta%ing averages.
Li%e constant loads& the 8)# load vector siFnI can be modeled as Daussian& i.e. siFnI T N
(iFnI& AiFnI). ! load vector siFnI R FFnI !FnI =FnI !=FnI BFnI L I
T
is the harmonic
power vector at discrete time n consumed by load i. ! constant load has constant mean
vector and covariance matri.. @n the other hand& because 8)#s do not have fi.ed steady
states& their means and covariances change with time. 1owever& to ma%e the Daussian
8)# model practical for the purpose of trac%ing 8)# fundamental powers& we have to
impose certain restrictions in the time-varying nature of its mean and covariance.
It is reasonable to assume that the covariance matri. is proportional to the suare of mean
real power7
I F
I F
I F
I F
3
/ &
/ &
r i
r i
i
i
n
n
n
n G
F
F
G
|
|
.
|

\
|
~ & (0./)
/44
where
I F
/ &
n
i
F
is the first component (real fundamental power) of the mean vector at time
n& and
I F
/ & r i
n F
and I F
r i
n G are reference mean real power and covariance matri. ta%en at
time nr& respectively. The references are usually ta%en when 8)# i is running under its
most common operating conditions. Thus& the covariance matri. at time n is determined
once the mean real power is %nown.
In estimating the mean power vector& we assume that there e.ist certain relations among
its components& i.e.

( ) . . I& F I& F & I& F I F
/ & / & / & &
n n n f n
j i j i i j i +
= F F F F
. (0.3)
1owever from the perspective of trac%ing fundamental powers using higher harmonics&
we are interested in the relationships between the higher harmonic means (especially fifth
and seventh) and the fundamental means. :e further assume that a higher harmonic mean
can be faithfully described by either the real or the reactive mean7

( ) ( ) B I F I F I F
3 & / & &
> ~ ~ j n g n f n
i i j i
F F F
. (0.=)
:e basically are assuming that there e.ist strong correlations among mean harmonic
powers so that one can be estimated reliably from another mean harmonic power. If the
above functional relationships are one-to-one& the real and the reactive mean can be
obtained from the measurement of a selected
I F
&
n
j i
F
& using inverse functions.
The mean of a harmonic power is typically measured by ta%ing averages of the observed
harmonic powers for a given duration. 1owever if the data are biased& the time average is
not a reliable estimate of true mean. -urrently& the harmonic powers generated by the
power coder are biased because of the phase bias. This phase bias is related to the
uncertainty in finding the e.act 2ero crossing of the fundamental voltage waveform
during sampling and the amount of bias increases with the harmonic number& as
e.plained in -hapter 3.
-urrently& we obtain biased harmonic powers from the power coder. 1owever& we still
need an unbiased measure of a higher harmonic power to %eep trac% of the fundamental
powers. @ne such a measure is the #th apparent power "#& defined in -hapter 37

3 3
# # #
! " + = . (0.?)

"# is simply the magnitude of the #th comple. power& whose real part is the #th real
power and imaginary part is the #th reactive power. "
/
has units of (8!). >ecause !# is
the magnitude of the #+T of the current waveform (with a scale factor)& it is free from the
time-shift of the reference voltage waveform.
:ith "# selected as an unbiased measure& we need to establish the relationships between
the higher harmonic apparent powers and the fundamental powers. !s an e.ample& +igure
0-= and 0-? show the correlation graphs between the fundamental powers and the fifth
and seventh apparent powers. The data were collected at the /04 )ansome )t. building on
/3E4BE43 in early morning. The graphs clearly show the positive relationships between the
fundamental powers and the higher harmonic apparent powers. 1owever the relationship
is not linear and the slope of a graph decreases with increasing or !. Thus when the
8)# set is running at its nominal operating condition& it tends to draw higher harmonics
less aggressively. This can be conceived as a %ind of saturation effect.

+igure 0-= -orrelation graphs between real power and selected apparent powers.

+igure 0-? -orrelation graphs between reactive power and selected apparent powers.
+igure 0-B shows another set of correlations ta%en from the supply fan 8)# of the
-hildren6s -ourt building in L! county. The data were ta%en on //E3?E43 after midnight&
while other loads are off& by manually running the 8)# at different speeds (as evidenced
/43
by distinct clouds). The power signal from that building is uite noisy& but the overall
correlation trend is similar. :e note that the correlation of "
A
is relatively flat and will be
less useful in trac%ing & compared to that of "
B
.

+igure 0-B -orrelation graphs between real power and apparent powers (L! building).
These correlations are comple. results of how a 8)# is constructed& what %ind of
physical load is attached to it (e.g. fan& duct and building space) and how the 8)# is
controlled. Theoretically& if we can accurately model these subsystems& we should be able
to predict these correlations. 1owever& overall physical system is often difficult to model
and its e.act modeling is unnecessary for our purpose. 1ere& we will obtain these
relationships empirically from the measurements.
The nature of the physical load& i.e. the building air handling system& e.plains why we
observe the saturation effect on the correlation graphs. +or e.ample& when the 8)# is
turned on in the morning& the building space is not pressuri2ed and thus the 8)# can
easily induce airflow. 1owever& as the space is being pressuri2ed& the 8)# has to wor%
harder to maintain airflow& changing the condition at which the 8)# is operated and
hence the correlations between the fundamental powers and the higher harmonic apparent
powers. This shows up as gradual changes in the correlation graph slopes. (specially& the
correlation graphs of ! seem to have three or four distinct operation ranges.
!lso& the time-varying nature of the physical load implies that the correlations are not
fi.ed. If the details of the physical load change& they will create a different operating
condition for a 8)# and thus will change the correlations. The duct and building space
will have different resistance& inertance and capacitance values depending on how air
terminal bo.es are regulated to meet varying heating& cooling and ventilation loads at
different locations in a building. Thus the correlations& in general& will differ day by day
and even vary slowly within a single day.
1owever& it also is true that there e.ist representative nominal operating conditions for
8)#s& especially the ones used for 18!- systems& and the variations from these
representative correlations are typically small& i.e. B J /4 O range. Thus& we use the
representative correlations obtained from measurements to %eep trac% of 8)# powers in
an appro.imate sense& although true correlations changes continually.
Now& the tas% is to obtain parameteri2ed estimations of the correlations. The estimations
will be used to establish the inverse relationships of (0.=)& in order to obtain fundamental
mean powers from higher harmonic mean powers. >ecause the slope of a correlation
graph gradually changes& we can not use a linear fit (straight line). ! higher order
polynomial may give an acceptable fit for the whole range& but the resulting function is
difficult to invert. @ne viable approach is to use a piecewise continuous function to
represent a correlation graph. This approach is appealing in the sense that a correlation
graph consists of different regions of operation conditions. ! linear fit within a region
may be acceptable& but will produce a poor result unless each region is sufficiently
narrow. ! better alternative is to use a power function with a coefficient and an e.ponent.
"iecewise power functions provide good fits to observed correlation graphs and three or
four different regions are typically good enough to represent the whole range.
+or simplicity& we denote the #th apparent& real and reactive mean power as "#& # and !#&
respectively. -onsider a correlation between and "# for a give regionG then "# can be
represented as a power function of 7

b
#
a f " = = ) ( & (0.B)
where a is the coefficient and b is the e.ponent of the power function& respectively. If we
ta%e logarithm on both sides&
b a "
#
ln ln ln + = . (0.0)
Thus it is a linear function in terms of the logarithms of powers.
)uppose that we are interested in estimating a and b from a length-N observation. Let
| |
T
# # #
N " " " I / F ln I / F ln I 4 F ln = . y and | |
T
N I / F ln I / F ln I 4 F ln = . @ . Then&
& ln 6a @ y = + ~ b a (0.A)
where
I F @ 1 6 =
and
T
b a I Fln = a . The estimation of a can be obtained easily using the
method of least suares (L))7
y 6 6 6 a
T T /
) ( U

= . (0.C)
The procedure can be repeated for different regions to obtain the set of parameters to
describe the correlation between "# and for the whole range. This procedure can also be
repeated for different apparent powers and for the reactive power !.

/4?
+igure 0-0 shows the parameter estimation result of a region for the correlations shown in
+igure 0-=. The resulting parameteri2ed correlations are shown in solid lines along with
the estimated parameters. The lower graph is in a logarithmic scale and the data points are
almost perfectly linear& demonstrating that the power function model is a reasonable
choice.

+igure 0-0 "arameter estimation for one of the regions shown in +igure 0-=.

region /
region 3
region =
region /
region 3
region =
+igure 0-A "iecewise correlations between fundamental powers and "B.
In this manner& correlations can be estimated for each region. These correlations can be
plotted over the whole range of or !& to obtain a piecewise continuous correlation
curve. +igure 0-A shows an e.ample of the correlations between and "B and between !
and "B& plotted for the whole ranges. The boundary between neighboring regions is
determined as where two correlation curves meet. The regional boundaries are shown on
the graphs as the vertical and hori2ontal dash-dot lines. !t each region& the innermost
curve is selected as the regional correlation. -onnecting these regional correlations yields
a piecewise continuous overall correlation curve. The overall correlation is one-to-one
and thus invertible.
:hen a mean estimation of "
B
is given& an appropriate region is first selected based on its
numerical value. Then& the corresponding mean value of or ! is computed using the
coefficient and e.ponent obtained from a loo%up table based on the selected region.
E.2 ;SD po?er estimation method
! simple 8)# load model was developed in the previous section for the purpose of
estimating 8)# fundamental powers from the measurement of higher harmonic powers.
In this section& we develop a systematic methodology to estimate 8)# powers from a
continuous stream of data and to condition the fundamental powers for further load status
analysis. The overall schematic is shown in +igure 0-C.

!
#
mean
estimation
loo%-up
table
whitening
filter
whitening
filter
load status
analy2er
!
#
8)# mean
of X
8)# mean
of "
X
"
9
9
J
J
8)# mean
removed
colored noise
removed
+igure 0-C )chematic of 8)# power estimation and signal conditioning.
The mean of a selected higher harmonic apparent power is first estimated. It is then fed to
the loo%-up table that yields the estimations of the mean fundamental powers consumed
by a set of 8)#s& based on the piecewise continuous correlations between the apparent
power and the fundamental powers. The 8)# mean estimations are subtracted from the
observations of and !& for further load status analysis& i.e. to identify the @NE@++
status of other constant loads. :ith the 8)# mean subtracted& and ! goes through
whitening filters& because 8)#s typically generate colored noise due to their mechanical
load imbalances and the slow rotation speeds of their end loads. The filtered fundamental
powers are free from the undulations caused by the time-varying 8)# mean& and free
from the colored noise. 1owever& any white noise added by the 8)# is retained and
/40
accounted for by the load status analy2er. The load status analy2er identifies the load
status of other constant loads by detecting and classifying load (@NE@++) events& and by
estimating the load states from the stream of fundamental powers& if applicable.
It also is possible to use more than one apparent power to estimate the fundamental
powers consumed by a 8)#. +or e.ample& we may use both "
B
and "
A
to estimate
fundamental powers separately and ta%e averages to obtain the final estimation results.
1owever& in this chapter& we use "
B
only for simplicity.
>ecause the mean of "# has to be estimated& the scheme has to be performed for a window
(or bloc%) of data. ! time-window of data is first collected& and the whole chain of 8)#
mean power estimation and signal conditioning is performed. Then the process is
repeated for the ne.t window of data. #epending on the circumstances& there may or may
not be an overlap between two neighboring windows.
The window si2e and the mean estimation method for "# are determined& in most part& by
the nature of the signal. +igure 0-< shows the time graphs of and "
B
& collected at /34
12& from the /04 )ansome building on /3E4?E43. The graphs show the undulating nature
of & and that the undulation of is readily reflected by "
B
. The undulation has a wea%
sinusoidal nature (with a period of appro.imately =4 seconds)& but its freuency and
phase are not fi.ed but change with time. The oscillation of time-wise power
consumption is related to the nature of the air handling system and how it is controlled.
!ppendi. > e.plains the observation as a limit cycle FDibson& /<0=I.

"
!B
+igure 0-< Time graphs of and "
B
.
(stimating the mean of "
B
as a sinusoidal function is& in general& not applicable because
its sinusoidal nature changes even within a cycle& ma%ing the faithful estimation of
sinusoidal parameters almost impossible. !nother option is to select a reasonable si2e
window& i.e. N R 3?4 T =04& and assume that the mean is described by a polynomial
function of time indices. ! higher-order polynomial will give a better fit for a given
window of data& but associated computational cost will be higher. !lso& too high a
polynomial order often results in an over-fit. In this chapter& we select the polynomial of
order one (line) to describe the mean of "
B
due to its simplicity and efficiency. The first
order polynomial mean estimator is described in !ppendi. -.

0 50 100 150 200 250 300
70
75
80
85
90
P
(
k
W
)
0 50 100 150 200 250 300
45
50
55
A5
(
k
V
A
)
sample number (120 Hz)
+igure 0-/4 #etailed view of and "B.
+igure 0-/4 shows a detailed view of and "B with the window si2e N R 3B0& ta%en at n R
A444 from +igure 0-<. The bottom figure also shows the result of the first order mean
estimation for "B as a dashed line. The graphs show large pea%s and they appear to have a
periodic nature. !n interesting observation is that these noise pea%s appear in phase with
the same periodicity in both and "B. :e have already seen that and "B pic% up the
large time scale undulation (T=4 sec) in unison. :e notice that they also e.hibit these
small time scale noise pea%s (T = 12) in phase. 1owever& they differ in amplitudes.

"
!B
+igure 0-// "eriodograms of signals in +igure 0-/4.
/4C
+igure 0-// shows the periodograms
//
of the +igure 0-/4 signals in a logarithmic
magnitude
/3
scale. The freuency is normali2ed so that the unit normali2ed freuency is
eual to the sampling freuency (/34 12). >oth periodograms show large spectral pea%s&
especially at low freuencies. ! closer e.amination of the graph reveals that the locations
of the spectral pea%s are the same for both and "
B
. +urthermore& the spectral pea%s of
are taller than those of "
B
almost by a constant amount in the logarithmic scale& which
corresponds to a multiplicative scaling factor in magnitudes. Thus& we conclude that
and "
B
have almost identical noise spectral distributions& which are different only in
magnitudes by a constant scaling factor.
These spectral components are often called interharmonics& or subharmonics& because
their freuencies are not the integer multiples of the fundamental freuency& i.e. 04 12
FKacamini& /<<0I. The interharmonics are typically caused either by a periodically
varying load or a freuency difference between two !- systems when they are connected
by a #- lin% FLi et al.& 344=I. The 8)# interharmonics can be caused either by a
mechanical load imbalance (periodic load) or by a #- ripple due to the freuency
mismatch between the rectifier and the inverter.
If an interharmonic is generated by a freuency mismatch between the rectifier and
inverter set& its freuency fi is given by the following formula FDunther& 3443I7

4 3 /
) / ( nf p f m p f
i
= & (0.<)
where f is the fundamental freuency of the rectifier supply voltage& f
4
is the inverter
output freuency& p
/
and p
3
are the pulse numbers of the rectifier and the inverter&
respectively& and m and n are integers. The first term is the harmonic freuency and the
second term is the freuency of the #- ripple& proportional to the output freuency. The
interharmonic freuency can be considered as a modulation of the harmonic freuency by
the ripple freuency.
:hen the interharmonic freuency is in the vicinity of a harmonic freuency& it will cause
a periodic change of the amplitude of voltage or current waveform& or beating. This
phenomenon is commonly %nown as flic#er& because it is believed to ma%e the
fluorescent lights flic%er. The beating freuency fb is given by the following euation7

# i b
f f f = & (0./4)
where f# R #f is a harmonic freuency close to fi. If we insert (0.<) into (0./4) and
rearrange terms7

4 3 /
) / ( nf p f # m p f
b
= . (0.//)
//
The periodogram is defined in -hapter A.
/3
The logarithmic magnitude of & is defined as &
/4
log /4 (decibel& d>).
:e note that the first term is a harmonic freuency and the second term varies linearly
with the output freuency. The beating freuency will vary linearly with the output
freuency but its range will be very short& because the beating occurs when two
freuencies are sufficiently close to each other. Thus& if the output freuency is increased
from 2ero& the beating should intermittently show up as the second term of (0.//) nears
one of the harmonic freuencies. !lso& the beating freuency should decrease and
increase as the second term passes by a harmonic freuency.
+igure 0-/3 shows the #+T magnitude plot of the /04 )ansome 8)# set start-up in a
morning. ! rectangular window of si2e /43? is applied on the time series data of to
compute the #+T. Then& the window is shifted by 3444 points (/0.A sec) to compute the
ne.t #+T. The graph clearly shows that the (first) spectral pea%& or beating freuency&
increases almost linearly with the ramp-up of the 8)# set. Thus& it is more li%ely that the
interharmonics are caused by the mechanical load imbalance of the fan output shaft(s).
The load imbalance causes a periodic pulsation in the 8)# power consumption and its
freuency is directly proportional to the shaft rotation freuency.

+igure 0-/3 #+T magnitude plot of 8)# ramp up.
The locations of the spectral pea%s suggest that the interharmonics& or colored noise& have
a harmonic nature& i.e. the locations are integer multiples of a fundamental freuency. The
harmonic nature is especially evident from the first si. pea%s at low freuency. The
lowest pea% occurs at appro.imately = 12 and ne.t pea%s occur with a = 12 interval.
>ecause the ma.imum speed of a 8)# is often the nominal speed of an induction motor
(=4 12 for a double-pole pair motor) and because there typically e.ists a /47/ speed
reduction between a motor and an air handling fan via a pulley-belt system& the speed of =
12 is appro.imately the rotational speed of a (centrifugal) fan. Thus& it is more plausible
to speculate that the colored harmonic noise is originated by the mechanical load
imbalance of an output shaft and its harmonic components are generated via a type of
resonance mechanism through the pulley-belt system (the pulley is mostly inertive (*)
while the belt has a compliance component (3)).
//4
If the colored noise is generated by a load imbalance& the estimation of its magnitude can
be used to uantify its severity. This information can be useful for a diagnostic purpose&
i.e. to monitor the state of a fan drive system and ta%e a maintenance action when
necessary. !nother implication is that the location of the first noise pea% indicates the
current speed of a fan. Thus& if we can reliably estimate the fundamental freuency of the
harmonic noise& we can monitor the rotational speed of the fan. In other words& we can
have a virtual tachometer. :e e.pect the location of the first pea% (or spacing between
pea%s) on the spectrum to change with time& reflecting the change in fan speed.
!lthough the harmonic colored noise can be useful for diagnosis and speed monitoring&
its large magnitude can interfere with the load status identification system. +or e.ample&
its large pea% can confuse the load event detector& which relies on a clear sharp change of
power levels to detect a load event. !lso& the (constant) load status analy2er e.pects white
noise& and thus can not guarantee optimality in its decision when colored noise is present.
:hite filtering is thus necessary for both and ! even after their 8)# means are
removed.
There can be two approaches in eliminating the 2ero-mean additive signal generated by a
8)# (set). @ne approach is to treat the colored noise as a signal of interest and try to
estimate the signal. Then& the estimated signal is subtracted from the original observation&
and the difference is the filtered signal. This approach is useful when we are interested in
using the estimated harmonic signal for diagnosis and speed monitoring. The other
approach is to treat the harmonic signal simply as a colored noise to be removed by
passing through a filter. The construction of an optimal whitening filter is closely related
to the estimation of the power spectral density (")#) of the signal.
In this chapter& we assume the presence of an ideal whitening filter to filter the colored
noise of 8)#s. The estimation and elimination of the 2ero-mean additive harmonic signal
generated by a 8)# will be fully discussed in the ne.t chapter.
E.3 Test o' ;SD po?er estimation method
!s a test& the overall 8)# mean power estimation and signal conditioning scheme is
performed for a span of data& collected at /34 12& from the /04 )ansome building on
/3E4?E43. The span consists of =4 non-overlapping rectangular windows each with the
si2e of N R 3B0.
+igure 0-/= shows the time waveform of "
B
(upper graph). It also shows the first order
mean estimation of "
B
as a solid line within the graph. >ecause the mean estimation is
piecewise continuous across windows but linear within a single window& it is not as
smooth as the natural undulation of "
B
. It nevertheless represents the time varying nature
of the "
B
mean reasonably well. The bottom graph shows the time waveform of "
B
after it
is white-filtered. The output of a whitening filter is often called innovations (process).

0 1000 2000 3000 4000 5000 6000 7000 8000
45
50
55
60
A5 and its first order mean estimation
A
5

(
k
V
A
)
0 1000 2000 3000 4000 5000 6000 7000 8000
45
50
55
60
A5 after its harmonic noise denoised
A
5

(
k
V
A
)
n, sample number
+igure 0-/= Time graph of "B with its first order mean estimation and innovations.
+igure 0-/? shows the time waveform of (upper graph). Its 8)# mean estimation from
the "B mean estimation via the loo%-up table is also shown as a solid line within the
graph. If we are interested in estimating the mean real power consumed by a 8)# (set)&
the proposed scheme can provide the estimation as a function of time. The bottom graph
shows the innovations of after its 8)# mean is subtracted and the difference is passed
through a whitening filter.

0 1000 2000 3000 4000 5000 6000 7000 8000
60
70
80
90
100
P and its VSD mean estimation
P

(
k
W
)
0 1000 2000 3000 4000 5000 6000 7000 8000
0
5
10
15
20
P after its harmonic noise/VSD mean removed
n, sample number
P

(
k
W
)
+igure 0-/? Time graph of with its 8)# mean estimation and innovations.
Ideally& we e.pect flat innovations of when other constant loads remain unchanged.
1owever& the innovations still show a slight undulation carried over from the original
waveform. The 8)# mean power estimation from the mean estimation of a higher
harmonic apparent power is not fully accurate& because both means do not increase by a
constant scaling factor. 1owever& on average& the overall scheme performs reasonably
//3
well in removing the effect of a 8)# (set) from & both the mean and the colored noise of
the 8)#. !lso& on that day& no other constant load was on e.cept bac%ground loads&
which consist of small loads with random @NE@++ events. Thus& the innovations are the
real power collectively consumed by these bac%ground loads& and it is possible that they
have a few activities within the span of the data we observe.
+igure 0-/B shows the time waveform of ! (upper graph). Its 8)# mean estimation from
the "
B
mean estimation is also shown as a solid line. The bottom graph shows the
innovations of ! as before. 1ere again& the 8)# mean estimation from "
B
does not
remove completely the undulation in the original waveform& although the result can still
be used for the load status analysis purpose. The innovations of ! show more undulations
then the innovations of do& mainly because ! generally has a smaller magnitude than .
Its undulation is not as aggressive as that of and when compensated by the mean
estimation from "
B
& whose undulation is more aggressive than that of !& the result is a
slight over-compensation.
!nother possibility is the phase bias of harmonic powers. !s e.plained in chapter 3& the
fundamental power has a phase bias of 3.C/ (degree). This is a relatively small angle&
which may be ignored. 1owever& when one component (in this case the reactive power or
imaginary part) has a substantially smaller magnitude than the other& the component with
a smaller magnitude suffers more from the phase instability. +or e.ample& if R C4 %:
and ! R 3B %8!,& with a 3.C/ degree phase uncertainty ! can change by /0 O of its
nominal value while changes only by /.B O.

0 1000 2000 3000 4000 5000 6000 7000 8000
15
20
25
30
35
Q and its VSD mean estimation
Q

(
k
V
A
R
)
0 1000 2000 3000 4000 5000 6000 7000 8000
0
2
4
6
8
10
Q after its harmonic noise/VSD mean removed
n, sample number
Q

(
k
V
A
R
)
+igure 0-/B Time graph of ! with its 8)# mean estimation and innovations.
+igure 0-/0 shows the standard deviations of the white filtered signals (innovations).
!lthough certain amounts of fluctuations e.ist& the standard deviation for a signal is
relatively stable& in part due to the effectiveness of the white filtering. The white noise
characteristic of the 8)# can be estimated from the innovations.

0 5 10 15 20 25 30
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Standard deviations after harmonic noise/VSD mean removed
window number (N=256)
(
k
W
,

k
V
A
R

o
r

k
V
A
)
P
Q
A5
+igure 0-/0 )tandard deviations of innovations for each window.

"
X
"
X
+igure 0-/A "eriodograms of and ! before and after white filtering.
+igure 0-/A shows the periodograms of and ! before and after white filtering for a
window of data (second window of the whole span). The top graph shows the
periodograms of and ! before filtering. !lthough the locations of spectral pea%s are the
same (i.e. same fundamental and harmonic freuencies)& the heights of the spectral pea%s
are largely different between and !. ! in general has smaller spectral pea%s and
relatively higher pea%s at mid-freuency range. !pparently ! is less noisy than . The
bottom graph shows the periodograms after white filtering. Large spectral pea%s
disappeared& and the periodograms are relatively flat for the overall freuency range.
Thus& we verify the effectiveness of the white filtering. The standard deviations of the
innovations (")# magnitude after white filtering) are the inherent properties of the 8)#
white noise.
//?
E.. Chapter summary
! practical 8)# power estimation method has been developed in this chapter& which
relies on the mean estimation of a higher harmonic apparent power and the correlation
between the apparent power and the fundamental powers. -ertain higher harmonic
powers are suitable for 8)# load trac%ing& because there e.ist strong correlations
between them and the fundamental powers consumed by a 8)# and because they are not
mas%ed by the activities of constant linear loads. Thus& the real and reactive power
consumed by a 8)# can be estimated reliably as a function of time& even in the presence
of the activities of other constant linear loads.
! Daussian random process model is proposed for a 8)# with a time-varying mean
vector and a time-varying covariance matri.. The model is simplified by assuming that
the covariance matri. is proportional to the suare of real power and that there e.ist
strong correlations between mean harmonic powers. The correlations were further
reduced so that there e.ists a one-to-one relationship between a higher harmonic mean
and a fundamental mean. The apparent power is introduced because of the phase bias of
the power coder.
The actual correlations were obtained from the observations and the piecewise continuous
power function model was adopted to represent a correlation curve over the whole range
of or !. ! power function becomes linear on a logarithmic scale and the method of
least suares is applied to estimate its parameters. The overall correlation curve can be
generated from these parameters& and the region boundaries are recorded in a loo%-up
table so that appropriate parameters can be selected based on the mean estimation of a
higher harmonic apparent power. The mean of a selected apparent power is estimated as a
first order polynomial.
The nature of the colored noise generated by a 8)# was discussed. The harmonic noise
generated by the 8)#s at the /04 )ansome building was attributed to the load imbalance
at the fan shaft(s). #ue to its physical origin& the magnitude of the colored noise can be
used to monitor the severity of the imbalance and its fundamental freuency can be used
to monitor the rotational speed. The strategy to estimate the harmonic signal was
discussed and the whitening filter was introduced to eliminate the colored noise. The
8)# mean and colored noise are removed from the real and reactive powers before they
are e.amined by a load status analy2er. The white noise generated by a 8)# still remains
in them and the analy2er has to ta%e it into account in analysis.
The test results show the effectiveness of the 8)# power estimation method based on the
observation of higher harmonic powers& thought the method is not fully accurate and
leaves many rooms for further refinements. Nonetheless& it is a great achievement in the
sense that no 8)# power estimation method has e.isted unless a dedicated power meter
is installed. @nce this method is established and adopted& subseuent improvements and
better alternatives will naturally follow.
Chapter 5
Estimation and Elimination o' 6armonic Signals 'rom Po?er 8a#e'orms
:hile any time-varying load can generate non-white noise& variable speed drives (8)#s)
are the most common source of colored noise on power waveforms. The colored noise of
a 8)# has a periodic nature and its spectral pea%s are often harmonically related to their
fundamental freuency proportional to the rotational speed of the 8)#. +urthermore& the
colored noise generally has a time-varying nature& i.e. the magnitudes of its spectral pea%s
and its fundamental freuency change with time. It is desirable to estimate the harmonic
signal generated by a 8)# to monitor the drive. It also is necessary to suppress the
colored noise generated by a 8)# for the functionality of a constant load status identifier&
which relies on the white noise model of the power signal generated by constant loads. In
this chapter& various techniues are introduced to estimate and eliminate the harmonic
signals generated by 8)#s.
5.1 ,ntroduction
Identifying load status is the most basic tas% of an electric load information system based
on the NI"M and provides a necessary basis to perform further diagnosis and
optimi2ation study. !naly2ing the status of loads has to rely on a certain signal model& so
that the current observation of harmonic powers can be represented as a sum of individual
signal sources& each of which describes the behavior of a specific load. The Daussian
random process model was introduced for constant electric loads in earlier chapters& in
which a load generates harmonic powers while it is on& following a Daussian distribution
with a fi.ed mean vector and a constant covariance matri..
! 8)# can also be modeled as a (Daussian) random process& but with a time-varying
mean and a time-varying covariance matri.. The elimination of the time-varying mean
was discussed in the previous chapter. #ue to the colored noise generated by a 8)#& the
Daussian model can not be applied in a strict sense& because it implicitly assumes that the
noise is white. Nonetheless& the Daussian random process model is still valid if we model
the 8)# as a sum of two signal sourcesJ a Daussian source with a white noise and a
colored noise sourceJ and if we can eliminate the latter. 4hite filtering refers to the
procedure of eliminating the additive colored noise of a signal and restoring its white
noise content. :hite filtering is necessary not only to apply a Daussian model for a 8)#
but also to guarantee the functionality of the constant load status identifier& which relies
on the Daussian model and the whiteness of the observations. In this chapter& we are
interested in developing a whitening filter for the colored noise generated by a 8)#.
The colored noise of a 8)# is often be related to the load imbalance of its end shaft.
1ere& the imbalance can be due to the rotor itself& e.g. shaft misalignment& uneven
distribution of rotor inertia& uneven friction& etc.& andEor to the medium impelled by the
rotor& e.g. air in a fan& refrigerant in a compressor& etc. This implies that the colored noise
reflects the nature of this load imbalance and that by estimating the colored noise we can
//0
analy2e its characteristic. +or e.ample& the severity of the imbalance is reflected by the
magnitude of the colored noise& and by constantly monitoring the magnitude we can
diagnose whether the mechanical drive of a 8)# is in a good shape.
#ue to its origin& the colored noise of a 8)# has a periodic nature& with its freuency
proportional to the rotational speed of its end shaft. +urthermore& the colored noise often
has a harmonic nature amplified through a resonance mechanism. The ")# of the colored
noise often has several harmonically related spectral pea%s with their fundamental
freuency eual to the rotational speed of the end device (fan& compressor& etc.). This
implies that if we can estimate the fundamental freuency of the harmonic signal
generated by a 8)#& we can monitor the rotational speed of the end shaft. In other words&
we can build a virtual tachometer solely from the non-intrusive power signals. This can
be helpful when one tries to build a fault detection and diagnosis system for a 8)#& in
which the rotor speed is constantly monitored and compared to a command input& if
available& or to a nominal operational speed range. !ny discrepancy is reported as a fault.
It may be unfair to treat the harmonic signal generated by a 8)# simply as a colored
noise due to its interesting nature& i.e. the mechanical load imbalance. Thus& we may treat
the harmonic signal as a signal of interest and estimate the signal. The estimation can be
further analy2ed for further monitoring and diagnosis. Nevertheless& the goal of filtering
the colored noise can still be achieved by subtracting the estimation from the observation.
! good estimation routine can perform the dual role of signal e.traction and elimination.
In this chapter& we will develop several methods to estimate the harmonic signal
generated by a 8)#.
-olored noise is not necessarily an e.clusive byproduct of a 8)#& but can be produced by
any motor load that has a dynamic (rotational) component. 1owever& we rarely observe a
colored noise from a constant motor load. ! single-speed induction motor typically runs
at 04 12 or =4 12 depending on the number of its pole pairs. (ven if there e.ists a certain
load imbalance& this is too high a speed to be captured by the /34 12 power coder unless
there is a significant gear reduction ratio. The resulting noise will be more white than
colored. !lso& small single-speed induction motors are typically used to pump liuid
(mainly water) through a closed pipe loop& whereas large induction motors powered by
8)#s impart movement to gas medium (air in fan& refrigerant in compressor& etc.). The
liuid has more inertia and viscosity than the gas& which in turn helps to stabili2e the
rotor. This in part e.plains why we do not observe colored noise from constant motor
loads& whose ma5ority are small water pumps.
This chapter first e.amines the method of detecting and estimating harmonic signals from
periodograms. Then& its two varieties are proposed. :e also develop a white filtering
method by treating the harmonic signal simply as a colored noise. The pros and cons of
these methods will be discussed at the end of the chapter.
5.2 Detection and estimation o' harmonic signal 'rom periodogram
To estimate a harmonic signal& we first have to detect its presence. ,unning a harmonic
signal estimation routine& when there is no such a significant signal& is a waste of resource
and further the routine will either fail or return a uestionable result. Thus& a harmonic
signal detector is a prereuisite to estimate harmonic signals reliably and accurately. !lso&
the development of a detector often gives rise to its accompanying estimator.
The estimation of a harmonic signal hinges on the estimation of its fundamental
freuency (or period). The estimation of other components& such as amplitudes and
phases& are usually straightforward once the fundamental freuency is available.
(stimating the fundamental freuency of a harmonic signal is closely related to
estimating its power spectral density (")#).
:hen &FnI is a real discrete-time wide-sense stationary (:))) random process& its ")#
&&(e) is defined by the :iener-$hinchine theorem F"apoulis and "illai& 3443I7
t e t e e s s =

= #
&& &&
# j # r ) e.p( I F ) ( & (A./)
where r&&F#I is the autocorrelation of &FnI defined as&
I) F I F ( I) F I F ( I F n & # n & ( # n & n & ( # r
&&
+ = = . (A.3)
Note that &&(e) is simply the discrete-time +ourier transform (#T+T) of r&&F#I and that it
is a nonrandom continuous function of e. The ")# can be interpreted as the average
power of &FnI in the narrow freuency band from e to e 9 de F$ay& /<CCI. 1owever& the
analytic evaluation of &&(e) reuires the complete %nowledge of the probability density
function ("#+) of &FnI& which is generally un%nown. Therefore& we often have to rely on
the estimations of the ")#& from a given finite length observation.
The periodogram& I&&(e)& of an N point seuence &FnI is defined as

3
3
/
4
) (
/
) e.p( I F
/
) ( e e e X
N
n j n &
N
I
N
n
&&
= =

=
. (A.=)
Note that the periodogram is a random continuous function of e and that it is the suared
magnitude of the #T+T of the signal& normali2ed by the signal length. In practice& the
periodogram is computed via the discrete +ourier transform (#+T). In this case& only
samples of I&&(e) at e# R 3t#EN are evaluated. The periodogram can also be defined as
FMarple& /<CAI


=

=
+ = =
Y Y /
4
/
) / (
I F I F
/
I F where & ) e.p( I F ) (
m N
n
&&
N
N m
&& &&
m n & n &
N
m m j m I e e
(A.?)
//C
The sample autocorrelation &&FmI is an estimate of the true autocorrelation r&&F#I& based
on a length-N observation of &FnI. #ue to the similarity in the definitions of &&(e) and I&&
(e)& the periodogram is called a natural estimate of the power spectral density. 1owever&
even though the periodogram is an asymptotically unbiased estimator of the ")#& it is not
consistent in the sense that its variance does not decrease with increasing data record
length F@ppenheim et al.& /<<<I.
)uppose that sFnI is a real harmonic signal with . as its fundamental period& where . is
a positive integer. The signal sFnI is a sum of harmonically related sinusoids at
freuencies fi R iE.& or

(

=
+ =
/
/
3
) 3 cos( I F
.
i
i i i
n f " n s | t & (A.B)
where each sinusoid amplitude "i is a ,ayleigh random variable with (("i
3
E3) R i. (ach
phase |i is uniformly distributed in the interval of F-t& t). !ll the random variables are
independent under this signal model. )uppose that | |
T
N & & & I / F I / F I 4 F = . @ represents a
length-N observation vector of a real discrete-time :)) random process &FnI. :e want to
detect whether &FnI contains sFnI7

= + =
=
& / & & / & 4 I F I F I F 7
I F I F 7
/
4
N n n w n s n & $
n w n & $
.
(A.0)
where wFnI is a 2ero-mean additive white Daussian noise (:DN) with variance o
3
. Thus&
under the first hypothesis ($
4
)& the signal is purely :DN. ;nder the second hypothesis&
the observation is the sum of a harmonic signal and a :DN.
Diven the above signal model& $ay F/<<CI proposed the following generali2ed li%elihood
ratio test (DL,T) to detect the presence of a harmonic signal with an un%nown
fundamental period .7

(
(
(
& ln
3
/
/
) (
ln
) (
ma.
ln
3
/
) G (
) &
U
G (
ln ma.
) & ( ma. ) (
3
/
/ E ) E ( & /
3 3
3
4
/
3
3

o o
o
>
|
|
.
|

\
|

|
|
.
|

\
|
=
|
|
.
|

\
|

=
=

> =
N
I I
N
$ p
$ p
. h T
.
. i I i
.
i
&& .
i
&&
.
.
.
.
.
&&
@
P @
@ @
(A.A)
where h(@& .) is the raw test statistic and is the threshold of the test. The raw test
statistic is evaluated for a set of possible fundamental periods and the ma.imum is
selected as the test statistic T(@) of the test. The test statistic is compared with the
threshold. If it is smaller than the threshold& $4 is selected (no harmonic signal).
@therwise& $/ is selected (presence of a harmonic signal). Note that in the latter case& the
fundamental period of the harmonic signal is determined from the argument of the
ma.imum raw test statistic.
The first term of the raw test statistic is the generali2ed li%elihood ratio (DL,) of the
hypothesis test& where
P
U
is the ma.imum li%elihood (ML) estimation of un%nown
parameters with
(
| |
T
.

/ 3 /
3

= . P
. The second term is the penalty factor based on the
minimum description length idea F,issanen& /<ACI. I&&() is the periodogram of &FnI.
! practical concern in computing the test statistic is the evaluation of the periodogram.
!s discussed before& only samples of the periodogram at freuencies f# R #EN are available
via the N-point #+T of the signal. The periodogram should have a sufficiently dense
freuency population so that it can be evaluated at various sample freuencies (iE.). Note
that as . increases& its freuency increment (/E.) gets smaller. !lso increasing . by one
results in totally different harmonic freuencies. +or a given range of .& N should be
sufficiently large to accommodate all possible sampling freuencies.
1owever& depending on the circumstances& the data length N may not be made arbitrarily
large. +or e.ample& the signal may have a time-varying nature or the system may need an
update at an interval shorter than N. In this case& one solution is to increase the
periodogram si2e (denser freuency population) by 2ero padding. In 2ero padding& the
periodogram si2e is increased to * (* Z N) by padding * J N 2eros at the end of the signal
and performing *-point #+T. 1owever& 2ero padding does not improve the freuency
resolution. It simply provides us with a method for interpolating the values of the
measured spectrum at more freuencies F"roa%is et al.& 3443I.
Diven .& the periodogram si2e reuirement can be e.pressed by the following simple
formula7
|
.
|

\
|

+
> |
.
|

\
|
+

*
#
*
#
. .
i
/
3
/
/ /
& (A.C)
where i is the harmonic number& # is the freuency number and * is the length of the
periodogram seuence. It simply states the fact that the freuency resolution provided by
the periodogram should be smaller than the resolution at harmonic number i. The scaling
factor 3 is introduced to allow rounding to the nearest integer for #. The worst case occurs
when i R /& i.e. in evaluating the fundamental freuency or period. :e rearrange the
ineuality with i R /& to get the simple result7
) / ( 3 + > . . * . (A.<)
:ith * roughly proportional to .
3
& as . increases& * has to increase uadratically. In our
application& the fundamental period is on the order of B4 with the /34 12 power coding
rate. Thus& * is typically set at C/<3.
/34
:hen the fundamental period . of a harmonic signal sFnI (corrupted by an additive white
noise) is available& its ML estimation is given by the following formulae F:ise et al.&
/<A0I7

& / I F
/
/ 4 I F
/
/
I F U
/
4
4
s s +
s s +
+
=

=
=
. # 5. N r. # &
5
5. N # r. # &
5
# s
5
r
5
r
(A./4)
where

.
N
5 = & i.e. the number of length-. cycles in the length-N observation of &FnI.
Note that
I F U # s
is an .-periodic seuence& i.e.
I F U I F U r. # s # s + =
& r e . Thus& the ML
estimation of a harmonic signal is simply the periodic-average of an observed waveform.
:hen N is an e.act multiple of .& only the second formula is used. @bviously& the
uality of estimation increases with 5& although depending on the circumstances it can
not be made arbitrarily large. In practice& three or four seems to be a reasonable lower
bound for 5 to yield a reliable estimation.
:hen multiple harmonic signals e.ist& it is customary to detect and estimate each
harmonic signal seuentially. (ach harmonic signal is detected& estimated and eliminated
from the observation starting from the one with the largest magnitude (strongest signal)&
until the detection test statistic of the remaining observation becomes smaller than a
threshold. ! simultaneous detection and estimation of multiple harmonic signals is
possible& but is computationally more e.pensive and reuires a priori %nowledge of the
number of harmonic signals& which may change depending on the circumstances. In this
chapter& we ta%e the seuential detection and estimation approach.
To eliminate the harmonic signal with a %nown fundamental freuency from an
observation& a certain type of filtering has to be performed. In cochannel spea%er
separation in speech processing& a comb filter with nulls at the harmonic freuencies is
applied to the observation to suppress the harmonic signal& while a comb filter with pea%s
at the harmonic freuencies is used to enhance the harmonic signal FMorgan et al.& /<<AG
Lim et al.& /<ACI. This techniue of harmonic enhancement and suppression wor%s
reasonably well for speech signals& in which only relative strengths of spectral
components are important (the absolute strength of an audio signal is easily changed with
amplification). 1owever& in our case& we want to conserve the absolute strength of both
the harmonic signal and the remaining observation for further analysis. Thus& instead of
applying a comb filter to suppress a harmonic signal& we simply subtract its ML
estimation from the observation (implicit comb filtering). The difference (or remaining
observation) is sub5ect to further harmonic signal detection test.
+igure A-/ shows the time graphs of the real power & the reactive power ! and the fifth
apparent power "B defined as
3
B
3
B
! + & i.e. the magnitude of the fifth comple.
harmonic power. The data were collected at a test building in /04 )ansome )t.& )an
+rancisco& -!& on /3E4?E43& at the /34 12 preprocessing rate. The building has two
8)#sJ one for a supply fan (/44 hp) and the other for a return fan (AB hp) in its air
handling system. The graphs clearly show periodic pea%s& implying the colored nature of
the signal generated by the 8)#s. ! closer e.amination of the graphs reveals that the
period and the phase of the periodic signal components are the same across the three
waveforms. 1owever& they differ in magnitudes. The real power shows the largest
amplitude& followed by "
B
and !.

0 50 100 150 200 250 300
70
80
90
100
P

(
k
W
)
0 50 100 150 200 250 300
20
25
30
35
Q

(
k
V
A
R
)
0 50 100 150 200 250 300
45
50
55
60
A
5

(
k
V
A
)
n, sample number
+igure A-/ Time graphs of power signals.

"
!B
+igure A-3 "eriodograms of power signals in +igure A-/ after their 8)# means removed.
+igure A-3 shows the periodograms of the signals in +igure A-/ after their 8)# mean
values are removed. The periodograms clearly show the harmonic nature of the periodic
signals and that the fundamental freuencies are the same across the power signals.
1owever& they differ in harmonic signal magnitudes. and "B differ almost by a constant
/33
amount in the logarithmic magnitude
/=
scale& implying a constant (multiplicative) scaling
factor in the original magnitude scale. ! has smaller pea%s than or "
B
& but has relatively
high magnitudes at mid and high freuency region. The figure suggests that we may use
either or "
B
to detect the presence of a harmonic signal across power waveforms (! has
too small magnitudes) and to estimate its fundamental freuency& but that the estimation
should be performed separately on each power signal because they contain different
amounts (magnitudes) of harmonic signals.

10 20 30 40 50 60 70 80 90 100
-500
0
500
1000
h
(
x
,
M
)
raw test statistic as a function of M to find a harmonic signal
10 20 30 40 50 60 70 80 90 100
-100
-50
0
50
h
(
x
,
M
)
10 20 30 40 50 60 70 80 90 100
-150
-100
-50
0
h
(
x
,
M
)
M, period
+igure A-= ,aw test statistics to detect the presence of a harmonic signal.
+igure A-= shows the raw test statistic at each different stage of the seuential harmonic
signal detection test& using the in +igure A-/ as an observation seuence. The top graph
shows the raw test statistic as a function of . for the original observation. The graph
shows two almost identical-height pea%s at . R ?/ and . R C3. 1owever& the pea% at . R
?/ is slightly higher and the test estimates the fundamental period of the first harmonic
set as ?/& which is indeed correct. Nonetheless& the test may yield . R C3 with a slight
change in the signal content. eriod doubling error is often observed and is a wea% point
of the harmonic signal detection and period estimation method based on the periodogram.
The second graph of +igure A-= shows the raw test statistic after the ML estimation of the
first harmonic signal is subtracted from the observation. The graph clearly shows the
presence of a second harmonic set with . R B4& although it is much wea%er than the first
harmonic signal. The upper graph of +igure A-? shows the periodogram of the
observation after the first harmonic set is subtracted. The periodogram clearly shows a
pea% at f R /E. R 4.43. :e remember that the building has two 8)#s. )o& it is natural to
e.pect the presence of two separate harmonic sets. In reality& the first strong signal is
related to the supply fan and the second wea% signal is related to the return fan. Thus& we
can detect and estimate two harmonic signals generated by two different physical loads&
which in turn helps to monitor the state of each load.
/=
The logarithmic magnitude of & is defined as &
/4
log /4 (decibel& d>).

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-40
-20
0
20
40
d
B
Periodogram of P after first harmonic signal removed
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-40
-20
0
20
d
B
Periodogram of P after second harmonic signal removed
normalized frequency
+igure A-? "eriodograms of observations after harmonic signals are subtracted.
The third graph of +igure A-= shows the raw test statistic after the ML estimation of the
second harmonic set is subtracted from the observation. It shows no conspicuous pea%
and decreases almost monotonically. :ith the threshold of appro.imately 2ero& the test
will conclude that there e.ists no more harmonic signal. The lower graph of +igure A-?
shows the periodogram of the observation after two harmonic sets are subtracted. It
shows no outstanding pea%& implying the absence of a harmonic signal. !lso& we note that
the periodogram is s%ewed rather than flat. 1ence& the harmonic signal subtraction is& in
effect& a harmonic suppression filter rather than a whitening filter.

0 50 100 150 200 250 300
-5
0
5
10
First harmonic set estimated from P
(
k
W
)
0 50 100 150 200 250 300
-4
-2
0
2
4
Second harmonic set estimated from P
(
k
W
)
n, sample number
+igure A-B ML estimations of harmonic signals from .
/3?

0 2 4 6 8 10 12 14 16 18 20
0
1
2
3
Amplitudes of the first harmoninc signal
P

(
k
W
)
0 5 10 15 20 25
0
0.5
1
1.5
Amplitudes of the second harmoninc signal
P

(
k
W
)
harmonic numbers
+igure A-0 !mplitudes of each harmonic set from .
+igure A-B shows the results of the harmonic signal ML estimation. The second set is
estimated after the first set estimation is subtracted from the observation. The graphs
clearly show that the first set has much larger magnitudes than the second set and that the
fundamental freuency of each set has the largest amplitude. +igure A-0 shows the
amplitude of each harmonic set at each harmonic number& using the relationship between
a sinusoid amplitude and its corresponding periodogram pea% height F$ay& /<<=I7
) (
?
U
3
# && #
f I
N
" = & (A.//)
where f# R #E. is the #th harmonic freuency and "# is the amplitude of the sinusoid with
the freuency f#. N is the length of the observation seuence &FnI. The fundamental
freuency of each set has the largest amplitude as evidenced in the time waveforms. The
amplitudes of the first set roll off uic%ly after the si.th harmonic. The amplitudes of the
second set become negligible after the tenth harmonic.
+igure A-A shows the harmonic signals estimated from !& with %nown fundamental
periods. +igure A-C shows corresponding harmonic amplitudes. !gain the first set has
larger magnitudes than the second set. The amplitude at each harmonic freuency and
their relative strengths are in general different from those of the harmonic sets from .
This section has applied the method of harmonic signal detection and its fundamental
period estimation based on the periodogram and the ma.imum-li%elihood harmonic
signal estimation to the power waveforms. The tests show that the ML estimation yields
good results& when at least four cycles (of a fundamental period) are included in a single
observation seuence. The periodogram-based harmonic signal detection and
fundamental period estimation also performs reasonably well. 1owever& it is slightly
unsatisfactory in the sense that it often generates period doubling errors and that it
reuires a priori %nowledge of the additive white noise covariance& which may change
with time. !lso& it is burdensome to have to compute a large si2e periodogram. The ne.t
section introduces a better alternative in detecting a harmonic signal and estimating its
fundamental period using autocorrelation.

0 50 100 150 200 250 300
-2
-1
0
1
2
3
First harmonic set estimated from Q
(
k
V
A
R
)
0 50 100 150 200 250 300
-2
-1
0
1
2
Second harmonic set estimated from Q
(
k
V
A
R
)
n, sample number
+igure A-A ML estimations of harmonic signals from !.

0 2 4 6 8 10 12 14 16 18 20
0
0.2
0.4
0.6
0.8
Amplitudes of the first harmoninc signal
Q

(
k
V
A
R
)
0 5 10 15 20 25
0
0.1
0.2
0.3
0.4
0.5
Amplitudes of the second harmoninc signal
Q

(
k
V
A
R
)
harmonic numbers
+igure A-C !mplitudes of each harmonic set from !.
5.3 Detection and pitch estimation o' harmonic signal using autocorrelation
#etecting and estimating a harmonic signal has been one of the central problems in
speech processing& because a (voiced) sound has a harmonic nature due to the periodic
pulsation of the glottis (an opening between the vocal folds) and the subseuent
e.citation (resonance) in the vocal tract F#eller et al.& 3444I. The fundamental freuency
of a voiced sound (hence a harmonic signal) is often called pitch. )everal pitch detection
/30
and estimation methods are %nown to the speech processing community& such as the
harmonic selection based on the histogram method F"arsons& /<A0I& the cepstral analysis
F)tubbs and )ummerfield& /<CCI& the modified covariance spectrum estimator based on
the autoregressive signal model FNaylor and "orter& /<</I& the mean-suare-error
estimator based on the sinusoidal speech model FMc!ulay and Xuatieri& /<<4I& delay
filtering and error minimi2ation in the linear predictive coding (L"-) F-hen et al.& /<<=I&
etc. 1owever& the ML pitch detector and estimator based on the autocorrelation is widely
used and performs well in the presence of additive noise FMorgan et al.& /<<AG Naylor and
>oll& /<CAI. In this section& we apply the same method to detect the harmonic signal and
estimate its pitch from a power waveform.
Diven a length-N real seuence &FnI& the raw test statistic ((@& p) to detect the presence of
a harmonic signal with a pitch period p is given by F:ise et al.& /<A0I7

=
=
*
#
&&
#p R
N
p
p (
/
I F
3
) & (@ & (A./3)
where

p p N * ) ( = and R&&F#I is the autocorrelation function& defined as


=
+ =
# N
j
&&
# j & j & # R
/
4
I F I F I F
. (A./=)
Note that the autocorrelation function is an unnormali2ed estimate of the true
autocorrelation of a real :)) random process. The test statistic is selected by
ma.imi2ing ((@& p) over a range of (integer) p. The ML estimation of the pitch period is
simply the argument of the ma.imum raw test statistic.
((@& p) is often interpreted as the inner product of the autocorrelation function and a comb
filter 3pF#I with intertooth spacing p& i.e.

otherwise. 4
& = & 3 & / & if /
I F
. = =
=
l lp #
# 3
p
(A./?)
Then the raw test statistic can be represented as

&&
T
p
N
p
p ( r c @
3
) & ( = & (A./B)
where | |
T
&& && && &&
N R R R I / F I / F I 4 F = . r and | |
T
p p p p
N 3 3 3 I / F I / F I 4 F = . c . The
comb filter is determined by ma.imi2ing the correlator output of 3pF#I and R&&F#I.
The ML pitch detector and estimator is often susceptible to pitch period doubling and
halving errors. To remedy this problem& the comb filter is modified to have teeth with
decreasing amplitude FMartin& /<C3I. In practice& a power weighting of #
4.C
is applied to
each element of the comb filter7

=
/
/
C . 4
I F I F
3
) & (
N
#
&& p
# R # 3 #
N
p
p ( @ . (A./0)

0 50 100 150 200 250 300
-2000
0
2000
4000
R
x
x
[
k
]
Autocorrelation function from P
0 50 100 150 200 250 300
-500
0
500
1000
R
x
x
[
k
]
0 50 100 150 200 250 300
-200
0
200
400
R
x
x
[
k
]
k
+igure A-< !utocorrelation functions computed from .

0 20 40 60 80 100 120 140
-50
0
50
E
(
x
,
p
)
Raw test statistic based on autocorrelation
0 20 40 60 80 100 120 140
-5
0
5
10
E
(
x
,
p
)
0 20 40 60 80 100 120 140
-2
0
2
E
(
x
,
p
)
p, pitch period
+igure A-/4 ,aw test statistics from autocorrelations in +igure A-<.
The first graph of +igure A-< shows the autocorrelation function computed from the real
power waveform in +igure A-/. The autocorrelation clearly shows the periodic nature
with decreasing magnitude and its period is eual to the pitch period of the harmonic
signal& i.e. p R ?/. The periodic character of the autocorrelation function is evident from
/3C
its definition& where products of observations with a lag # add constructively when the lag
is an integer multiple of an underlying pitch period. The first graph of +igure A-/4 shows
the corresponding raw test statistic as a function of p. The ma.imum clearly occurs at the
correct pitch period& although multiples of the pitch period also have pea%s with
substantial but smaller heights. Thus& we verify that the prospect of period doubling is
significantly reduced& an advantage over the method based on the periodogram in the
previous section.
The first harmonic set is estimated using the ML estimator discussed in the previous
section and subtracted from the observation. New autocorrelations are computed from the
difference and shown in the second graph of +igure A-<. The autocorrelation still shows a
periodic nature& implying the presence of the second harmonic set. The second graph of
+igure A-/4 shows the corresponding raw test statistic. The second harmonic signal with
p R B4 is detected& estimated and subtracted from the observation. The third graph of
+igure A-< shows the autocorrelation of the observation after two harmonic sets are
subtracted. It does not show a conspicuous period and its corresponding raw test statistic
in the third graph of +igure A-/4 shows no dominant pea%.

0 5 10 15 20 25 30
40
40.2
40.4
40.6
40.8
41
Period estimation of first harmonic set
s
a
m
p
l
e

n
u
m
b
e
r
0 5 10 15 20 25 30
50
51
52
53
Period estimation of second harmonic set
window number (N=256)
s
a
m
p
l
e

n
u
m
b
e
r
+igure A-// (stimated pitch periods for a series of observation.
+igure A-// shows the ML pitch estimation for each harmonic set from the data recorded
at the /04 )ansome building on /3E4?E43. The estimation was conducted for thirty
consecutive windows with no overlap& each with a length of N R 3B0. The first graph
clearly shows the oscillatory nature of the pitch period of the first harmonic set. !s
e.plained before& the pitch of a harmonic signal is proportional to the rotational speed of
a corresponding 8)#. Thus& the graph suggests that there is an oscillation in the
rotational speed of the supply fan. >ecause the pressure in the supply air duct regulates
the fan speed& this in turn suggests that there e.ists a pulsation in the duct air pressure.
This phenomenon can be e.plained as a limit cycle created by the combination of the
physical characteristic of the air duct system and the specific control logic used to control
the fan speed based on the duct air pressure& as e.plained in !ppendi. >. 1owever& the
oscillation in the fan speed is not fi.ed but varies with time.
The second graph of +igure A-// shows the pitch period estimation of the second
harmonic set. ;nli%e the pitch period of the first set& it fluctuates rather randomly
between B4 and B=& though its true period seems to lie somewhere between B/ and B3.
>ecause the speed of the return fan (source of the second harmonic signal) is regulated by
the static pressure measured at a location in large building space& it should be relatively
constant. The une.pected random nature of the second harmonic pitch may have a
physical origin& such as a noise or disturbance in the velocity control loop of the return
fan.
1owever& it is more li%ely that the randomness of the second harmonic pitch is caused by
the spectral lea#age in the process of windowing and harmonic subtraction. !ny
windowing forces a sharp impulse-li%e freuency component to be dispersed around its
nominal freuency. !lso& the first harmonic set is not necessarily confined at its nominal
harmonic freuencies for various reasons. :indowing is obviously one of the reasons.
The time-varying nature of the signal actually creates a very narrow band of pitch& not
necessarily an impulse. !lso& the true pitch period of a harmonic signal is not necessarily
an integer. Thus& the first harmonic set has bands of harmonic freuencies& and their ML
estimations are& in a sense& weighted averages of each band with the constraint that the
pitch period is an integer. :hen the ML estimation of the first harmonic set is subtracted
from the observation& not all of its spectral components are eliminated but some of them
lea% into the remaining observation. The converse is also true. )ome of the spectral
components of the second harmonic set may be pic%ed up by the first harmonic set
estimation if their freuencies are located sufficiently close to the harmonic freuencies
of the first set.

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Locations of harmonic frequencies, p1 = 41, p2 = 52 (solid circles)
normalized frequency
+igure A-/3 1armonic freuencies of two harmonic sets with pitch periods of ?/ and B3.
+igure A-/3 shows the harmonic freuency locations of two harmonic sets when pitch
periods are ?/ (open circles) and B3 (solid circles). The graph demonstrates that the first
and fourth harmonics of the first set& which typically have large amplitudes& are located
/=4
very close to the harmonics of the second set. It is inevitable that some elements of the
second set are included to the first set estimation and the lea%age from the first set
changes the spectral content of the second harmonic signal. #epending on the nature of
true spectral content (or ")#) of each harmonic set& the degree of the spectral lea%age
may vary. This in turn creates a distortion in the spectral content of the second harmonic
set embedded in the remaining observation and shows up as a fluctuating pitch period
estimation. Improvements are possible by a better windowing (1amming window& etc.)& a
simultaneous pitch estimation& etc. These issues will be e.plored in future. The current
seuential detection and estimation techniue still yields reliable and useful results and
the pitch period estimation is usually within BO of its nominal value.
+igure A-/= shows the standard deviation of each power waveform after its two harmonic
sets and 8)# means are removed. The standard deviations e.hibit substantial
fluctuations partly due to the time varying (noise) nature of the 8)#s and mostly due to
the incompleteness of the harmonic set removal. !s e.plained before& spectral content of
the two harmonic sets lea%s into the remaining observation. !lso& the harmonic de-
noising by subtraction performs poorly towards the boundary of a window. The ML
signal estimation tends to align itself to the harmonic waveform in the middle of a
window& if the signal has a time varying nature. :e also remember that the harmonic
subtraction is not a white filtering in a strict sense. ! separate white filtering may be
performed to reduce the fluctuations in the standard deviations.

0 5 10 15 20 25 30
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Standard deviations after harmonic noise/VSD mean removed
window number (N=256)
(
k
W
,

k
V
A
R

o
r

k
V
A
)
P
Q
A5
+igure A-/= )tandard deviations of observations after their harmonic signals and 8)#
means are removed.
+igure A-/? shows the amplitude at each harmonic number of the first harmonic set
estimated from the real power waveform at each window. The amplitude becomes
negligible after the si.th harmonic and the first and the fourth have dominant magnitudes.
The amplitude distribution is relatively stable over time& implying the consistency of the
underlying load imbalance mechanism (or the stable nature of the resonanceEe.citation
mechanism). +igure A-/B shows the amplitudes of the second harmonic set. The
amplitude distribution is also relatively stable& although the amplitudes at high harmonic
numbers are slightly uneven& possibly due to the spectral lea%age.

+igure A-/? !mplitudes of first harmonic set at each window.

+igure A-/B !mplitudes of second harmonic set at each window.
+igure /0 shows the total amplitude of each harmonic set defined as7

=
=
*
i
i
" "
/
3
& (A./A)
where * is the ma.imum harmonic number and "i is the ith amplitude of a harmonic set.
!lthough there e.ist fluctuations& we observe that the total magnitude of each set follows
/=3
a certain mean value. The total magnitude is a suitable measure of the e.tent of the load
imbalance and a monitoring routine can be installed which detects a fault when a low-
pass filtered version of the total magnitude e.ceeds a certain threshold.

0 5 10 15 20 25 30
3.6
3.8
4
4.2
4.4
4.6
Total amplitude of first harmonic set
P

(
k
W
)
0 5 10 15 20 25 30
1.2
1.4
1.6
1.8
2
2.2
Total amplitude of second harmonic set
P

(
k
W
)
window number
+igure A-/0 Total amplitude of each harmonic set.
In this section& we applied the ML pitch detection and estimation method based on the
autocorrelation to detect and estimate the pitch periods of the harmonic signals generated
by 8)#s. The ML detector performs better than the periodogram based detector in the
previous section& in the sense that the former generates less period doubling errors& that a
priori %nowledge of noise variance is not reuired and that the computational burden is
slightly reduced
/?
. 1owever& the ML pitch estimation is still unsatisfactory because the
pitch period is constrained to be an integer. In the ne.t section& we further refine the
method to allow a fractional pitch period.
5.. (ractional pitch period estimation and least squares signal estimation
The ML pitch estimator provides a reasonably accurate pitch period estimation for most
applications. 1owever& its performance is limited due to the integer pitch period
constraint. ! small round up error in the pitch period can result in significant biases in the
higher harmonic freuency estimations. Natural processes do not necessarily have integer
periods (of sampling intervals)& especially when they have time-varying natures.
+ractional pitch (+") period estimation is often employed to improve the uality of
estimation& obviously at the cost of additional computation and comple.ity. In an +"
period estimation& the pitch period is allowed to be an arbitrary positive real number. The
/?
+or a length-N seuence& the autocorrelation reuires appro.imately 3 ) / ( + N N multiplications and
3 ) / ( + N N additions. ! si2e * fast +ourier transform (++T) reuires * *
3
log additions and *
*
3 3
log
multiplications. :ith N R 3B0& =3C<0 3 ) / ( = + N N . :ith * R C/<3& /40?<0 log
3
= * * .
+" period estimation can be performed directly from the +ourier transform of a signal&
such as by pea# pic#ing FXuatieri and #anisewic2& /<<4I& or through an additional
refinement procedure based on an integer pitch period estimation. The latter includes an
interpolation filter F-hen et al.& /<<=I& in which the sampling rate is interpolated to match
a proposed +" period& and the iterative minimi2ation of the difference between the
original spectrum and the estimated spectrum to obtain the optimal estimations of signal
parameters (including the +" period) FDriffin and Lim& /<CCI.
There also e.ist general-purpose sinusoidal parameter estimation techniues such as
"isaren%o6s harmonic decomposition method F"isaren%o& /<A=I& M;)I- F)toica and
Nehorai& /<C<I and ()",IT F,oy and $ailath& /<C<I. These methods& which rely on the
signal subspace idea& provide high freuency-resolution estimations of multiple sinusoids
in an additive white noise. 1owever& they are not specifically designed for a harmonic
signal& in which sinusoidal freuencies are harmonically related. They also reuire very
accurate estimations of the autocorrelation matrices for the eigen-decompositions. Thus&
they are not good candidates for a harmonic signal with a time-varying nature.
In this section& we apply a variation of the harmonic enhancement and suppression
method& in which a comb filter is constructed with a +" period. This method can yield an
+" period estimation with a reasonable accuracy without the burden of the iterative
minimi2ation FMorgan et al.& /<<AI.
+or a given fractional pitch period p& the corresponding second-order harmonic
suppression filter (1)+) $

(,) is given by the following formula7

3
) / (
) / )( / (
) (
o
o o
+

=

p p
, ,
, $ & (A./C)
where 4 S o S /. The variable o is chosen close to one& and o R 4.<< is normally used. @n
the freuency domain& $

(,) is simply a comb filter with nulls at the harmonic

freuencies of /Ep.
! length-N observation seuence &FnI is filtered through h

FnI to yield a harmonically

suppressed signal sFnI7
sFnI R &FnI h

FnI& (A./<)
where denotes circular convolution. +or each +" period from a set of candidates& sFnI is
computed and the one that gives minimum signal energy is selected.
The circular convolution can be implemented efficiently in the discrete freuency domain
via the #+T&
I F I F I F # $ # X # %

= . (A.34)
/=?
The #+T of the 1)+ is evaluated as (due to the noninteger circular shift property of the
#+T)7

& / / Y ) (
4 Y ) (
I F
3 ) E ) ( 3 e.p(
3 ) E 3 e.p(
s s +
s s
=
=
=

N # , $
# , $
# $
N
N N # j ,
N
N # j ,
t
t
(A.3/)
assuming an even N. The second norm of %F#I is minimi2ed to select the best +" period.
:hen an integer pitch period estimation p is given& the fractional pitch period candidates
pi can be obtained either to limit the error at the highest harmonic freuency less than
one-half bin of an N-point #+T& or with a fi.ed incrementEdecrement interval from p. The
former is represented by FMorgan et al.& /<<AI7

( ) I i I p p
N
i
i
s s + =
3
/ & (A.33)
where i is an integer and I is a bound. The latter is simply given as7
I i I p i p p
i
s s + = _ & (A.3=)
where p _ is an increment.
+igure A-/A shows an e.ample of the 1)+ result when p R ?/ and Ap R 4./B. The norm of
the filtered signal shows a minimum at pi R ?/.?B.

38.5 39 39.5 40 40.5 41 41.5 42 42.5 43 43.5
180
185
190
195
200
205
210
215
220
225
norm of filtered output
fractional periods
|
|
S
[
k
]
|
|
+igure A-/A Norm of harmonic suppressed signal as a function of fractional pitch period.
Diven a +" period estimation& we would li%e to estimate the corresponding harmonic
signal from the observation. 1owever& we can not use the ML signal estimator because it
relies on the integer pitch period assumption. The ML estimator may be applicable if the
signal is upsampled via interpolation so that the +" period becomes an integer. The
iterative minimi2ation method FDriffin and Lim& /<CCI can be applied to estimate a
harmonic signal with an appropriate sinusoidal signal model. The iterative minimi2ation
can be costly because it tries to see% a global minimum in the difference between the
original spectrum and a synthesi2ed spectrum. !n alternative is to see% a local minimum
at a harmonic freuency via a freuency warping and repeat the procedure for every
harmonic freuency F-hoi and Koun& 3443I.
1owever& these methods are generally comple. and computationally intensive. In this
section& we develop a simple least suares (L)) signal estimator based on a harmonic
sinusoidal signal model. The L) estimation of a harmonic signal in the freuency domain
was investigated by Xuatieri and #anisewic2 F/<<4I.
-onsider a sinusoidal harmonic signal sFnI with a pitch period p7

=
+ =
*
i
i i i
n f " n s
/
) 3 cos( I F | t & (A.3?)
where * is the ma.imum harmonic number and fi R iEp is the ith harmonic freuency.
(ach cosine term can be e.panded into the sum of a cosine and a sine using the
trigonometric identity7
n f b n f a n f "
i i i i i i i
t t | t 3 sin 3 cos ) 3 cos( + = + & (A.3B)
where
i i i
" a | cos = and
i i i
" b | sin = . :e assume that a length-N observation seuence
&FnI is the sum of sFnI and a 2ero-mean additive noise wFnI7
/ 4 I F I F I F s s + = N n n w n s n & . (A.30)
In a vector-matri. format&
? 6J @ + = & (A.3A)
where
| | | | | |
.
) / ( 3 sin ) / ( 3 cos ) / ( 3 sin ) / ( 3 cos
4 3 sin 4 3 cos 4 3 sin 4 3 cos
& & I / F I / F I 4 F & I / F I / F I 4 F
/ /
/ /
3 3 / /
(
(
(

=
= = =
N f N f N f N f
f f f f
b a b a b a N w w w N & & &
* *
* *
T
* *
T T
t t t t
t t t t
.
. . . .
.
. . .
6
J ? @
(A.3C)
Note that 6 is an N3* matri. and has a ran% of 3* when 3* S N.
The L) estimation of the parameter vector is given by7
@ 6 6 6 J
T T /
) (
U

= . (A.3<)
/=0
Then& each harmonic amplitude estimate can be computed from the L) parameters&

3 3
U
U
U
i i i
b a " + = . (A.=4)
#ue to the Dauss-Mar%ov theorem& the L) estimator is the same as the best linear
unbiased estimator (>L;() when the additive noise is 2ero-mean white Daussian F$ay&
/<<=I. 1owever& no optimality can be claimed when wFnI is not white Daussian.

0 50 100 150 200 250 300
-5
0
5
10
First harmonic set estimated from P, p = 41, p
i
= 41.45
(
k
W
)
LS
ML
0 50 100 150 200 250 300
-4
-2
0
2
4
Second harmonic set estimated from P, p = 50, p
i
= 51.05
(
k
W
)
n, sample number
LS
ML
+igure A-/C 1armonic signals estimated from (ML estimations same as +igure A-B).

0 2 4 6 8 10 12 14 16 18 20
0
1
2
3
Amplitudes of first harmonic set from P
(
k
W
)
LS
ML
0 5 10 15 20 25
0
0.5
1
1.5
Amplitudes of second harmonic set from P
(
k
W
)
harmonic number
LS
ML
+igure A-/< !mplitudes of each harmonic set.
+igure A-/C shows the L) estimations of harmonic signals from the real power in +igure
A-/. The figure also shows the corresponding ML estimations with integer pitch periods
(reproduction of +igure A-B). (ach fractional pitch period estimated from the 1)+ is also
shown in the figure. The graphs suggest that the L) estimator tends to estimate a
harmonic signal less aggressively than the ML estimator. :e also note that two
estimations tend to align in the middle of a window and grow out of phase towards
window boundaries. +igure A-/< shows the amplitudes of each harmonic set& both from
the L) and the ML estimation. The graphs verify that the L) estimator underestimates the
harmonic signal amplitudes.
+igure A-34 shows the periodograms of the original observation and the residual
observation after the L) harmonic signal estimation is subtracted. :e observe that the
residual still has relatively large pea%s in the vicinity of the third& fourth and fifth
harmonic freuencies of the first harmonic signal.

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-60
-50
-40
-30
-20
-10
0
10
20
30
Periodograms of P
normalized frequency
(
d
B
)
original
after LS denoised
+igure A-34 "eriodograms of real power before and after de-noising.
There can be various reasons for this unsatisfactory result. !s discussed before& the
harmonic signal actually has a narrow band for each harmonic freuency rather than a
fi.ed impulse-li%e location due to the time-varying nature. The 1)+ may yield an +"
period estimation well adapted to a few large pea%s& but the resulting harmonic
freuencies can be off by significant amounts at other pea%s. The L) signal estimation is
very sensitive to freuency locations& thus a small offset between the center of a
freuency band and freuency estimation can yield a significant bias in the signal
estimation. !nother concern is that a harmonic signal does not necessarily have pea%s at
e.act harmonic locations& often 5ustifying the use of a freuency-warping techniue. !lso&
the harmonic signal in power waveforms has a relatively low signal-to-noise ()N,) ratio
(compared to speech signals). The combination of these causes results in a relatively flat
band of minimum values in +igure A-/A& reducing the credibility in determining the e.act
+" period.
The presence of the second harmonic signal can affect the +" period estimation of the
first harmonic signal& because their first pea%s are located close to each other. The second
harmonic signal also acts as a non-white noise in the estimation of the first harmonic
signal. ! better windowing may be necessary to minimi2e the signal distortions. !lso& the
/=C
power signal itself has certain distortions introduced by the power coder& during the
course of interpolation and resampling of the raw voltage and current waveforms.

0 5 10 15 20 25 30
39.5
40
40.5
41
41.5
42
Period estimation of first harmonic set
s
a
m
p
l
e

n
u
m
b
e
r
0 5 10 15 20 25 30
48
50
52
54
Period estimation of second harmonic set
window number (N=256)
s
a
m
p
l
e

n
u
m
b
e
r
+igure A-3/ (stimated fractional pitch periods for a series of observation.

0 5 10 15 20 25 30
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
Standard deviations after harmonic noise/VSD mean removed
window number (N=256)
(
k
W
,

k
V
A
R

o
r

k
V
A
)
P
Q
A5
+igure A-33 )tandard deviations of residue observations (+" periods used).
+igure A-3/ shows the +" period estimation of each harmonic set from the same data as in
+igure A-//. In spite of the increased resolution& it is difficult to see any clear periodicity&
compared to +igure A-//. +igure A-33 shows the standard deviations of the residue
observations after the L) estimations of the +" period harmonic signals are subtracted.
-ompared to +igure A-/=& it generally shows larger standard deviations& implying a poor
signal estimation and subtraction result.
! fractional pitch period estimation coupled with an appropriate harmonic signal
estimation techniue can potentially yield a better estimation of the signal and hence a
better harmonic de-noising of an observation. 1owever& when the signal has a time-
varying nature and a low )N,& it is difficult to estimate the +" period accurately and thus
the signal estimation and de-noising can not be performed reliably. !lso& the power signal
has distortions created by the power coding and the method itself introduces various
distortions& ma%ing the +" period estimation unreliable. The current +" period estimation
and harmonic signal estimation techniue may be applicable& when the circumstances are
more favorable. :e may also develop a more sophisticated method in future. ;ntil then&
we will use the ML harmonic signal estimator with an integer pitch period estimation&
due to its simplicity and robustness.
5.D Autoregressi#e modeling o' harmonic signal and ?hite 'iltering
In this section& we ta%e a different approach from the previous sections by treating the
harmonic signal present in a power waveform as colored noise& rather than the signal of
interest to be estimated. (stimating and eliminating a harmonic signal can be difficult
when it has a time-varying nature andEor a low )N,. )ometimes& we may simply want to
filter out the harmonic noise before the power signal is fed to a load status analy2er&
which relies on the whiteness of the noise present in the observation.
-onstructing a whitening filter& nonetheless& reuires a certain amount of %nowledge
about the nature of a targeted harmonic signal. In this sense& the harmonic noise is still a
signal of interest. 1owever& the estimation of the signal characteristics is performed
implicitly in the conte.t of white filtering. The specific properties of a harmonic signal&
such as its pitch and harmonic amplitudes& will not be available in general.
:e adopt the linear system model for a real 2ero-mean discrete-time :)) random
process &FnI& in which a linear system 6(,) generates &FnI from a 2ero-mean white
Daussian noise (:DN) process input wFnI. +igure A-3= shows the linear system model. 6
(,) is assumed to be time-invariant.

6(,)
wFnI &FnI
+igure A-3= Linear system model of a discrete-time :)) process.
:ith this linear model& the ")# of &FnI is determined solely by the ")# of wFnI and 6(,)
F"apoulis and "illai& 3443I7
) E / ( ) ( ) ( ) E / ( ) ( ) (
3
, 6 , 6 , , 6 , 6 ,
ww &&
o = = & (A.=/)
where o
3
is the ")# of wFnI. Thus the ")# of &FnI is completely determined by 6(,)
within a scaling factor. In this sense& 6(,) is called the shaping filter of &FnI. +or a given
/?4
")#& infinitely many choices of 6(,) are generally possible. 1owever& we are interested
in a unique 6(,) that is causal and stable& and has a causal and stable inverse.
:e can also conceive an inverse linear system in which &FnI is filtered to wFnI as shown
in +igure A-3?. It can be shown that $(,) R /E6(,)& where 6(,) is the uniue one so that $
(,) is causal and stable. $(,) is called the whitening filter of &FnI and wFnI is often
referred to as the innovations of &FnI.

$(,)
&FnI wFnI
+igure A-3? Inverse linear system to whiten &FnI.
>ecause $(,) is the reciprocal of 6(,)& constructing a whitening filter for &FnI is
euivalent to finding a uniue 6(,)& via spectral factori,ation& to represent the ")# of &
FnI. >ecause the true ")# of a random process is often un%nown& designing a whitening
filter is intimately tied to the estimation of its ")#.
There e.ists a variety of ")# estimation methods& such as the nonparametric methods
based on periodograms& the minimum-variance spectral estimation methods and the
parametric methods F$ay& /<CCG $ay and Marple& /<C/I. !mong these& the parametric
methods usually give the best result F"roa%is et al.& 3443I and suit our need because we
want to construct a whitening filter e.plicitly with the obtained parameters. The
eigenanalysis algorithms& such as ()",IT and M;)I-& are often considered as ")#
estimation methods& although they can be more appropriately termed as sinusoidal
parameter estimation methods. They also do not provide the ")# in a parameteri2ed
form& ma%ing it difficult to construct a corresponding whitening filter e.plicitly.
Three types of the parametric model of a random process are used& namely autoregressive
(!,)& moving-average (M!) and autoregressive moving-average (!,M!). The !,
model is good at representing spectral pea%s& while the M! model is good at spectral
valleys. The !,M! model is well adapted when the ")# of a process has both pea%s and
valleys& and in general can yield a ")# estimation with a fewer model order (sum of its
!, order and M! order) than the !, or M! model. 1owever& the estimation of an
!,M! model is more costly than the !, model of an euivalent order& because its M!
part is emulated by a larger order !, process to estimate its parameters F$ay& /<CCI.
!lso& any M! or !,M! process can be represented by an !, process of possibly infinite
order F:old& /<B?I. +or these reasons& the !, model is the most widely used in
estimating the ")# of a random process and is appropriate for our need to model the
harmonic signal that has a pea%y spectrum.
Diven an order p !, model& !,(p)& the random process &FnI is described by the
following time series euation7
I F I F I F
/
n w # n & a n &
p
#
#
+ =

=
& (A.=3)
where a# is the #th coefficient of the !,(p) model. Its shaping filter is all-pole7

+
= =
p
#
#
#
, a
, "
, 6
/
/
/
) (
/
) (
. (A.==)
Note that $(,) is eual to "(,) and that $(,) is all-2ero. Thus& the whitening filter is
simply an M! process with the same coefficients7

=
+ =
p
#
#
# n & a n & n w
/
I F I F I F . (A.=?)
:hen the coefficients are %nown& the ")# of &FnI is given by7
3
3
) (
) (
e
e
o
j
j
&&
e "
e =
. (A.=B)
;nder the !,(p) model& the goal is to select the model order p and the coefficients a# to
best represent the true ")# of &FnI& which in turn allows the construction of the whitening
filter of &FnI. The !, or all-pole model is widely used in speech processing to model a
sound FLim and @ppenheim& /<ACI and in geophysical signal processing to model a
reverberated seismic source FMc-lellan& /<CCI.
If we multiply both sides of the autoregressive euation by &Fn lI (l > 4) and ta%e
e.pectations& we obtain the following euation of correlations7
I F I F I F
/
l r # l r a l r
w&
p
#
&& # &&
+ =

=
& (A.=0)
where ( ) I F I F I F l n & n & ( l r
&&
= is the autocorrelation of the process at lag l. The euation
can be further simplified by evaluating the crosscorrelation7

( )
( )
. I F
I F I F I F I F I F
I F I F I F I F I F I F
3
3
l g
# l n # n g # w l n w ( # n g
# w # n g l n w ( n & l n w ( l r
# #
#
w&
=
+ = + =
|
|
.
|

\
|
+ = + =

=
o
o o
(A.=A)
/?3
>ecause gF#I is a causal filter (gF#I R 4& # S 4)& the crosscorrelation is non-2ero only when
l R 4. -ombining the results& we obtain the celebrated Kule-:al%er euations7

<
>
= +
=

=
=
& 4 I F
4 I F
4 I F
I F
/
3
/
l l r
l # l r a
l # l r a
l r
&&
p
#
&& #
p
#
&& #
&&
o
(A.=C)
where the last euality follows from the symmetric property of a real autocorrelation. In
matri. format&
(
(
(
(

=
(
(
(
(
(

(
(
(
(

+
+
I F
I 3 F
I / F
I 4 F I 3 F I / F
I 3 F I 4 F I / F
I / F I / F I 4 F
3
/
p r
r
r
a
a
a
r p r p r
p r r r
p r r r
&&
&&
&&
p && && &&
&& && &&
&& && &&
. .
.
. . .
.
.
and

=
+ =
p
#
&& # &&
# r a r
/
3
I F I 4 F o . (A.=<)
>ecause the correlation matri. is Toeplit2& it can be efficiently inverted by the Levinson
algorithm FMc-lellan& /<CCI to find a#. The Levinson recursion is summari2ed below7

. I 4 F ) / (
I 4 F
I / F
3
//
3
/
//
&&
&&
&&
r a
r
r
a
=
=
o
(A.?4)
+or # R 3& =& L& p&

& ) / (
/ & & 3 & /
I F I F
3
/
3 3
& / & /
3
/
/
/
& /

=

=
= + =
+
=

# ## #
i # # ## i # #i
#
#
l
&& l # &&
##
a
# i a a a a
l # r a # r
a
o o
o
.
(A.?/)
where
3
#
o is the estimation of
3
o and a#i is the estimation of ai at the #th iteration. The
recursion yields the solution of the Kule-:al%er (K:) euations when # reaches p.
! practical concern is the estimation of r&&F#I& because the true autocorrelation is usually
un%nown. It can be shown that the ML estimation of the !, parameters can be found by
using the following biased sample autocorrelation function F$ay& /<<=I7

| | | |

>
s +
=


=
. 4
/
/
I F U
Y Y /
4
N #
N # # n & n &
N # r
# N
n
&&
(A.?3)
The resulting !, model is stable.
@ther !, parameter estimation methods are also available& such as the >urg method (or
ma.imum entropy spectral estimation) and the unconstrained least-suares method
F"roa%is et al.& 3443G ;lrych and -layton& /<A0I. The >urg method usually yields a better
freuency resolution& at a slightly higher computational cost& than the K: method&
especially when the data length is short. 1owever& the >urg method e.hibits spectral line
splitting at high )N, and spurious pea%s at high order models F"roa%is et al.& 3443I. The
unconstrained L) method generally yields a better ")# estimation with sharp spectral
pea%s than the K: and >urg methods& and has a computationally efficient algorithm
comparable to the Levinson recursion FMarple& /<C4I. ;nli%e the >urg and K: methods&
the unconstrained L) method does not guarantee the stability of the !, model.
:hen the stability of the !, model is not a concern& the unconstrained L) method is
often the best choice. 1owever& e.periences show that when the signal has a combination
of a few large spectral pea%s and many small pea%s& the unconstrained L) method with a
moderate model order tends to focus on the former& creating very sharp pea%s& while
ignoring the latter. >ecause the harmonic signal present in power waveforms has
numerous small pea%s at higher freuencies& the resulting whitening filter constructed
from the unconstrained L) method often over-compensates low freuency components
while under-compensating high freuency components. !lso& the underlying time-varying
nature of the harmonic signal creates relatively broad pea%s at harmonic freuencies&
negating the necessity of sharp pea% estimations. Thus& the unconstrained L) method is
better suited for a stationary signal with a few large amplitude sinusoids. +or the
harmonic signal we observe in the power waveforms& the K: method or the >urg
method is a better candidate. >ecause we can select the data length of a reasonable si2e
(N T 344)& we adopt the K: method for the remainder of this section.
!nother aspect in the use of the !, model is the selection of the model order p. The
model order can be determined by trial and error& but it is helpful to have an order
selection criterion. It also is necessary to chec% the model order occasionally in real time
when the signal has a dramatic time-varying character& i.e. when the harmonic amplitudes
and the number of harmonic sets change with time. The "#ai#e information criterion
(!I-) is one of the well-%nown !, model order selection criterions F!%ai%e& /<A?I. It
minimi2es

N
p
p p
3
) ( U ln ) ( !I-
3
+ = o & (A.?=)
where ) ( U
3
p o is the estimation of o
3
at model order p. !s p increases& ) ( U
3
p o decreases
rapidly& reducing the overall !I-. 1owever& as p passes a certain threshold value& the
/??
decrease in the noise variance becomes negligible& increasing the !I-. 1ence& the !I-
reaches a minimum at a certain p& which is selected as the !, model order. @ther criteria
are also available& such as the minimum description length (M#L) criterion F,issanen&
/<C=I and the autoregressive transfer (-!T) function F"ar2en& /<A?I.

20 25 30 35 40 45 50 55 60
0.54
0.55
0.56
0.57
0.58
0.59
0.6
0.61
0.62
0.63
AIC for a window of real power (N=256)
p, AR model order
+igure A-3B !%ai%e information criterion of a 3B0-point window of real power containing
two harmonic sets.
+igure A-3B shows the !I- applied to a window of real power waveform holding two
harmonic signals (same data as in +igure A-/). The !I- becomes minimum at p R ?=.
Tests show that p R ?4 is the minimal !, order to estimate reliably the ")# of the power
waveform containing two nominal harmonic sets. In this chapter& p R B4 is typically
employed for white filtering.
>ecause both the real and the reactive power reuire white filtering& we have to construct
two filters. :e can either construct a single whitening filter& e.g. from "
B
& and use it for
both channels& or construct two filters separately based on a separate ")# estimation of
each channel. The former is attractive when computational budget is tight. 1owever& its
white filtering result is slightly inferior to that of the latter& because and ! have
different harmonic amplitude distributions although they share the same harmonic
freuencies. In this chapter& we employ the separate estimation and filtering approach for
the real and reactive power channels.
:hen the !, parameter estimation and white-filtering is performed on a continuous
stream of data& the data are divided into si2e-N windows& with or without overlap& and the
estimation and filtering is e.ecuted on a bloc%-by-bloc% basis. The bloc% processing
method is simple to implement and yields satisfactory results in most cases. 1owever due
to the discontinuity of !, parameters at bloc% boundaries& the innovations typically show
larger variances in the beginning of a window. )euential estimation methods (and
subseuent white filtering) also e.ist& which update !, model parameters seuentially as
a new data point becomes available. They rely on the lattice filter reali2ation of an all-
pole system and the recursive estimation of the filter parameters (or prediction
coefficients) F"roa%is et al.& 3443I. In this chapter& we use the bloc% processing method
for its simplicity.

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-40
-20
0
20
40
Periodogram and AR PSD of P before white filtering
(
d
B
)
periodogram
AR PSD
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-30
-20
-10
0
10
Periodogram of P after white filtering
(
d
B
)
normalized frequency
+igure A-30 ")# estimation of real power based on all-pole model and periodograms
before and after white filtering.
+igure A-30 shows the result of the ")# estimation for a segment of real power (+igure A-
/ data) based on the !, model of order B4& along with the periodograms before and after
white filtering. The ")# conforms very well to the first five large pea%s of the
periodogram (top graph)& without creating too sharp or too high pea%s& and generally
follows the contour of the small pea%s at high freuencies. The comparison of the ")#
and the periodogram hints that the !,(B4) model combined with the K: method is
amenable to estimating the true ")# of the real power. The bottom graph shows the
periodogram of the real power after white filtering. The periodogram is relatively flat at ?
d> (whitened)& implying a good white filtering result.
+igure A-3A shows the results for a window of reactive power (+igure A-/ data). !gain&
the ")# estimated from the !,(B4) model via the K: method conforms the shape of the
periodogram very well. :e also notice the difference in the ")#s between the real power
and the reactive power& demonstrating the necessity to construct two separate white
filters. The bottom graph shows the periodogram after white filtering. It is flat at 4 d>.
+igure A-3C shows the standard deviations of the channels after white filtering. Thirty
windows of length-3B0 are used without overlap. -ompared to the case where the
harmonic signals are estimated and subtracted& e.g. +igure A-/=& the standard deviations
are substantially reduced and show much less variations over windows. Thus& we verify
that the white filtering based on an !, signal model is a better option to condition the
observation for a further analysis& when the embedded harmonic signals do not need to be
estimated.
/?0

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-30
-20
-10
0
10
20
Periodogram and AR PSD of Q before white filtering
(
d
B
)
periodogram
AR PSD
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-30
-20
-10
0
10
Periodogram of Q after white filtering
(
d
B
)
normalized frequency
+igure A-3A ")# estimation of reactive power based on all-pole model and periodograms
before and after white filtering.

0 5 10 15 20 25 30
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Standard deviations after white filtering
window number (N=256)
(
k
W
,

k
V
A
R

o
r

k
V
A
)
P
Q
A5
+igure A-3C )tandard deviations after white filtering.
5.E Chapter summary
In this chapter& four different methods were investigated to eliminate the harmonic signals
embedded in the power waveforms. The first three methods treat the harmonic signal as
the signal of interest and rely on its e.plicit estimation and subtraction from the
observations. The last method treats it as a colored noise and constructs the whitening
filter based on the e.plicit estimation of the ")# with a parameteri2ed model. !ll four
methods achieve the common goal of suppressing the harmonic noise& which was often
viewed as a nuisance in processing power signals for load monitoring.
>ecause the first three methods estimate the harmonic signal e.plicitly& they are well
adapted for load monitoring and fault detection& even though they perform poorly from
the viewpoint of white filtering. #ue to the physical origin of a harmonic signal& i.e. the
load imbalance of a rotating machine and accompanying resonance mechanism& it can be
used to monitor the operational status of its source. Its pitch estimation can be used as a
virtual tachometer and the amplitude estimation can be employed to assess the severity of
the imbalance. :ith an appropriate postprocessing decision-ma%ing algorithm& the
%nowledge gained from the harmonic signal estimation can play a significant role in the
overall electric load information system based on non-intrusive power monitoring& for the
ob5ectives of load status identification& fault detection& optimi2ation& etc.
!n interesting observation is the analogy between the harmonic signal in the power
waveforms and human speech. !lthough they have remar%ably different physical origins&
rotating machine vs. vocal tract& they share the common mechanism of pulsation and
resonance. )peech is used to convey an e.traordinary amount of information& and thus
has a very sophisticated structure& e.g. pitch& formant& voicedEunvoiced& word& sentence&
etc. The harmonic signal is not comparable to speech in terms of comple.ity& but
nonetheless can carry a certain amount of low-level information. )trictly spea%ing& the
harmonic signal has the uality of pitch and formant found in the speech. 1umans can
distinguish different spea%ers based on their distinctive pitches and formants. 1armonic
sets can be identified by their characteristic fundamental freuencies and amplitude
distributions. @ne can perceive that a person is in a different mental or physical state
based on his or her abnormal pitch and formant. The electric load information system can
monitor the health of a load by analy2ing the harmonic signal generated by the load.
The first method introduced a harmonic signal detector and its integer pitch period
estimator based on the periodogram of the observation. The signal was retrieved by the
ML estimator. The harmonic detector performs reasonably well& but it is cumbersome
because it reuires a priori %nowledge of the additive white noise covariance and a large
freuency population in computing the periodograms. !lso& it often shows period
doubling errors& ma%ing the pitch estimation completely unreliable.
To overcome the problems associated with the harmonic detector based on periodograms&
the second method employed the ML pitch detector based on autocorrelations. The tests
show that the ML detector is less prone to the period doubling error. The ML detector is
also more convenient to use& because it does not reuire the variance of the additive noise
and is computationally more efficient. ! seuential harmonic detection and estimation
strategy was adapted to handle multiple harmonic sets present in an observation.
Interference in estimating two harmonic sets was also observed. )pectral lea%age occurs
due to windowing& spectral pro.imity between the signals and their time-varying nature.
! better windowing can reduce the spectral lea%age& but can not eliminate it completely.
:e observed the fluctuations in the pitch estimation of the second harmonic set& due to
the spectral lea%age in estimating two harmonic sets seuentially.
The third pitch period estimator was introduced to allow a fractional pitch period
estimation& based on the harmonic suppression filtering. ! simple least suares harmonic
/?C
signal estimator was also developed& because the ML estimator can not handle a
fractional pitch period (without interpolation). !lthough the +" period estimation and
signal estimation is an improvement compared to the former method& the test results are
not satisfactory in the sense that the standard deviations after the subtraction are increased
and that the +" period estimations showed large variations. The L) signal estimator can
be partially blamed because it is not optimal and typically underestimates signal
amplitudes. 1owever& the ambiguity of the +" pitch estimation is more liable to the
underperformance& caused by the combination of the time-varying nature& low )N,&
windowing& interference between harmonic sets and signal distortion introduced by the
power coder. These issues need to be investigated in depth to develop a better +" period
estimation and harmonic signal estimation techniue. ;ntil then& the ML detector and
signal estimator based on an integer pitch period assumption is the best tool to capture the
harmonic signal present in power waveforms.
The fourth method changed the perspective and treated the harmonic signal as colored
noise to be removed by white filtering with a filter constructed from the parametric
estimation of its ")#. The signal was modeled as the output of an all-pole system e.cited
by white noise and the Kule-:al%er method was used to obtain the !, parameters via
the Levinson recursion. The model order of B4 was chosen from the evaluation of the
!%ai%e information criterion and e.perience. The resulting ")# estimation shows a good
conformity to the periodogram& without creating too sharp or too high pea%s. :hen the
observation is filtered by the whitening filter constructed from the !, parameter
estimations& its spectrum is relatively flat& indicating the high performance of the white
filter. :e also note that the white filtering removes the formant (overall ")# contour)
present in an observation& which is impossible in the former three methods because they
can not reshape the ")# over the whole spectrum but only comb at a few harmonic
freuencies.
!nother advantage of the white filtering is that it can be used irrespective of the presence
of a harmonic signal. Thus& it does not reuire a harmonic detector and can be placed in
front of a load status analy2er to whiten the incoming signal at all conditions (possibly
with a real-time model order selector). The white filtering based on the !, ")#
estimation yields smaller and more stable standard deviations of the filtered signal
compared to the harmonic signal estimation and subtraction methods. Thus& it is a better
alternative to eliminate the harmonic noise when the signal does not need to be estimated.
It also is possible to combine the white filtering and one of the harmonic signal
estimation and subtraction methods. +or e.ample& the residual signal from the second
method may be white filtered& to remove the formant and eliminate any harmonic spectral
content lea%ed into the residue.
Chapter H
Test -esults3 Conclusion and (uture 8ork
H.1 Test results
! real-time constant load disaggregator was developed& which can handle multi-load
events. The disaggregator displays current power consumptions graphically on its web
page& and identified load events can be viewed from a lin%ed web page. Its software
architecture is e.plained in !ppendi. #. The disaggregator was tested using a set of
offline data& recorded from /04 )ansome )t.& )an +rancisco& -!. +igure C-/ shows an
e.ample of its event history page. The list of the monitored loads is given in the top
portion of the page. (ach identified load event is e.plained both by a graph and a tag. The
graph displays an event by a red line and a matched e.emplar by green dots. The tag is
shown to the right of a graph& including the number& identification and time of the event.
The load status after the event is also displayed. +igure C-3 shows another e.ample where
a multi-event was successfully captured along with a series of other events.

+igure C-/ (vents identified by constant load disaggregator.

@n the other hand& an NI"M is installed at the laundry room of Ne.t 1ouse& an MIT
undergraduate dormitory located at B44 Memorial #rive& -ambridge& M!. The laundry
room has a three-phase 34C 8 service to power up its ten dryers. ! dryer has a /34 8
single-phase motor (phase-to-neutral) and a 3?4 8 heater (phase-to-phase). The NI"M
currently samples the phase-! current only and nine dryers were observable from the
power signal. !mong those& three dryers have their motors connected to the phase-!. The
rest si. dryers appear as pure heaters in the power signal.
/B4

+igure C-3 Multi-event identified by constant load disaggregator.
The dryer is an e.emplary multi-stage load (@++& Motor and 1eater @N& Motor @N)& and
its heater goes through a series of @N-@++ cycles& based on its thermostat setting& while
its motor runs continuously. !lso& although the dryers can be grouped based on their
steady-state and transient responses& they typically show almost identical responses within
a group. Thus& to identify the load status of a place li%e the Ne.t 1ouse laundry room& the
load disaggregator should be able to handle the multi-stage and multiple identical loads.
The load disaggregator developed in this thesis currently can not handle these situations.
1owever& to test the capability of the load disaggregator in a real situation& it is modified
to identify load events only without an underlying load state model& so that it can be used
as an event trac%er at Ne.t 1ouse. It can recogni2e a dryer @N or @++ event& but it can
not answer how many dryers are on at a given moment. >ecause it does not have a state
model& the ML classifier based on the Daussian random process model is not applicable.
Instead& a simple (uclidean distance classifier is used to classify step change
observations. The transient pattern classifier based on the L) error criterion is still
applicable.
+igure C-= shows an event history page of the event trac%er. The first event (event no.
=0B) is a dryer @N event& whose motor is connected to the phase-!. The ne.t event is
another dryer @N event& but because its motor is not connected to the phase-!& it is
simply reported as a heater @N event. !n event tag shows the event number& time&
description and step magnitudes. +igure C-? shows the power consumption behavior of
last twenty hours. It illustrates that Ne.t 1ouse students are active in laundry after
midnight& in the morning and in the afternoon (on )unday).

+igure C-= Load events identified by event trac%er.

+igure C-? (vent trac%er history page.
! real-time web based 8)# load trac%er was developed and successfully tested at the
)ansome building. Its software architecture is also e.plained in !ppendi. #. +igure C-B
shows its main page. The bottom graph shows the fifth apparent power in red line. Its first
order mean estimation is shown in solid green line. The red line in the upper left graph is
the real power and the solid green line is the estimation of the real power consumed by
the 8)# obtained from the "
B
mean estimation. The real power consumed by non-8)#
loads (bac%ground load in this case) is shown in blue line. Note that the non-8)# power
/B3
is white filtered& based on the autoregressive parameter estimation of the ")# of "
B
. The
upper right graph shows the 8)# trac%ing result of the reactive power.

+igure C-B 8)# trac%er main page (/C73B ?E3 344= ()T
).

+igure C-0 8)# trac%er ")# estimation page (/C73B ?E3 344= ()T
).

(ach graph is based on /43? data points& and with the /34 12 power coding rate& the
main page is updated at every C.B= second. The time server of the computer has a
problem& and as a result& the time stamps in the graphs were not correct. -orrect
observation time is given in the caption of the figure.
+igure C-0 shows the ")# estimation page of the 8)# trac%er. The left graph shows the
periodogram of "
B
in red line obtained from its /43? data points. The parameters of the
autoregressive model of order B4 are estimated from "
B
and the resulting ")# estimation
is shown in solid green line. The periodogram clearly shows si. main harmonic pea%s&
and the ")# estimation follows the contour reasonably well. The right graph shows the
periodograms of the white filtered and ! (blue lines in +igure C-B). The parameters
obtained from the "
B
")# estimation are used to construct a whitening filter. !s a result&
the periodograms shows slight height changes& rather than being flat& because of the
differences in the formant of "
B
and those of and !. 1owever& we note that the
harmonic pea%s are completely removed.

+igure C-A 8)# trac%er history page (3/7BA ?E= 344= ()T
).
+igure C-A shows a history page of the 8)# trac%er. )amples are ta%en in one minute
intervals (as the average of /43? points). #ue to large downsampling and averaging& the
fifth apparent power seems to be almost constant. The graphs show that there was an
activity of a constant load& evidenced by the step changes in the real and reactive power
waveforms. 1owever& the 8)# fundamental power estimations were unchanged& because
the fifth apparent power was unaffected. !s a result& the step changes are conserved in the
non-8)# leftover power signals. This e.ample clearly shows that the 8)# trac%er can
separate the fundamental powers consumed by 8)#s and non-8)# loads.
+igure C-C shows another history page. This figure is interesting in many aspects. ! large
step change set in the afternoon is a chiller activity. The /04 )ansome building has two
/B?
large chillers to meet its cooling load. (ach chiller has a constant speed compressor& but
ad5usting its inlet vane position can vary its capacity. Thus& the chiller can be categori2ed
as a non-8)# variable load. ;nli%e constant loads& non-8)# variable loads can draw
higher harmonic powers& as evidenced by a dip in the "
B
history graph. 1owever& the
higher harmonic power consumed by a non-8)# variable load is almost constant during
its operation& although its fundamental power consumption can vary. +igure C-< shows
this behavior more clearly.

+igure C-C 8)# trac%er history page (3=74B ?EC 344= (#T
).

12.5 13 13.5 14 14.5 15 15.5 16 16.5 17
0
50
100
150
200
250
3/27/02
time (hr)
(
k
W

o
r

k
V
A
)
P
A5
A7
+igure C-< -hange in higher harmonic powers when chiller is on
.
In order to estimate the fundamental powers consumed by 8)#s reliably& any (constant)
amount of higher harmonic power consumed by non-8)# variable loads needs to be
detected and compensated. +igure C-C shows that there are sudden drops in the 8)#
fundamental power estimations when the chiller is on& because the compensation was not
performed. The compensation will reuire a type of change-of-mean detection and
estimation in the stream of "
B
. The general change-of-meanEvariance detector developed
in !ppendi. ! may be employed for this purpose.
!nother observation is that the reactive power of the 8)#s is slightly overestimated. This
is evident from the fact the leftover non-8)# reactive power is almost 2ero& whereas at
nights the reactive power is of substantial amounts. There may have been some changes
since the correlation graph was measured. ;pdating the correlation graph will partially
solve the problem. !nother observation is that the reactive power is more overestimated
in the morning. The 18!- load in the early morning is different from that in the
afternoon& i.e. when the building is fully occupied and the ambient temperature is high.
Thus& the 8)# set will see different load conditions. It may be that the 8)# set has two
different correlation curves and as the loading condition changes& the correlation moves
from one curve to the other. Thus the solution to this problem will be a set of correlation
curves and educated decisions about which correlation curve the 8)# set is currently on&
based on various observations.

H.2 Conclusion
This thesis is a great accomplishment& because it has shown that information about the
status of electric loads can be obtained from the non-intrusive power signals and because
several signal models and signal processing and information handling algorithms were
successfully developed to e.tract the information. The idea of monitoring electric loads
non-intrusively has been around for more than two decades. >ut it has not been widely
received by the academic or technical community& because it has lac%ed a concrete
scientific foundation and because the benefit or usefulness of the information offered by a
non-intrusive monitor has not been clear. Its practicality was also a lingering uestion.
(ven when the monitor was employed as a residential load disaggregator& its performance
was limited because of its rudimentary steady-state load model. It could not handle
overlapping load events and lac%ed a state estimation capability. The variable loads were
declared intractable. This thesis is a partial answer to these challenges.
Traditionally& only a few power terms& such as the real and reactive powers& were
measured to monitor the electric loads. The current harmonic powers& defined in this
thesis& are a more complete and systematic tool to characteri2e a current waveform that
contains rich information about the electric loads. They have a solid mathematical
bac%ground of +ourier analysis and modern discrete-time signal processing. They can be
viewed as an efficient coding process to capture most of the information present in a
/B0
current waveform with a few selected +ourier bases. !naly2ing current harmonic powers
is more economical than analy2ing raw current waveforms& due to the reduction in the
number of samples& and loo%ing at the behaviors of the electric loads from the harmonic
dimension provides an insight and a mathematical basis to characteri2e them.
The concept of voltage harmonic intensity was also introduced& to uantify the harmonic
content of a voltage waveform. It can be computed using the same method of the current
harmonic power& and its time series will possibly provide additional information about
the electric loads. The current harmonic power and the voltage harmonic intensity can be
used in a complementary manner to better characteri2e electric loads and to e.tract
relevant information.
(arlier wor%s considered electric loads to be more or less deterministic& as evidenced
from the approaches to trac% constant loads based on the steady-state power level
changes& although people agreed that there can be variations in their observations. This
thesis introduced a stochastic model to describe a power signal generated by an electric
load or a collection of them. :hen the power signal of an electric load is modeled as a
random process& it has both the property of a mean to describe its steady-state power
consumption and the property of higher order moments& such as variance& to describe its
noise behavior. Thus& each load becomes more distinct when modeled by a random
process& than when modeled by a deterministic process. The stochastic process has a rich
mathematical bac%ground and ma%es it possible to describe the situations& such as the
temporal overlap of multiple events and the steady-state of multiple loads& in clear
mathematical terms. The mathematical description in turn allows developing various
detection& estimation and classification algorithms to identify those situations.
The harmonic powers generated by a constant load at its steady state were modeled as a
Daussian random process. The model is not perfect in the sense that a real statistical
distribution is not necessarily a strict Daussian and even so& its characteri2ing moments&
i.e. mean and variance& can change with time. 1owever& this model is crucial in
e.pressing the observations in various situations in mathematically tractable terms. >ased
on this model& event observations (step changes) and steady-state observations can be
e.plained as linear combinations of individual Daussian processes& which in turn follow
Daussian distributions. >ased on these Daussian distribution models of events and steady-
states& the uni-event classifier& multi-event classifier and load state estimator were
successfully developed.
The transient response of a load is another asset often overloo%ed in the past. !lthough a
few researchers ac%nowledged its potential value& no significant effort has been made to
e.ploit it systematically and in a complimentary manner to the steady-state step changes.
The transient response is not e.plicitly modeled as a stochastic process& due to the
difficulty in modeling and in obtaining its parameters. ,ather& a few representative
responses are used as e.emplars or templates& and the classifiers based on the method of
least suares are developed. The advantage of the least suares method is that it relies on
an implicit statistical noise model and that when the noise is white Daussian it is the same
as the optimal estimator based on an e.plicit noise model. The transient pattern classifiers
were developed to handle both single and simultaneous load events.
>y utili2ing two different aspects of a constant load& steady state and transient& in a
constructive and complementary manner& it was possible to design a more accurate load
disaggregator& which can handle single load events more reliably. It can also handle
multi-load events and identify load states without load events& both of which were
impossible before. !lso& the disaggregation is performed locally in real time& as compared
to the previous central and batch processing. The load disaggregation scheme proposed in
this thesis is a uantum leap compared to the past practices.
The 8)#s were previously considered intractable& because they do not have clearly
defined steady states. 1owever& by watching the behavior of a 8)# from the whole
dimension of current harmonic powers& it is possible to trac% its fundamental power
consumptions. There e.ists a correlation between a higher harmonic power and a
fundamental power& and the former is unobstructed by the activity of other constant linear
loads. The 8)# is still modeled as a stochastic process& but its mean and covariance
change with time and its noise is not white Daussian. To simplify the model& the
covariance matri. was assumed to be proportional to the suare of the real mean power&
and the harmonic mean powers were considered as correlated. The correlations were
obtained from the observations and they were successfully used to estimate the
fundamental mean power consumptions from the estimation of a harmonic mean power.
The noise generated by a 8)# was modeled as the sum of a white Daussian source and
colored noise. The colored noise was removed by white filtering. @nce the mean and
colored noise of a 8)# are removed from the fundamental power observation& the left-
over is simply the fundamental power consumed by non-8)# loads plus the white noise
of the 8)#. )ubseuent (constant) load disaggregation can be performed on this non-
8)# signal& because it now follows a Daussian model& with the residual white 8)# noise
ta%en into account. The 8)# trac%ing scheme proposed in this thesis is a first practical
solution to estimate the fundamental 8)# powers non-intrusively& and is compatible with
the constant load disaggregation scheme proposed in this thesis.
The estimation and elimination of the harmonic signal observable on the power
waveforms was first started as a byproduct of the 8)# trac%ing research& because it was
the colored noise to be removed. 1owever& it has the physical origin of the mechanical
load imbalance of the output shaft of a 8)# and thus carries the information about the
status of the drive. The estimation of the harmonic signal allows us to uantify the
severity of the imbalance from its amplitude and monitor the shaft rotation speed from its
fundamental freuency. Its elimination ma%es it possible to apply the various constant
load disaggregation techniues on the leftover signals.
The harmonic signal was also modeled as the output of a stochastic source. @nce this
model was adopted& numerous estimation and filtering techniues were readily available
because the estimation of a harmonic signal and its ")# has been one of the central
problems in signal processing. Three different methods to estimate a harmonic signal
were presented7 periodogram& autocorrelation and fractional pitch period. The
periodogram-based method wor%s reasonably well& but it suffers period doubling errors
and its computational cost is rather high. The autocorrelation method is the best choice
/BC
and is also widely used in speech processing. The fractional pitch period estimation
coupled with the least suares signal estimation has the potential to improve the overall
uality of the harmonic signal estimation. 1owever& its test result in this thesis is
unsatisfactory& possibly because of the time-varying nature of the signal and the biases
introduced by the power coder. @ne e.plicit white filtering method was also introduced&
which relies on the autoregressive signal shaping model and the estimation of its
parameters using the Kule-:al%er euations to construct the inverse whitening filter.
8arious accomplishments have been made in this thesis& such as putting the current
harmonic powers on a sound mathematical basis& characteri2ing electric load behaviors in
the harmonic dimension& adopting random process models to e.plain steady-state power
observations& combining transient and steady-state responses for the common goal of load
identification& etc. )pecifically& a real-time constant load disaggregator was developed
which can handle single load events& multiple load events and state estimations without
load events. ! 8)# load trac%er was developed for the first time in the history of NI"M.
8arious methods to estimate and eliminate harmonic signals were also developed.
Ket& there still are numerous unsolved problems to develop a commercial-grade (LI)
based on the NI"M. The issues presented in this thesis are the most important ones&
selectively chosen by the author& so that their solutions can advance the status quo of the
NI"M research significantly. These solutions lay a solid foundation& based on which the
solutions of other unresolved problems shall be developed by future researchers. This
thesis will also inspire and motivate them to come up with new approaches for coming
challenges and better methods for the problems already presented in this thesis. The
information has been and will be the %ey in developing various algorithms and will
remain to be the %ey motivation in developing NI"M-based applications.

H.3 (uture ?ork
Many issues need to be addressed in coming years to ma%e the (LI) based on the NI"M
more complete and practical. )ome of the issues were conceived in the beginning of the
research but the author did not have enough time and resources to handle them& and the
rest emerged along the course of the research. This section briefly outlines these issues
and proposes preliminary solutions or research directions where possible.
The current harmonic power is a great idea in the sense that it can compress most of the
information present in a current waveform to a few selected power terms. 1owever& it is
blind to the information that may be embedded in a voltage waveform. It is true that the
current waveform carries most of the information about the electric loads& but the voltage
waveform also transmits a substantial amount of information& especially about the
transient behavior of a load. 8oltage waveforms were monitored traditionally from the
power uality viewpoint& i.e. to monitor voltage dip& swell& transient& interruption& etc.
The voltage harmonic intensity can be used to uantify these power uality indices and to
monitor electric loads. Thus a future (LI) will e.tract richer information from both the
voltage harmonic intensity and the current harmonic power generated by the NI"M& and
ma%e better decisions concerning the attributes of the monitored electric loads.
The current harmonic powers are computed via the #+T of current waveform samples&
following the method presented in -hapter 3. !s discussed& it generally is difficult to
obtain a consistent number of evenly spaced samples per cycle& due to the generator-
sampler synchroni2ation problem. It is currently solved by oversampling& followed by
filtering and resampling. 1owever& this approach inevitably introduces errors& such as the
phase biases of the harmonic powers due to the sampling position uncertainties. The
current harmonic powers will continue to be computed from the #+T. Thus& it is
necessary to develop a more effective solution to this synchroni2ation problem.
The Daussian random process model was introduced in this thesis to describe the steady-
state power observations of constant loads in clear mathematical terms. The Daussian
model was decisive in developing various estimation routines for constant load
disaggregation& and the resulting estimators were very effective. 1owever& it is still
limited in the sense that a constant load can have slightly different means at different
instances and its noise characteristic can change with time. ! better stochastic model will
be reuired to address this time-varying nature of a constant load. !lso& the steady-state
power model does not necessarily have to be stochastic. )omeone may come up with a
better non-stochastic model in describing the steady-state power observations.
The transient response of a constant load was the other half in developing the load
disaggregator. 1owever& the statistical properties of responses were not derived e.plicitly
and only the simple least suares estimator was used to classify patterns. It obviously is
true that a better classifier can be designed if the statistical distributions of the transient
patterns are %nown. +uture classifiers may have a training period to collect enough
information about the pattern statistics and& based on the statistical characteri2ation& come
up with a classification metric& such as the ma.imum-li%elihood.
! multi-event handler was developed in this thesis based on the event space search and
the transient pattern composition. The event space search can not be said to be optimal&
because it relies on the heuristic search of a binary tree to visit possible state candidates.
Nonetheless& it usually does not impose too much a computational burden and thus is
acceptable. 1owever& the transient pattern composition is not a polynomial algorithm&
because of the permutation of the pattern seuences& although dynamic programming was
used to reduce the computational comple.ity significantly. 8arious heuristic methods will
have to be developed to further reduce the comple.ity. !lternatively& the problem may be
formed from a totally different viewpoint& which may do away with the non-polynomial
algorithm.
The constant load disaggregation scheme based on both steady state and transient was
very effective in distinguishing similar loads& i.e. /4 hp water pumps. 1owever& there are
situations where two loads are e.actly identical and where similar loads have to be treated
as identical for practical reasons& i.e. when there are too many individual loads but they
can be categori2ed to several distinct groups. 1ow to design a constant load disaggregator
under this situation is an unresolved challenging but interesting problem.
/04
! constant load has only two states7 @N and @++. 1owever& there are loads that consume
a fi.ed amount of power at a state but the number of states is more than two. #ishwashers
and laundry machines are good e.amples. (ach unit has an assembly of smaller electric
devices& such as motors& heaters& valves& etc. #epending on which device is on during the
course of its programmed use& it consumes different amounts of power at different stages
while it is @N. They can be called multi-stage loads. ! multi-stage model is reuired to
classify load events& which can update the stages of relevant loads and thus determine the
states of electric loads (and observation).

! very simple 8)# trac%er was implemented in this thesis& using the correlations
between the fundamental powers and higher harmonic powers. !lthough effective& it
sometimes shows a poor estimation result (within about /4 O of true value)& due to the
time varying nature of the correlations. +uture improvements may include collecting
different correlation curves and ma%ing the trac%er to ma%e an educated decision which
curve to choose based on observations. The 8)# power estimation from higher
harmonics can become more accurate by employing a time-varying representation of
correlations and an in situ selection of the most li%ely correlation.
The proposed 8)# trac%er assumes that there e.ists only a single set of 8)#s. It is an
interesting problem how to develop a power estimator for multiple 8)# sets that can
estimate powers either collectively or separately for each set. :hen multiple 8)#s are
running& there e.ist certain constraints for the composition of the overall harmonic
powers due to the characteristic of each 8)#. The observed harmonic powers may be
decomposed& using the constraints& into individual harmonic powers generated by each
8)# and thus the corresponding fundamental powers may be estimated.
The feasibility of the multi-8)# power trac%ing method will depend in large part on the
availability of a new power coder with a sufficiently small phase bias. The proposed
constraints in harmonic powers will dictate necessary phase precision. @n the other hand&
the phase accuracy is necessary even to observe the presence of the constraints and to
estimate them reliably.
1ow to handle non-8)# variable loads is also an interesting problem. They may have
signatures in a higher harmonic domain& although constant& as was the chiller at the
)ansome building. These signatures may be used for detection purposes. If no good
method is available to trac% their individual power consumptions& their collective power
consumption& at least& can be estimated by subtracting the powers consumed by the
constant loads and 8)#s from the observations.
8arious estimators and classifiers reuire a priori information of the signal& such as mean
and variance& to ma%e necessary decisions. In this thesis& the statistical parameters were
obtained separately from manual analyses. 1owever& a commercial grade (LI) will be
reuired to collect a priori information by itself from observations& to minimi2e labor-
intensive human interventions.
This thesis has concentrated on the load disaggregation aspect of the (LI) based on the
NI"M. @ther possibilities were not specifically investigated& such as the diagnosis& power
uality monitoring& optimi2ation and power management. To ma%e the NI"M a more
competent and diverse tool& the researches to materiali2e these applications should be
pursued in future. :ith relentless endeavors& the NI"M will soon become a ubiuitous
electric load information system.
/03
Appendi@ A
:eneral Change!o'!meanK#ariance Detector
! change-of-mean detector identifies the presence of a change-of-mean event and
estimates the location of the event and the mean and variance after the event (and before
the event if necessary)& from a window of data. The detector may be able to identify
multiple events or a single event depending on the construction. The change-of-mean
detector performs an abstraction on the incoming data and groups them as segments& each
of which is identified by a uniue mean and variance.
In the conte.t of the steady state load classification& the change-of-mean detection can be
performed on a single stream (e.g. or ! only)& multiple streams or a composite uantity
of the current harmonic powers. 1owever& in most cases& it is suitable to use the stream
only& because virtually all loads consume real power and because any turn-on or turn-off
event will accompany a change in the real power level.
The design of a change-of-mean detector depends mostly on whether the location of
change and the mean and variance of the segment before the change are %nown. It is
natural to assume that the location of change is not %nown& and furthermore that whether
there is a change-of-mean event also has to be tested.
! simple binary hypothesis test based on a generali2ed li%elihood ratio (DL,) can be
designed to detect the presence of a change-of-mean event& when the mean and variance
of a prior-event segment is %nown and the variance does not change after the event F#.
Luo& 344/G >asseville and Ni%iforov& /<<=I. The prior-event mean and variance need to
be estimated from a pre)event window& and thus they depend on the si2e and location of
the window.
If one tries to detect a steady-state event by suppressing transients with adeuate
downsampling and filtering& the aforementioned DL, test can be used with a careful
choice of pre-event and detection window lengths and mean and variance estimation
techniues. 1owever& it is restrictive in the sense that it can detect only a single change-
of-mean event& that the mean and variance before the event should be %nown& and that it
can not detect a change-of-variance event.
In this chapter& a general change-of-meanEvariance detector is developed. The detector is
general in the sense that it can detect an arbitrary number of mean andEor variance
changes from a given data window (with an upper limit) and that it does not reuire any
prior information (i.e. means and variances) to perform its 5ob.
The following composite multiple hypothesis test can be conceived for an N-point
seuence &FnI7

&
I) F var( / & & I F
I) F var( / & & I F
I) F var( / & & 4 I F
I F 7
.
I) F var( / & & I F
I) F var( / & & 4 I F
I F 7
I) F var( / & & 4 I F I F 7
ma.
3
/
3
/ / 4 /
3
4 4 4
3
/ 4 /
3
4 4 4
/
3
4 4 4
#
# # #
#
$
n w q N n n n w "
n w q n n n n w "
n w q n n n w "
n & $
n w q N n n n w "
n w q n n n w "
n & $
n w q N n n w " n & $
.
.
.
.
.
.
.
.
.
o
o
o
o
o
o
= = +
= = +
= = +
=
= = +
= = +
=
= = + =

(!./)
where #ma& is the ma.imum number of meanEvariance changes tested for the given si2e N
window and q (q > 4) is the number of transition points to be ignored between two
segments. p is the minimum length of a segment (p > 3). In general& if j parameters are to
be estimated from a segment& p > j. p and q should satisfy the following relationship7 (#ma&
9 /)(p 9 q) s N. "i is an un%nown but constant mean of segment i. The 2ero-mean
additive white Daussian noise (:DN) wFnI has an un%nown but constant variance oi
3
at
the segment i. n#-/ is the starting inde. of the segment # (with n-/ R 4 and
ma.
#
n
R N).
+or a hypothesis $i& un%nown parameters are
& I F and I F & I F
/ 4 / / 4 / 4
T
i i
T
i i
T
i i
n n n " " " . . . = = =

L n A (!.3)
where j R oj
3
. ! single parameter vector can represent these parameters7
| | .
T
T
i
T
i
T
i i
L n A J = (!.=)
If li denotes the number of un%nown parameters for $i (or the length of ui)&
. 3 = + = i l
i
The probability density function ("#+) of @ given $i is denoted by p(@G ui Y $i). The "#+
can not be evaluated due to the lac% of the %nowledge of ui.
! typical multiple hypothesis test chooses the hypothesis for which the conditional
density p(@ Y $i) is ma.imum& which is %nown as the ma.imum li%elihood (ML) rule.
1owever& it generally is impossible to obtain the conditional density e.actly
& ) ( ) & Y ( ) Y (
i i i i i
d p $ p $ p J J J @ @
}
= (!.?)
/0?
because the integration is usually difficult to perform and reuires the %nowledge of ui.
Instead& an appro.imation to the original ML rule is widely used F$ay& /<<CI7

( ). )
U
( det ln ) Y
U
G ( ln where
& ma. arg
U
3
/
i i i i
i
$
#
$ p
$
i
J , J @ =
=

(!.B)
i
J
U
is the ML estimation of
i
J
assuming $i is true. ,() is the +isher information matri.
F"apoulis and "illai& 3443I. The above criterion is often referred to as the generali,ed .*
rule.
>ecause the evaluation of the +isher information is rather involved& the test statistic i of
the generali2ed ML rule is appro.imated by the minimum description length (M#L)
approach F,issanen& /<ACI7
. ln ) Y
U
G ( ln
3
/
N l $ p
i i i i
~ J @ (!.0)
The first term of i is an estimated li%elihood of the data given $i and the second term is
the penalty factor for having to estimate un%nown parameters. The penalty factor
increases linearly with the number of un%nown parameters and logarithmically with the
data length. :hen more parameters are to be estimated from a given data& the credibility
of the li%elihood decreases& the penalty factor punishes this tendency.
The q points per segment to be ignored are treated as un%nown deterministic uantities&
and thus are eliminated in computing the li%elihood. In turn& li is increased to
accommodate the number of these points
. ) / ( 3 = q i i l
i
+ + + = (!.A)
#ue to the assumption of additive :DN (!:DN)& ) Y
U
G (
i i
$ p J @ is easily e.pressed as a
Daussian "#+7

( )
& ) ( ) (
3
/
e.p
) det( ) 3 (
/
) Y
U
G (
/
3
/
3
) / (
|
.
|

\
|
=

+ i i
T
i
i
q i N i i
$ p m @ G m @
G
J @
t
(!.C)
where
T
q n N
i
q n n q n
i
i
" " " m
(
(

=


I / / / F
U
I / / / F
U
I / / / F
U
/ 4 / 4
/ 4
_
. .
_
.
_
. and
. I / / / F
U
I / / / F
U
I / / / F
U
diag
/ 4 / 4
/ 4
|
|
|
.
|

\
|
(
(

=


T
q n N
i
q n n q n
i
i
_
. .
_
.
_
. G
diag(a) is a diagonal matri. with the
diagonal of a and 2ero off-diagonal terms.
j
"
U
and
j

U
are the ML estimations of the mean
and variance of the segment j F$ay& /<<=I& i.e.

=

=
q n
n n j j
j
j
j
n &
q n n
"
/
/ /
I F
/
U
and
( ) .
U
I F
/
U
/
3
/ /

=
q n
n n
j
j j
j
j
j
" n &
q n n

(!.<)
The test statistic (!.0) becomes
( )
. ln
3
/
)
U
I F (
U
/
)
U
I F (
U
/
)
U
I F (
U
/
3
/
U
ln ) (
U
ln ) (
U
ln ) (
3
/
3 ln
3
) / (
/
3
/
3
/
/
/
4
3
4
4
/ / 4 / 4 4
/
/
4
N l " n & " n & " n &
q n N q n n q n
q i N
i
q N
n n
i
i
q n
n n
q n
n
i i i
i
o

|
|
.
|

\
|
+ + +
+ + +
+
=


=

=

=

t
.
.
(!./4)
Ma.imi2ing i is eual to minimi2ing
i i
3 = '
and now (!.B) becomes& with the new
test statistic
i
'
&

& min arg
U
i
$
#
i
$ ' =
(!.//)
where
( ) ( )
. ln )
U
I F (
U
/
)
U
I F (
U
/
)
U
I F (
U
/
U
ln ) (
U
ln ) (
U
ln ) ( 3 ln ) / (
/
3
/
3
/
/
/
4
3
4
4
/ / 4 / 4 4
/
/
4
N l " n & " n & " n &
q n N q n n q n q i N
i
q N
n n
i
i
q n
n n
q n
n
i i i
i
o
+
|
|
.
|

\
|
+ + + +
+ + + + + = '


=

=

=

t
.
.
(!./3)
>ecause n
i
is un%nown& the test statistic can not be readily evaluated. >ecause there e.ists
no e.plicit formula to determine ML indices& each nj needs to be found by enumerating all
possible change indices. It generally reuires '(N
i
) evaluations of
i
'
under $i. #ue to
the large computational load& the enumeration scheme is not feasible e.cept for a small i.
@ne approach to reduce the computational load in finding change times is to use the
techniue of dynamic programming (#") F$ay& /<<CI. #" in general can reduce the
computational comple.ity down to '(iN)& i.e. an e.ponential comple.ity is changed to a
polynomial comple.ity.
:e note that the first and last terms of the test statistic do not depend on change times.
Thus the ML change times can be found by minimi2ing the following cost function7
/00
( )
& )
U
I F (
U
/
)
U
I F (
U
/
)
U
I F (
U
/
U
ln ) (
U
ln ) (
U
ln ) ( ) &
U
&
U
(
/
3
/
3
/
/
/
4
3
4
4
/ / 4 / 4 4 &
/
/
4
|
|
.
|

\
|
+ + + +
+ + + =


=

=

=

q N
n n
i
i
q n
n n
q n
n
i i i i i
i
o
" n & " n & " n &
q n N q n n q n 0


.
. n L A
(!./=)
i.e.
) &
U
&
U
( min arg U
& i i i i
0
i
n L A n
n
=
.
The following additive cost function is defined7

. )
U
I F (
U
/
U
ln ) ( I / & F
/
3
/ /
/

+ = A
q n
n n
j
j
j j j j j j
j
j
" n & q n n q n n

(!./?)
The overall cost function is simply the sum of additive costs

. I / & F ) &
U
&
U
(
4
/ &
=

A =
i
j
j j j i i i
q n n 0 n L A
(!./B)
The following intermediate optimal cost up to the segment # and inde. * is defined7

& I / & F min I F
4
/
/ & 4
& & &
/
/ / 4

=

+ + = =
A =

#
j
j j j
q * n n
n n n
#
q n n * I
#
#
.
(!./0)
where n#-/ 9 (p /) s *. Note that the minimum of the overall cost function is Ii FN /
qI. It can be shown that the intermediate optimal cost function can be written in a
recursive form F>ertse%as& /<CAI

( ). I & F I / F min I F
/ / /
/
* n q n I * I
# # # #
n
#
#

A + =

(!./A)
The euation (!./A) says that the intermediate optimal cost up to the segment # and inde.
* is the minimum of the sum of the intermediate cost up to the segment # J / and inde.
n#-/ J / J q and the additive cost between the indices n#-/ and *.
/
U
#
n & the starting inde. of
the segment # at which I#F*I becomes a minimum& is found during the course of the
minimum search.
:ith the aid of the dynamic programming&
i
'
& the test statistic for $i& is computed via the
following procedure7
/. Let # R 4 and I
J/
FI R 4. +or * R (p /)& (p /) 9 /& L & (N /) i(p 9 q)& compute

& )
U
I F (
U
/
U
ln ) / ( I & 4 F I F
4
3
4
4
4 / 4 4
=

+ + = = A =
*
n
" n & * * n * I

where

=
+
=
*
n
n &
*
"
4
4
I F
/
/
U
and

=

+
=
*
n
" n &
*
4
3
4 4
)
U
I F (
/
/
U
.
3. Let # R /. +or * R (p /) 9 #(p 9 q)& L & (N /) i(p 9 q) 9 #(p 9 q)& compute

( ). I & F I / F min I F
/ / /
) / ( ) / ( ) )( / ( ) / (
/
* n q n I * I
# # # #
p * n q q p # p
#
#

s s + + + +
A + =

!lso& for each *& determine

/
U
#
n at which I#F*I becomes minimum and call it n#-/F*I.
=. Let # # 9 / and iterate until # R i.
?. Let # R i. +or * R (N J /) J q& compute

( ). I & F I / F min I F
/ / /
) / ( ) / ( ) )( / ( ) / (
/
* n q n I * I
# # # #
p * n q q p # p
#
#

s s + + + +
A + =

!lso& find I / F U
/ /
q N n n
i i
=

& i.e. the start inde. of the last segment.
Now Ii FN / qI is the desired minimum of the overall cost function and
i
'
can be
computed by adding two remaining terms& i.e.

( ) . ln 3 ln ) / ( I / F N l q i N q N I
i i i
+ + + = ' t
(!./C)
The change times are found by the following bac%ward recursion7

. I / U F U
I / U F U
I / F U
/ 4 4
/ 3 3
/ /
q n n n
q n n n
q N n n
i i i
i i
=
=
=


.
(!./<)
@nce the change times are found& the ML estimations of the mean and variance of each
segment are easily obtained by using (!.<).
The test statistic
i
'
is computed for each i& i e Q4& /& L & #ma&P& and an $i which gives a
minimum test statistic is selected from the composite multiple hypothesis test. Note $i
indicates that there e.ist i meanEvariance changes for the given data of length N and (i 9
/) constant meanEvariance segments. The start inde.& mean estimation and variance
estimation of each segment are obtained as byproducts of the hypothesis test.
The computational time to complete the hypothesis test increases substantially with #ma& (
( )

=
ma.
4
3
T
#
#
# N ' ). Thus a reasonable number should be chosen for #ma&& considering the
length of the data record and the relative freuency and nature of load activities.
/0C
!s an e.ample& consider the test signal shown in +igure !-/ with N R /44. The original
signal sFnI is composed of three segments with start indices at 4& 34 and ?4. There e.ist
two transition points between the segments. sFnI is corrupted with an additive 2ero-mean
:DN whose variance is / and changes to /0 starting at n R C4. Thus the resulting signal
&FnI R sFnI 9 wFnI has four segments& the last of which is a pure variance change. +or the
test& following parameters are used7 #ma& R ?& p R B and q R 3.

0 10 20 30 40 50 60 70 80 90 100
0
1
2
3
4
5
s
[
n
]
0 10 20 30 40 50 60 70 80 90 100
-10
-5
0
5
10
sample number, n
x
[
n
]
s[n] corrupted with zero-mean WGN
+igure !-/ Test signal for change-of-mean detector (Top7 original signal& >ottom7
corrupted by 2ero-mean :DN).
+igure !-3 shows the test statistic computed for each hypothesis. The test statistic clearly
becomes minimum at i R =& and it successfully detected three meanEvariance changes
present in the test signal.

0 0.5 1 1.5 2 2.5 3 3.5 4
360
380
400
420
440
460
480
500
520
test statistic to detect number of mean/variance changes, min selected
hypothesis number (# of mean/var changes)
x
i
+igure !-3 Test statistic for each hypothesis with the signal of +igure !-/.
The following is the screen dump of the change-of-mean detector output7
n estimation:
21 40 82 100
A estimation:
-2.4928e-001 5.4169e+000 -2.4563e-001 3.0583e-001
var estimation:
1.1539e+000 6.0716e-001 7.6817e-001 1.6012e+001
The change time estimates are reasonably accurate considering two points are dismissed
between the segments. The resulting mean and variance estimations are also acceptable.
:hen the change-of-meanEvariance detector is used in the conte.t of the steady-state load
classification& it has to be applied to the continuous stream of a selected power. !
window of si2e N is applied to the stream to identify change-of-meanEvariance events&
then the window is shifted by * and the process continues. -onsidering the fact that the
boundary of a window can be placed arbitrarily with respect to a turn-onEoff response& it
is necessary to allow a small amount of overlap and ad5ust the overlap si2e& if necessary&
while %eeping * fi.ed.
The general change-of-meanEvariance detector developed in this chapter can detect an
arbitrary number of meanEvariance changes for given data. (.cept for test parameters&
such as the ma.imum number of changes& the minimum length of a segment and the
number of transition points to be ignored& it does not reuire any prior information about
the data. It detects the number of changes& and estimates the change inde.& mean and
variance of each segment. It converts the stream of a power into the series of constant
meanEvariance segments.
The computational load was reduced substantially by the introduction of dynamic
programming. 1owever& computations can still be a problem when a large #ma& is tested. If
#ma& is set to one& the detector functions similar to the conventional DL, detector& and also
with a comparable computational cost.
/A4
Appendi@
Air 6andling "nit *odeling
!n air handling unit (!1;) is designed to provide thermal comfort and ventilation to the
occupant in a building. The !1; consists of fans to induce airflow& heating and cooling
coils to deliver thermal energy& dampers to regulate air flow& thermostats to control
temperature& ducts to provide air passage and building space. There e.ist many varieties
of !1; systems. 1owever& they can be classified mainly either as a constant volume
(-8) system or a variable-air-volume (8!8) system. ! -8 system typically has a single
supply fan& whereas a 8!8 system has both a supply and a return fan to meet the varying
air flow load FDraf and Lee& /<<CI. >ecause most commercial buildings have 8!8
systems& we will concentrate on their modeling and analysis.
.1 ;A; model
+igure >-/ shows the schematic of the 8!8 system. The combination of heating and
cooling coils transfers energy to the airflow reuired to meet the thermal load of a
building. The supply fan delivers the conditioned air to each 8!8 terminal bo. through
the supply duct. (ach 8!8 bo. regulates the airflow& by varying its damper position& to
maintain a constant set temperature of the space or 2one being served. The return fan
returns air from the space to the mi.ing 2one.

2one
T
2one
T
2one
T
relief
air
damper
returnEe.haust
fan
outside air
damper
return air
damper
heating coil
cooling coil
supply fan
Thermostat controls temperature
by varying air volume
delivered by the 8!8 bo.
"
"ressure regulator
controls supply fan
capacity to maintain
constant duct pressure.
8!8 bo.
8!8 bo.
8!8 bo.
"
"ressure regulator
controls return fan
capacity to maintain
constant building
static pressure.
+igure >-/ )chematic of variable-air-volume system.
#epending on the position of the three dampersJ the outside air& return air and relief air
dampersJ 4 to /44 O of the return air can be recirculated to the supply fan. +or e.ample&
if the outside air and relief air dampers close fully and the return air damper opens fully&
the system recirculates /44 O the return air. If the dampers are in the opposite positions&
the return air is fully vented to the environment& and /44 O outside air is supplied to the
mi.ing 2one. The fan capacity is controlled to maintain a constant pressure at a
designated position& either by varying speed or varying inlet damper position. +or
e.ample& both fans can be controlled to maintain a constant supply duct pressure& or the
supply fan maintains a supply duct pressure and the return fan maintains a building static
pressure.
:e present a very simple model of the 8!8 !1; system. +igure >-3 shows the
schematic of the model.

es
er
,/
L/
,3
L3
,=
-/
"s X/ "/ X3
supply
fan
supply
duct
building
static
return
duct
return
fan
es7 pressure created by supply fan
er7 pressure created by return fan
"s7 supply duct pressure
"/7 building static pressure
X/7 flow rate along supply duct
X37 flow rate along return duct
L/7 inertance (inductance) of supply duct
L37 inertance (inductance) of return duct
-/7 capacitance of building
,/7 resistance of supply duct
,37 resistance of building space
,=7 resistance of return duct
+igure >-3 )imple model of !1; system.
! set of assumptions is made in developing the model7
/. The main characteristic of the air duct is described by the inductance-resistance (*-R)
model& while the building space is represented by the resistance-capacitance (R-3) model.
>ecause a duct has a very long& yet confined air passage& it is difficult to develop flow
rate (inertial effect) whereas it is relatively easy to build up pressure. It is observed in
buildings that sudden rises in pressure often damage duct systems& when fans are
accelerated too rapidly. ! gradual ramp up of a variable speed fan& during start up& is
reuired to prevent an abrupt pressure rise and hence duct damage. @n the other hand& the
building space has a much larger volume. Thus it is relatively easy to establish flow rate&
but it is difficult to increase the space pressure (capacitance effect). This capacitance
effect is the result of the inherent air compressibility& though minor. @f course& the duct
has a certain amount of capacitance and the building space has an inertance. >ut they are
not the dominant characteristic of each system and ignored in this model. @n the other
hand& resistance is present in every system.
/A3
3. The presence of dampers (relief air& outside air and return air) is ignored for simplicity.
In this configuration& the supply and return fans are connected directly to the
environment. They are open to atmospheric pressure. This is the same as the /44 O
outside air supply case (return damper closesG relief and outside damper open). To add the
dampers to the model& the supply and return fans should be connected with a variable
resistance and with two tees& with their own varying resistances& to the atmosphere. The
influence of the dampers will be considered in future models. >ecause the reference
pressure is atmospheric& all the pressures in this model are gauge pressures (negative for
vacuum).
=. The time-varying nature of the system is ignored during the development. The system
has an inherent time-varying nature. +or e.ample& the resistance of the building static
varies depending on loads. If there are more thermal& or ventilation& loads& 8!8 bo.es
will move dampers towards the fully open positions& hence reducing the resistance to the
air flow. @r& a branch of the supply duct may be closed if the corresponding space of a
building is not in use& increasing the duct resistance. The inertance and the capacitance of
the system are believed to remain relatively unchanged. :e describe the system by first
assuming a time-invariant nature. Later& we will discuss the effect of the time variance.
?. The duct and building space networ% are simplified. The duct system is not a single air
stream passage. It typically divides into many branches& each of which serves a different
portion of a building. The return duct 5oins many branches and forms a trun% before it
reaches the return fan. The same is true in the building space. It is divided into many
floors and 2ones& providing multiple passages for the air flow. 1owever& the model
describes the networ% in an average sense.
B. #ynamics of the fans and motors are ignored. The duct and space air passage system
has more inertance and hence a slower time constant than the fans and motors. The
relatively fast dynamics of these devices are ignored for simplicity. +ans are treated as
effort& or pressure& sources that instantly create pressure rises. ! more sophisticated
model may include the dynamics of fans and possibly of motors.
0. !ir infiltration is ignored. In the model& the air delivered by the supply fan& minus any
amount stored by the building capacitance& is received by the return fan without any loss.
In real buildings& depending on the pressure difference& there e.ists an air lea% to or
infiltration from the environment. The effect of infiltration can be modeled by adding a
tee with a resistance to the building static.
The bond graph of the described 8!8 !1; model is shown in +igure >-=. >ond graphs
are often used to describe and model comple. dynamic systems. They have the advantage
of showing the inherent nature of a system& such as power flow& causality& state euation&
etc.& in a simple compact graphical format F$arnopp et al.& /<<4I.

4 / /
es er
L/ L3
,/
,3
,=
-/
)e )e
.. ..
.. ..
..
X/ X3
"/
+igure >-= >ond graph of !1; system model.
! state space euation describing the system can be obtained from the bond graph. In
bond graph terms& each energy storage element& * or 3& with an integral causality (which
simply means that it maintains its own uniue characteristic even interacting with other
elements) contributes a state variable. >ecause we have three energy storage elements
with integral causality in the system& we have three state variables. (ach variable& namely

/
2 !
/
and !
3
& is related to its corresponding energy storage element. The system is
described by the following state space euation7

.
/
4
4
/
4 4
4
/
4
/
/ /
4
3
/
3
/
/
3
=
3
/
3 /
/
3 /
3
/
/
(

(
(
(
(
(
(
(

+
(
(
(

(
(
(
(
(
(
(

=
(
(
(

r
s
e
e
*
*
!
!

*
R
*
*
R R
*
3 3
!
!

`
`
`
(>./)
er is the pressure created by the return fan (thus negative). )er gives a positive vacuum
pressure. 1owever the above euation is not convenient in the sense that the flow rate is
not readily measurable. Thus& we introduce the supply duct pressure s in place of !
/
using the relationship7 s 7 R
3
!
/
8
/
. The new state space euation is given by7

=
+ =
C@ y
u A@ @`
& (>.3)
where C R , (, is the identity matri.) and y is the output of the system. The state vector&
input vector and coefficient matrices are given by the following formulae7
/A?

| | | |
.
/
4
4 4
4
and
/
4
/ / /
/
)
/
( )
/
(
& &
3
/
3
3
=
3
/ 3 / 3 /
/ 3 / /
/
/
3 /
3 /
3 /
(
(
(
(
(
(

=
(
(
(
(
(
(
(

=
= =
*
*
R
*
R
*
3 R 3 R 3
3 R 3 *
R
*
R R
R 3
e e !
T
r s
T
s
A
u @
(>.=)
The characteristics of the system can be e.amined by obtaining the open-loop response of
the system. #efine the open-loop transfer function :(s) R 9(s)E"(s)& where 9(s) and "(s)
are the Laplace transforms of y and u& respectively. :(s) is obtained by the matri.
euation& :(s) R C(s, J A)
-/
& where s is the Laplace operator. !fter wor%ing out all the
algebraic details& :(s) is given by7

.
/ / /
/ /
/ / / /
) det(
/
) (
) (
) (
) (
) (
) (
) (
) (
) (
) (
) (
) (
) ( ) (
) ( ) (
) ( ) (
) (
/ / /
3 / 3
3 3 3 / /
3
/
3 /
/ 3 3
=
3 / /
3
/ /
/
/ 3 / 3 /
3
3
=
3 / /
3
3 3
/ /
=3 =/
33 3/
/3 //
(
(
(
(
(
(
(
(

|
|
.
|

\
|
+
+
+
|
|
.
|

\
| +
+
|
|
.
|

\
|
+
|
|
.
|

\
|
+ +
|
|
.
|

\
|
+
|
|
.
|

\
|
+

=
(
(
(
(
(
(
(
(

=
(
(
(

=
* 3
s
*
R R
s
* * R 3 *
R
*
R R
s
3 * *
R
s
R 3 *
R
* 3
R
s
3 * 3 * *
R
*
R
s
R 3
s
*
R
s
s e
s !
s e
s !
s e
s
s e
s
s e
s
s e
s
s 6 s 6
s 6 s 6
s 6 s 6
s
r s
r s
r
s
s
s
A ,
:
(>.?)
The determinant is given by the following formula7

.
/ / ) (
) det(
3 / /
= 3 /
3 / / / 3 /
= 3 / 3
/
3 /
3
= =
* * 3
R R R
s
* 3 * 3 * *
R R R
s
*
R R
*
R
s s
+ +
+
|
|
.
|

\
|
+ +
+
+
|
|
.
|

\
| +
+ + = A ,
(>.B)
The closed-loop response depends on what type of control logic is used. +or simplicity&
we assume a simple proportional control with an offset. )pecifically& we control the
supply-fan pressure to maintain a constant duct pressure and the return-fan pressure to
maintain a constant building static pressure.

&
4
4 4
4 4
) (
3 3
/ /
3
/
(
(
(

+
(
(
(

= + =
r
s
r
r
s sr
m
m
! !


#
#
m @ r 2 u
(>.0)
where r is the reference& or control set point& and m is the offset vector. 2 is the gain
matri.. In this case we do not feed bac% the flow rate& which is difficult to measure& to
control the fan pressures. +igure >-? shows the bloc% diagram of the control system.

@ M A@ N u
.
C 2
r
N
!
N
N
m
u @
y
+igure >-? >loc% diagram of !1; control system.
The closed-loop system response is given by7
)& ( ) ( ) ( ) ( ) ( s s s s s * 6 - : 9
c
+ = (>.A)
where
2 2&O %A , CP
-
9
:
1
c

= = s
s
s
s
) (
) (
) (
and
2&O %A , CP
*
9
6
1
= = s
s
s
s
) (
) (
) (
.
9(s)& -(s) and *(s) are the Laplace transforms of y& r and m& respectively.
:ith this type of control logic& the open-loop transfer functions 6
//
(s) and 6
33
(s) play
important roles in determining the closed-loop response of the system. (ach transfer
function describes the behavior of the state variable as a function of the corresponding
control input& fan pressure& which in turn is used to drive the variable to the desired
reference point.
The /04 )ansome building !1; is controlled in almost the same way as we described
the closed loop control system& i.e. the supply-fan speed is controlled to maintain a
constant supply duct static pressure and the return-fan speed is controlled to a constant
building static pressure. The following is the control logic of the building.

+ =
+ =
= . /C ) ( 3?
3B . /< ) ( /4
/ /
u
u
r r
s sr s
& (>.C)
/A0
where us and ur are the velocity command inputs to the supply and the return fan 8+#&
respectively (? J 34 m!). "ressures are measured in inch 1
3
@. The presence of large
offsets& close to the upper limit& is noteworthy. These offsets serve as the open loop inputs
to the system& deriving the pressures& hence the state variables& to certain #- values&
which are usually close to the desired set points. The proportional control with a feedbac%
is added to the control input to regulate the pressure to a desired set point. 1owever& the
control logic of the building is not the same as the control law of the model (>.0). In the
model& u is the pressure input vector to the system. Thus appropriate gains& and the
transfer functions of the fan system should be multiplied to the control logic to obtain an
euation applicable to the model.
.2 Sustained oscillation! limit cycle
It will be helpful to glimpse the behavior of the variables of the real !1; system& from
the /04 )ansome (nergy Monitoring and @ptimi2ation ((M@) website. #ata are
displayed in real time using the (lectric (ye software. The minimum reliable refreshing&
or sampling& interval is B second.

0 50 100 150 200 250 300
1.4
1.45
1.5
1.55
1.6
Supply duct static pressure, 12/03/01, 15:00
i
n
c
h

H
2
O
time (sec)
0 50 100 150 200 250 300
18.5
19
19.5
20
SF control command (4-20 mA)
m
A
time (sec)
+igure >-B )upply duct static pressure and )+ control command.
+igure >-B shows an e.ample of the behavior of the state and control variables of the
!1;& recorded from the website. +rom the figure& it is clear that the supply duct pressure
has a steady oscillation with a period of =4 sec& the same fan oscillation period observable
from the stream. !lso we can infer that the supply duct pressure set point is /.B inch
1
3
@& because it is the average value of the measurement. The )+ control signal is /C4
out of phase with the duct pressure (the pressure is multiplied by J/ in the control logic).
The presence of a sustained oscillation is an interesting problem. In linear system theory&
a feedbac% system can show an oscillation at steady state if it has poles located on the
imaginary a.is in the comple. plane. 1owever& if there is a slight disturbance& the pole
can move to the right half comple. plane& meaning that the system becomes unstable
(oscillation %eeps growing in amplitude). Thus& there is no safety net to allow a sustained
oscillation in a linear system. @n the other hand& it is unreasonable to believe that the
control engineer places the system poles e.actly on the imaginary a.is after many tuning
attempts& even accidentally. Thus& linear system theory alone can not e.plain the
sustained oscillation behavior of the feedbac%-controlled system such as the !1; system.
)omething is nonlinear. @therwise& the behavior can not be e.plained. There can be many
sources of nonlinearity. ! duct may have an inertance that is a function of flow rate and
pressure& rather than a constant. There possibly can be a transport delay in the control
loop. >ut& we eye on the actuatorsJ the )+ and the ,+. They can saturate.
!1; fans are typically selected& in terms of capacity& based on their steady state load.
They are seldom selected to meet the transient load& or acceleration. In a sense& !1;s are
underpowered. It is uite possible that if there is a changing demand& the actuators
become saturated because controllers demand more than they can produce& hence entering
a nonlinear region. (In modern computer controlled systems& output from a control logic
is already limited before it is given to an actuator. In this case& it is the saturation of the
control logic& but the result is the same). This is plausible for the /04 )ansome !1;
system& because both fans& especially the )+& are operated close to their upper limits.

$
r
9
-
y
D(s)
e u v
v
u
a
N
-a
-N
saturation effective gain of saturation bloc%
#
NEa
a
YvY
input magnitude
g
a
i
n
feedbac% control system with saturation
+igure >-0 (ffect of saturation in feedbac% control system.
+igure >-0 shows the effect of saturation on a feedbac% control system F+ran%lin et al.&
/<<?I. :hen input to the saturation bloc% is less than a threshold value& the bloc% gives
an output proportional to the input (gain # R NEa& slope of the saturation curve). Note the
overall gain of the system is $#. 1owever& if the input magnitude e.ceeds the threshold
value& the output is no longer proportional to the input. The output is the constant limit
value and the effective gain is reduced to NEYvY. !s the magnitude of input grows& the
effective gain diminishes. The behavior of the effective gain with varying input
/AC
magnitude is depicted in the graph. The feedbac% control system has a varying gain that
depends on the magnitude of the error signal (e). :hen the error signal is small& the
system has a constant gain. >ut as the magnitude of the error signal grows the effective
gain of the system is reduced. (+or e.ample& the system shows a high gain at its steady
state& but a low gain during acceleration or deceleration.)
+igure >-A shows an e.ample of system with oscillation due to saturation. The root locus
of the linear portion of the system is also shown in the picture. If we set the gain $ R 4.B&
the root locus shows that the system has two unstable closed-loop poles in the right half
plane (,1"). Thus the system should be unstable and the output y will diverge to infinity.
1owever& the system shows an oscillatory behavior in the limit. +igure >-C shows the
result of the unit step response simulation of the system including the effect of saturation.
!fter the initial ramp up& it starts to show an oscillatory behavior. The oscillation builds
up to a fi.ed amplitude and stops growing. The freuency is roughly / radEsec (cross-over
points at the root locus are 4.<</j).

$
r
9
-
y
e u v
/ 3 . 4
/
3
+ + s s s
/
4./
4./
+eedbac% system with an oscillatory mode due to saturation
-4.B -4. ? -4.= -4. 3 -4./ 4 4./ 4. 3 4.=
-/.B
-/
-4.B
4
4.B
/
/.B
,eal !.is
I
m
a
g

!
.
i
s
,e(s)
I
m
(
s
)

$ R 4.B
crossing occurs
when $ R 4./<A ,oot locus of the system
e.cluding saturation
+igure >-A )ystem with sustained oscillation due to saturation.
!lthough the system has a gain of 4.B& at the beginning of the step command the
controller is saturated due to the large amount of error. Thus& initially the system has a
very low effective gain and has poles on the left half plane (L1"). !s system builds up
output& the error reduces and the degree of saturation diminishes& increasing the effective
gain of the system. The system poles start to move towards the ,1". !s they near the
crossing points on the imaginary a.is& the system starts to show oscillatory behavior (or
they can actually 5ump into the ,1"). The oscillation builds up until the poles reach the
imaginary a.is. 1owever& once they move into the ,1"& the output amplitude starts to
increase due to instability. >ut on the other hand& the increased output reduces the system
gain due to the saturation of controller. The saturation moves the poles bac% toward the
L1". (ventually& they will settle e.actly on the imaginary a.is& resulting in a stable&
sustained& and fi.ed freuency and amplitude oscillation& called a limit cycle FDibson&
/<0=I.

0 20 40 60 80 100 120
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
step response of a limit cycle system
time (sec)
y
+igure >-C ;nit step response of system showing limit cycle.
The limit cycle is a stable euilibrium of the system response. ! disturbance may cause
the system poles to move away from the imaginary a.is& but eventually they come bac% to
the euilibrium position. In the e.ample& two poles on the imaginary a.is contribute to
the sustained oscillation& while one pole on the real a.is is responsible for the #- value
of the step response.
1owever& the argument in the previous paragraph is based on the heuristic e.planation of
the system behavior. :e need a simple analytic tool to describe the nonlinear behavior of
a system. -onsider the following single-input single-output system7

C@
A@ @
=
+ =
y
u `
(>.<)
with an output nonlinearity u 7 (y). The linear part of the system is controllable and
observable. The linear system Laplace transform function

) (
) (
) (
) (
) (
s
s !
s
s 9
s :
s 6
,eroI3s
= = ~

AO , CP
1
(>./4)
is strictly proper& i.e. (s) is a polynomial of higher order than the polynomial !(s). If y(t)
is periodic with period T& the following mth order harmonic balance condition is satisfied
F8incent and Drantham& /<<AI7
& & & / & 4 & 4 ) ( m # c j# 6 a
# #
. = = e (>.//)
/C4
where a# is a coefficient of the +ourier series of y and c# is a +ourier series coefficient of
u. They are formally defined as&

. )) ( (
/
& )) ( ( ) (
) (
/
& ) (
4
4
}

=
= = =
= =
T
j#
#
#
t j#
#
T
j#
#
#
t j#
#
d e y ;
T
c e c t y ; t u
d e y
T
a e a t y
t t
t t
t e e
t e e
(>./3)
The harmonic balance euation appro.imates the output and the nonlinearity with an mth
order +ourier series. The simplest appro.imation is to ta%e m R /& which yields the
describing function method FTru.all& /<BBI. The method is widely used to describe the
nonlinearities present in otherwise linear system.

$
r
9
-
9
9
m
u
.
D(s) /
v
u
u9
u-
v R $(r - .) 9 m
v
+igure >-< +eedbac% control system with saturation nonlinearity
Let6s consider the system shown in +igure >-<. The system has an upper saturation limit
u8 and a lower saturation limit u) for an input signal u. >oth of the limits are positive. :e
assume the e.istence of a sinusoidal plus #- level signal for &.

. ) (
3
sin
4 4

+ = = + =
t j t j t j#
#
e e
j
a
a e a t a a &
e e e
e
(>./=)
Thus the output & is described by the first order +ourier series and a
/
7 a/3j. Then the
input signal v to the saturation bloc% is given by
& sin ) (
4 4
t 5a m a r 5 v e + = (>./?)
where r< is the desired constant set point. The output from the saturation bloc% u depends
on the magnitude of v& which in turn depends on 5& r and &. There can be many
possibilities for the shape of the u waveform. 1owever we consider two e.treme cases for
simplicity.
-ase /)
+
> + u 5a m a r 5 ) (
4 4
In this case& the input to the saturation bloc% is always greater than the upper limit of
saturation. Thus the output is constant u8
. 3 / for 4 &
4
. 2 2 # c u c u e c u
#
t j#
#
= = = = =
+ +
e
(>./B)
-ase 3)

> + u 5a m a r 5 ) (
4 4
and
+
s + + u 5a m a r 5 ) (
4 4

The input to the saturation bloc% is in the linear region. Thus the output of the bloc% is the
same as the input.

. 3 for 4 &
3
&
3
& ) (
sin ) (
/ / 4 4 4
4 4
. 2 # c
j
5a
c
j
5a
c m a r 5 c
t 5a m a r 5 e c u
#
t j#
#
= = =

= + =
+ = =

e
e
(>./0)
In -ase /& u is described by the ,eroth-order +ourier series. In -ase 3& a first-order +ourier
series is enough to describe u. In an intermediate case& the input signal v is partially
saturated and the output has a clipped sinusoidal waveform& which generally reuires
more coefficients in higher orders to describe with a +ourier series.
In -ase 3& the harmonic balance euation with # R / becomes

. 4 )) ( / ( 4
3
) (
3
= + =

e e j 56 a
j
5a
j 6
j
a
(>./A)
>ecause a is non2ero (a is the amplitude of oscillation)& the term inside the parenthesis
should be 2ero. >ut readers will recogni2e the euation as nothing but the root locus
formula (a root locus is a series of points which satisfies / 9 56(s) R 4)& when the poles
are located on the imaginary a.is (s R je). The assumption of the e.istence of a sinusoidal
oscillation led to the root locus euation on the imaginary a.is. In other words& if there is
a sustained oscillation within the saturation limits& the system poles are located on the
imaginary a.is& and the oscillation freuency& in radEsec& itself is the closed-loop system
pole.
If 5 is real& / 9 56(je) R 4 can be rewritten as&
/ 9 5,eF6(je)I 9 jImF6(je)I R 4. (>./C)
Thus&
/ 9 5,eF6(je)I R 4 and ImF6(je)I R 4. (>./<)
/C3
The second euation can be used to find the freuency of oscillation and the first euation
to find the gain 5& or an un%nown of 6(s) if the gain is %nown.
:ith # R 4& the -ase 3 harmonic balance euation gives

. ) 4 (
4 ) 4 (
4
4
4
4 4
m
a
c
a
6
c 6 a
~ =
=
(>.34)
Thus if the #- value of the signal is near the set point (r
4
~ a
4
)& 6(4) is appro.imated by
the ratio of the signal #- value to the input offset.
:e have shown heuristically that a sustained oscillation can develop and e.ist stably in a
feedbac% system with saturation by means of the varying effective gain. :e have also
shown that if there is a continued oscillation& the closed-loop poles are located on the
imaginary a.is and the freuency is the magnitude of the pole.
.3 Conditions to e@hi$it limit cycle
)o far we have developed the 8!8 !1; model& and e.plained the nature of limit cycle&
which can be present in a feedbac% control system with saturation. To e.plain the
oscillatory behavior of the /04 )ansome !1; system& we have to show that there e.ists a
limit cycle for at least one of the transfer functions of :(s) (remember the matri. :(s) is
composed of si. scalar transfer functions). 1owever& this is an e.tremely difficult tas%&
because we do not %now the numerical values of the parameters. :e will simply show
and e.plain under which conditions a limit cycle may develop.
:e choose 6//(s) R s(s)Ees(s)& which describes the portion of supply duct pressure
generated by the )+ pressure input& as a candidate to show a sustained oscillation. )upply
fan pressure is the main driver of the !1; system and [pushes6 air to the duct and space
networ%. In comparison& the return fan [pulls6 air from it. ! pull system is inherently
more stable than a push system (If you tow a trailer& you %now it is much easier to haul
than bac% up). 6//(s) is a push system. The algebraic form of 6//(s) is repeated here again&
.
/ / ) (
/ /
) (
) (
) (
3 / /
= 3 /
3 / / / 3 /
= 3 / 3
/
3 /
3
= =
/ 3 /
3
3
=
3 / /
3
//
* * 3
R R R
s
* 3 * 3 * *
R R R
s
*
R R
*
R
s
3 * *
R
*
R
s
R 3
s
*
R
s e
s
s 6
s
s
+ +
+
|
|
.
|

\
|
+ +
+
+
|
|
.
|

\
| +
+ +
+
|
|
.
|

\
|
+
|
|
.
|

\
|
+
= =
(>.3/)
:hen the loop is closed by feeding s bac% to the controller with a gain #& the poles of the
closed loop system are located on / 9 #6//(s) R 4& i.e. on the root locus plot. 6//(s) has
two 2eros and three poles. The locations of poles and 2eros depend on the numerical
values of the parameters. 1owever& to e.hibit a limit cycle& the root locus should cross the
imaginary a.is. The only possible configuration for a two-2eroEthree-pole system is when
2eros are located close to two symmetric poles but farther away from the real a.is. +igure
>-/4 shows the root locus of such a system& as an e.ample. The system has open loop
poles at J4./ and J4.3 j4.B and 2eros at J4.3 j3.

-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Real Axis
I
m
a
g

A
x
i
s
root locus, z = -0.2 +/- 2j, p = -0.2 +/- 0.5j, -0.1
+igure >-/4 ,oot locus of third-order system with crossover points.
Heros of 6
//
(s) are easily obtained&
.
/ /
?
/ /
3
/
/ 3
3
3
=
3 / 3
=
3 /
3 * *
R
R 3 *
R
R 3
,
|
|
.
|

\
|

|
|
.
|

\
|
+ = (>.33)
Heros will have imaginary components if the term inside the suare root is negative. The
readers will recogni2e /E*
3
3
/
as the suare of a natural freuency of oscillation
/B
. Thus
6
//
(s) has a potential to show a limit cycle if the freuency due to the reactive
components (*
3
and 3
/
) is greater than the one due to resistive components (R
3
and R
=
).
Note the components involved are those of the system pushed by 6
//
(s) (past the supply
duct).
+inding poles of 6
//
(s) is a much more difficult tas% unless numerical values are given&
because the third order denominator is not easily decomposed into simple factors. Instead
of trying to find poles& we use the euation& ImF6
//
(je)I R 4& to see under which condition
the solutions e.ist (poles are on the imaginary a.is).
:e rewrite 6
//
(s) as&
.
) (
) (
3 =
3
//
3 =s "s s
( 1s s
s 6
+ + +
+ +
=
o
(>.3=)
/B
In a simple *)R)3 circuit& the comple. poles& if they e.ist& are given by s 7 )a je & where a R RE3*& e
3
R
en
3
J a
3
and en
3
R /E*3. en is called natural frequency of oscillation and e is damped frequency of
oscillation.
/C?
!fter wor%ing out all the algebraic details&
ImF6
//
(je)I R -e(e
?
J (= 9 ( J "1)e
3
9 (31 J =()) R 4. (>.3?)
e R 4 is a trivial solution (one system pole is always on the real a.is). To show a limit
cycle& there should be four crossover points on the root locus. Thus the term inside the
parenthesis should have four real roots& i.e.&

3
) ( ? ) ( ) (
3
3
=( 31 "1 ( = "1 ( = + +
= e
(>.3B)
should have two real positive solutions. It is possible when
. 4 ) ( ? ) ( and 4
3
> + > + =( 31 "1 ( = "1 ( = (>.30)
The first condition is worth e.amining.

.
) (
/
/
3 /
3 / =
3 / /
3 /
3
3
=
/
3 /
/ 3
|
|
.
|

\
|
+
+
+
+
|
|
.
|

\
|

|
|
.
|

\
|
+
+
= +
* *
R R R
R * 3
R R
*
R
*
* *
3 *
"1 ( =
(>.3A)
The first term of the euation is the freuency suared due to the reactive& energy storage&
components of the system. The second is the freuency suared due to the resistive&
energy dissipative& components. (both in an aggregate sense.) Thus the system can e.hibit
a limit cycle if the aggregate reactive freuency is greater than the aggregate resistive
freuency.
The second condition is difficult to analy2e due to the suare term involved. 1owever& we
surely %now that the condition is satisfied if 31 J =( is nonpositive.

. /
/ ) ( /
3
=
3 /
3 /
3 / / 3
= 3 / =
3 /
= 3 /
3 / /
|
|
.
|

\
|
+
+

|
|
.
|

\
| + +
+
+ +
=
R
R
* 3
* *
* * 3 *
R R R R
R 3
R R R
* * 3
=( 31
(>.3C)
The condition is satisfied when the second term& reactive& is greater than the first term&
resistive.
+inally& we loo% at the harmonic balance euation with # R 47

.
4 ) 4 ( ) 4 (
= 3 /
/
= 3 /
= 3
4
43
= 3 /
/
4/
= 3 /
= 3
4
43 /3 4/ // 4
r s
m
R R R
R
m
R R R
R R
a
c
R R R
R
c
R R R
R R
a
c 6 c 6 a
+ +

+ +
+
~
+ +

+ +
+
=
=
(>.3<)
:e need two transfer functions in this euation& because the supply duct pressure s is
contributed by two inputs. a
4
is the #- value of s& and c
4/
and c
4/
are the ,eroth +ourier
series coefficient of es& the first component of u vector& and >er& the second component of
u vector& respectively. Thus we obtain an euation about the resistive components of
system& if we %now a
4
& ms& and mr& which are easy to measure. (a
4
is obtained by averaging
the supply duct pressure measurements obtained for a period of time. ms and mr are
obtained by driving the )+E,+ 8+#s in open loop mode with the commands of fi.ed
offset values (/<.3Bm! and /C.= m!) and measuring the pressures right after the fans.)
:e have developed a very simple model of 8!8 !1; system& and tried to e.plain the
sustained oscillation behavior as a limit cycle due to the saturation of actuators. :e have
not proven that there e.ists a limit cycle& because the numerical values of the system are
un%nown at this point. 1owever& we have showed a variety of conditions that lead to the
development of limit cycle. !lso a few euations are derived& which in turn can be used
to find system parameters. +or e.ample& we %now that e R 3tE=4 radEsec is the solution to
ImF6
//
(je)I& because it is the limit cycle freuency. :e also %now the relationship
between the resistive components& the signal #- value and the input offset. +urther
efforts are reuired to prove the model& such as finding numerical parameter values
andEor improving models to include the effect of dampers& etc. 1owever& once proved& the
model can generally be applied to any type of 8!8 !1; system and in turn can be used
to find a few essential system parameters.
1aving gone through all the gory details of modeling and analysis& a lingering uestion is
how we can eliminate the oscillation& if necessary
/0
. @ne obvious solution is to reduce the
gain so that the closed system poles never cross the imaginary a.is. 1owever& this can
lead to a poor system performance and essentially is no different from running the system
in an open loop mode. !nother option is to design a more sophisticated controller& so that
the closed loop poles will never have a chance to cross the imaginary a.is. This is
possible when the model is proved and the necessary parameters are obtained.
/0
:e also haven6t as%ed the uestion that if the system is described in such a way& which is the best way to
control the system. This is the issue of designing the optimal control logic for an !1; system.
/C0
Appendi@ C
Estimation o' (irst 7rder Polynomial *ean o' a -andom Process
This short appendi. develops a simple method to estimate the time varying mean of a
random process& as a first order polynomial of time. The random process is first modeled
as white Daussian and the ma.imum-li%elihood (ML) estimator is derived. It is shown
that the least-suares error criterion leads to the same formula and that the ML estimator
can also be used with a non-Daussian additive noise in a least-suared error sense. The
first order polynomial mean estimator is mostly used to estimate the mean of a time series
of a higher harmonic power& in order to estimate the mean of the fundamental power
consumed by a variable speed drive through a correlation between two powers.
)uppose that @ is a length-N observation of a random process& e.g. an apparent power
(!#)&
| |
T
N & & & I / F I / F I 4 F = . @ (-./)
:e assume that &FnI is the sum of a first order polynomial mean and a white noise7
/ 4 I F I F I F s s + = N n n w n m n &
&
& (-.3)
where | o + = n n m
&
I F and
3
I) F var( o = n w . :ith this signal model& the observation
vector @ follows a Daussian distribution with the probability density function ("#+) of7

( ) & I F
3
/
e.p
) 3 (
/
) ( ) (
3
/
e.p
) ( det ) 3 (
/
) & Y (
/
4
3
3
/
3
3
/
3
|
.
|

\
|
=
|
.
|

\
|
=

N
n
N
T
n n &
p
N
N
| o
o
o t
t
| o
@ @ @
@
m @ G m @
G
@
(-.=)
where o and | are the parameters of the distribution. The ma.imum li%elihood (ML)
estimations of these parameters are found by ma.imi2ing the logarithm of the "#+ with
respect to o and |. >ecause the logarithm of the "#+ is a uadratic function of o and |&
the ma.imum occurs where the first derivatives are 2ero.
4 ) I F (
/ ) & Y ( ln
/
4
3
= =
c
c

=
N
n
n n & n
p
| o
o o
| o @
(-.?)
4 ) I F (
/ ) & Y ( ln
/
4
3
= =
c
c

=
N
n
n n &
p
| o
o |
| o @
.
>y combining these two euations& we obtain the following single matri. euation7
$ Aa = & (-.B)
where

(
(
(
(

=
(

=
(
(
(
(

=
(
(
(
(

=
/
4
/
4
/
4
/
4
/
4
/
4
3
I F
I F
& &
3
) / (
3
) / (
0
) / 3 )( / (
/
N
n
N
n
N
n
N
n
N
n
N
n
n &
n & n
N
N N
N N N N N
n
n n
$ a A
|
o
. (-.0)
The matri. A is invertible whenever N > 3 and does not depend on &FnI. Thus the ML
estimation of a can be computed easily by inverting A& which is a function of N only.
The same formula can be derived by using the least-suares (L)) error criterion& i.e.
6a 1 n @ = + ~ | o & (-.A)

where
I F 1 n 6 =
and
T
N I / / 4 F = . n . The L) estimation of a is given by7
@ 6 6 6 a
T T /
) ( U

= . (-.C)
,eaders can easily verify that 6
T
6 R A and 6
T
@ R $. This is not a coincidence because
the L) estimator is the best linear unbiased estimator (>L;() when the additive noise is
2ero-mean white Daussian due to the Dauss-Mar%ov theorem F$ay& /<<=I. The >L;( is
a minimum variance unbiased (M8;) estimator constrained to be linear. @n the other
hand& an ML estimator is appro.imately the M8; estimator when its record si2e is
sufficiently large.
Thus when the additive noise is white Daussian& we have an ML first order polynomial
mean estimator& which is appro.imately the M8; estimator. :hen the noise is not white&
the same mean estimator is still valid in the L) error sense. The latter point of view is
especially helpful because power signals typically carry colored noises& such as those
generated by variable speed drives.
:hen the first order mean estimation is performed on each succeeding window& we may
impose an additional constraint that the mean estimation is continuous across window
boundaries. )uppose that m
@& j is the length-N mean vector of the jth window and that m
@& j-
/ is the length-N mean vector of the preceding (j-/)th window. The constraint can be
stated as m&& j-/FNI R m&& jF4I. In terms of parameters7

j j j
N | | o = +
/ /
. (-.<)
/CC
Thus when the continuity constraint is applied& the parameter | of a current window is
determined solely by the parameters of its preceding window. Then the remaining
parameter oj is obtained by solving the first derivative condition&

0
) / 3 )( / (
3
) / (
I F I F
/
4
/
4
3
/
4
/
4

=
N N N
N N
n & n
n
n n & n
j
N
n
N
n
N
n
j
N
n
j
| |
o . (-./4)
Appendi@ D
E),S So't?are Architecture
This appendi. gives a brief overview of the (LI) software architecture at bloc% diagram
level. The (LI) software performs signal analysis and information e.traction on the non-
intrusive power signals generated by a power coder. It also displays acuired information
to users on hyper te.t mar%up language (1TML) pages that are accessible via the internet.
The (LI) is currently written with -99 and e.ecuted on a Linu. platform.

The power coder performs sampling of raw voltage and current waveforms and
computation of current harmonic powers. It currently is configured to output eight power
terms (& !&
=
& !
=
&
B
& !
B
&
A
& !
A
) at /34 12. The power terms become input to the (LI)
via a pipe
/A
. The pipe also allows opening a prerecorded data file and supplying the
readings as input to the (LI)& which is useful to test (LI) programs offline.
! real-time data handling software typically consists of three main componentsJ
initiali2ation& main loop and termination. The main loop is called repeatedly either at a
fi.ed freuency or with the occurrence of a triggering event. The (LI) also has these
three components and its main loop is called with a fi.ed interval. +igure #-/ shows the
bloc% diagram of the (LI).


Initiali2ation
Main Loop
Termination
+igure #-/ Three main components of (LI).
The initiali2ation bloc% reads load database and process parameters (classification
methods& thresholds& options& etc.)& assigns memory and creates variables. It also
initiali2es 1TML pages for webcasting. The 1TML pages are generated as te.t files and
interpreted by web browsers. +igure #-3 shows the bloc% diagram of the (LI)
initiali2ation.
-urrently& the constant load disaggregator and the 8)# trac%er are implemented as
separate entities& although they will be merged as a single (LI) program in future. The
main loop of the load disaggregator performs data acuisition& event detection and
classification and 1TML page update. The load disaggregator has three 1TML pagesJ
/A
"ipe is a file-handling interface in Linu.. It simply connects one process6s output to another process6s
input. [Y6 is the shell command used to create a pipe.
/<4
main& event and history. +igure #-= shows the bloc% diagram of the main loop of the
constant load disaggregator.

,ead load database
,ead process parameters
!llocate memory and create variables
Initiali2e 1TML pages
+igure #-3 Initiali2ation bloc% of (LI).

,ead a stage of data
Terminate if (@+ reached
-ompute downsampled and !
#etect change-of-mean events
!ssociate events
-lassify events and update event 1TML page
;pdate current and ! plot (main 1TML page)
;pdate history plots (history 1TML page)
+igure #-= Main loop of constant load diaggregator.
The main loop first reads a stage of data. ! stage typically consists of /43? or /344 rows
of power terms. Thus& with the /34 12 power coding rate& the main bloc% is e.ecuted at
every C.B= or /4 sec. If an (@+ (end of file) is reached during readout& the main bloc% is
terminated immediately and the program control is forwarded to the termination bloc%.
!fter a stage of data is successfully retrieved& and ! streams are downsampled for the
change-of-mean detection.
The change-of-mean detector e.amines the incoming stream and converts it to a series
of constant-mean segments. The event associator analy2es the series and locates load
events. The load events are forwarded to the event handler that classifies the events and
posts the results on the event 1TML page. !n event figure is created by gnuplot and its
lin% and event tag are inserted to the 1TML page. !lso& the plots of real and reactive
powers are generated for the main 1TML page and the plots of the time history of various
powers are created for the history 1TML page. +inally& the main bloc% goes bac% to read
a stage of data and the process is repeated.
The 8)# trac%er has three 1TML pagesJ main& ")# estimation and history. Its main
bloc% is generally similar to that of the constant load disaggregator& but has different
signal analysis routines. +igure #-? shows the bloc% diagram of the 8)# trac%er main
loop.

,ead a stage of data
Terminate if (@+ reached
-ompute "
#
(stimate 8)# mean powers
;pdate & ! and 8)# power plots (main 1TML page)
(stimate ")# and perform white filtering
;pdate ")# plots (")# estimation 1TML page)
;pdate history plots (history 1TML page)
+igure #-? Main loop of 8)# trac%er.
The 8)# trac%er main loop first reads a stage of data and chec%s if an (@+ is reached.
Then& it computes necessary higher harmonic apparent powers from the stage of data.
+irst order mean estimation is performed on a selected apparent power& and the mean
fundamental power consumed by 8)#s is estimated using a %nown correlation. The
estimation results are displayed as figures on the main 1TML page. The !, ")# of
current stage signal is estimated and its parameters are used to construct a whitening filter
for leftover real and reactive powers after their 8)# means are compensated. The ")#
estimation and white filtering results are again displayed as figures on the ")# estimation
1TML page. +inally& history plots are updated for the history 1TML page.
The termination bloc%s of both the constant load disaggregator and the 8)# trac%er
simply release dynamically allocated memory to prevent memory lea%.
/<3
i$liography
,. !. !bdoo& 'Industry ,estructuring for #eregulation and -ompetition&* I((( ower
(ngineering Review& pp. /0 J /C& Muly /<<<.
1. !%ai%e& '! New Loo% at the )tatistical Model Identification&* I((( Transactions on
"utomatic 3ontrol& 8ol. /<& No. 0& pp. A/0 A3=& /<A?.
M. !rrillaga& M. 1. M. >ollen and N. ,. :atson& '"ower Xuality following #eregulation&*
roceedings of the I(((& 8ol. CC& No. 3& pp. 3?0 J 30/& +ebruary 3444.
>. ,. >ar%ovich and #. 8. 1aw%& '-harting a New -ourse in -alifornia&* I(((
%pectrum& pp. 30 J =/& Muly /<<0.
M. >asseville and I. 8. Ni%iforov& 1etection of "brupt 3hanges& "rentice 1all&
(nglewood -liffs& NM& /<<=.
1. M. >eides and D. T. 1eydt& '#ynamic )tate (stimation of "ower )ystem 1armonics
using $alman +ilter Methodology&* I((( Transactions on ower 1elivery& 8ol. 0& No. ?&
pp. /00= J /0A4& @ctober /<</.
#. ". >ertse%as& 1ynamic rogramming& "rentice 1all& (nglewood -liffs& NM& /<CA.
M. !. -. >ingham& The Theory and ractice of .odem 1esign& :iley& New Kor%& /<CC.
>. $. >ose& '8ariable +reuency #rives- Technology and !pplications*& -onference
"roceedings of the I((( International %ymposium on Industrial (lectronics& I)I(6<=&
>udapest& pp./ J /C& /<<=.
D. >ucci& (. +iorucci and -. Landi& '#igital Measurement )tation for "ower Xuality
!nalysis in #istributed (nvironments&* I((( Transactions on Instrumentation and
.easurement& 8ol. B3& No. /& pp. AB J C?& +ebruary 344=.
8. ). >udhra5a& '-alifornia6s (lectricity -risis&* I((( ower (ngineering Review& pp. 0
J /?& !ugust 3443.
). ,. >ull& ',enewable (nergy Today and Tomorrow&* roceedings of the I(((& 8ol. C<&
No. C& pp. /3/0 J /330& !ugust 344/.
L. -armichael& 'Nonintrusive !ppliance Load Monitoring )ystem&* (RI 0ournal&
(lectric "ower ,esearch Institute& pp. ?B J ?A& )eptember /<<4.
1. -hen& :. -. $ong and -. -. $o& '-omparison of "itch "rediction and !daptation
!lgorithms in +orward and >ac%ward !daptive -(L" )ystems&* I(( roceedings on
3ommunications2 %peech and "udio Vision& 8ol. /?4& No. ?& pp. 3?4 3?B& /<<=.
K.-+. -hen and T.-#. -hiueh& '>aseband Transceiver #esign of a /3C-$bps "ower-Line
Modem for 1ousehold !pplications&* I((( Transactions on ower 1elivery& 8ol. /A&
No. 3& pp. ==C J =??& !pril 3443.
K.-). -hoi and #.-1. Koun& '+ast 1armonic (stimation Method for 1armonic )peech
-oders&* (lectronics *etters& 8ol. =C& No. A& pp. =?0 =?A& 3443.
!. I. -ole and !. !lbic%i& '#ata (.traction for (ffective Non-Intrusive Identification of
,esidential "ower Loads&* roceedings of the I((( Instrumentation and .easurement
Technology 3onference& 8ol. 3& pp. C/3 J C/B& May /<<C.
!. I. -ole and !. !lbic%i& '!lgorithm for Non-Intrusive Identification of ,esidential
!ppliances&* roceedings of the I((( International %ymposium on 3ircuits and %ystems
(I)-!))& 8ol. =& pp. ==C J =?/& /<<C.
>. -oo%& 'Introduction to +uel -ells and 1ydrogen Technology&* I(( (ngineering
%cience and (ducation 0ournal& pp. 34B J 3/0& #ecember 3443.
T. 1. -ormen& -. (. Leiserson& ,. L. ,ivest and -. )tein& Introduction to "lgorithms&
)econd (dition& MIT "ress& -ambridge& Massachusetts& 344/.
M. L. -row& '"ower )ystem #eregulation&* I((( otentials& pp. C J <& Manuary 3443.
". $. #ash& ,. $. Mena& D. "anda and !. ,outray& '!n (.tended -omple. $alman +ilter
for +reuency Measurement of #istorted )ignals&* I((( Transactions on
Instrumentation and .easurement& 8ol. ?<& No. ?& pp. A?0 J AB=& !ugust 3444.
M. ,. #eller& M. 1. L. 1ansen and M. D. "roa%is& 1iscrete)Time rocessing of %peech
%ignals& I((( "ress& "iscataway& NM& 3444.
(. L. #raper& '!ssessment of #eregulation and -ompetition&* I((( ower (ngineering
Review& pp. /A J /C& Muly /<<C.
). #ren%er and !. $ader& 'Nonintrusive Monitoring of (lectric Loads&* I((( 3omputer
"pplications in ower& 8ol. /3& No. ?& pp. ?A J B/& @ctober /<<<.
M. #riesen& T. 8an -raenenbroec% and #. 8an #ommelen& 'The ,egistration of
1armonic "ower by !nalog and #igital "ower Meters&* I((( Transactions on
Instrumentation and .easurement& 8ol. ?A& No. /& pp. /<B J /<C& +ebruary /<<C.
M. (l-1abrou% and M. $. #arwish& '#esign and Implementation of a Modified +ourier
!nalysis 1armonic -urrent -omputation Techniue for "ower !ctive +ilters using
#)"s&* I(( roceedings on (lectric ower "pplications& 8ol. /?C& No. /& pp. 3/ J 3C&
Manuary 344/.
!. (madi& M. (hsani and M. M. Miller& '!dvanced )ilicon ,ich !utomotive (lectrical
"ower )ystems&* roceedings of the (ighteenth 1igital "vionics %ystems 3onference&
8ol. 3& pp. C.>./-/ J C.>./-C& /<<<.
!. (. (manuel& '"owers in Nonsinusoidal )ituations7 ! ,eview of #efinitions and
"hysical Meaning&* I((( Transactions on ower 1elivery& 8ol. B& No. =& pp. /=AA J
/=C<& Muly /<<4.
!. (. (manuel& M. !. @rr& #. -ygans%i and (. M. Dulachens%i& '! )urvey of 1armonic
8oltages and -urrents at #istribution )ubstations&* I((( Transactions on ower
1elivery& 8ol. 0& No. ?& pp. /CC= J /C<4& @ctober /<</.
!. (. (manuel& M. !. @rr& #. -ygans%i and (. M. Dulachens%i& '! )urvey of 1armonic
8oltages and -urrents at the -ustomer6s >us&* I((( Transactions on ower 1elivery&
8ol. C& No. /& pp. ?// J ?3/& Manuary /<<=.
!. (. (manuel& '1armonics in the (arly Kears of (lectrical (ngineering7 ! >rief
,eview of (vents& "eople and #ocuments&* roceedings of the Ninth International
3onference on $armonics and !uality of ower& 8ol. /& pp. / J A& 3444.
/<?
L. (rup& +. M. Dardner and ,. !. 1arris& 'Interpolation in #igital Modems- "art II7
Implementation and "erformance&* I((( Transactions on 3ommunications& 8ol. ?/& No.
0& pp. <<C J /44C& Mune /<<=.
M. +arooue and 1. -. Maru& '+uel -ells- The -lean and (fficient "ower Denerators&*
roceedings of the I(((& 8ol. C<& No. /3& pp. /C/< J /C3<& #ecember 344/.
D. +. +ran%lin& M. #. "owell and !. (mami-Naeini& ?eedbac# 3ontrol of 1ynamic
%ystems& Third (dition& !ddison :esley& ,eading& M!& /<<?.
$. +u%unaga& Introduction to %tatistical attern Recognition& !cademic "ress& NK& /<A3.
+. M. Dardner& 'Interpolation in #igital Modems- "art I7 +undamentals&* I(((
Transactions on 3ommunications& 8ol. ?/& No. =& pp. B4/ J B4A& March /<<=.
M. (. Dibson& Nonlinear 3ontrol %ystem& McDraw-1ill& NK& /<0=.
1. ). Dill& 8. +ilipovic-Dled5a& 1. ". Mu%her5ee& T. ). )idhu and L. -ui& ')oftware
Xuality Tools for the #eregulated Mar%et&* I((( 3omputer "pplications in ower& pp.
== J =C& @ctober 344/.
(. 1. Draf and M. ). Lee& '!ll-!ir )ystems&* in $V"3 %ystems and 3omponents
$andboo#& N. ,. Drimm and ,. -. ,osaler& (ds& McDraw-1ill& NK& /<<C.
#. :. Driffin and M. ). Lim& 'Multiband (.citation 8ocoder&* I((( Transactions on
"coustics2 %peech and %ignal rocessing& 8ol. =0& No. C& pp. /33= /3=B& /<CC.
(. :. Dunther& 'Interharmonics- ,ecommended ;pdates to I((( B/<&* I((( ower
(ngineering %ociety %ummer .eeting& 8ol. 3& pp. <B4 J <B?& 3443.
". D. 1arris& 'Impacts of #eregulation on the (lectric "ower Industry&* I((( ower
(ngineering Review& pp. ? J 0& @ctober 3444.
D. :. 1art& 'Nonintrusive !ppliance Load Monitoring&* roceedings of the I(((& 8ol.
C4& No. /3& pp. /CA4 J /C</& #ecember /<<3.
I((( Industrial !pplications )ocietyE"ower (ngineering )ociety& 'I((( ,ecommended
"ractices and ,euirements for 1armonic -ontrol in (lectrical "ower )ystems&* I(((
%tandard @AB)ABBC& NK& /<<=.
I((( :or%ing Droup on Nonsinusoidal )ituations& '! )urvey of North !merican
(lectric ;tility -oncerns ,egarding Nonsinusoidal :aveforms&* I((( Transactions on
ower 1elivery& 8ol. //& No. /& pp. A= J AC& Manuary /<<0.
I((( :or%ing Droup on Nonsinusoidal )ituations7 (ffects on Meter "erformance and
#efinitions of "ower& '"ractical #efinitions for "owers in )ystems with Nonsinusoidal
:aveforms and ;nbalanced Loads7 ! #iscussion&* I((( Transactions on ower
1elivery& 8ol. //& No. /& pp. A< J /4/& Manuary /<<0.
M. Ilic& ". )%ant2e and ". 8isudhiphan& '(lectricity Troubles in -alifornia7 :ho6s
Ne.t\* I((( %pectrum& pp. // J/=& +ebruary 344/.
(. Macobsen and ,. Lyons& 'The )liding #+T&* I((( %ignal rocessing .aga,ine& 8ol.
34& No. 3& pp. A? J C4& March 344=.
:. -. $arl& ). >. Leeb& L& !. Mones& M. L. $irtley and D. -. 8erghese& '!pplications of
,an%->ased +ilters in "ower (lectronics*& I((( Transactions on ower (lectronics& 8ol.
A& No. =& pp. ?=A J ??=& Muly /<<3.
#. -. $arnopp& #. L. Margolis and ,. -. ,osenberg& %ystem 1ynamicsD " 9nified
"pproach& )econd (dition& :iley-Inter)cience& NK& /<<4.
M. D. $assa%ian& 1.--. :olf& M. M. Miller and -. M. 1urton& '!utomotive (lectrical
)ystems circa 344B&* I((( %pectrum& pp. 33 J 3A& !ugust /<<0.
M. D. $assa%ian& '!utomotive (lectrical )ystems- The "ower (lectronics Mar%et of the
+uture&* ?ifteenth "nnual I((( "pplied ower (lectronics 3onference and (&position&
8ol. /& pp. = J <& 3444.
). M. $ay and ). L. Marple& ')pectrum !nalysis- ! Modern "erspective&* roceedings
of the I(((& 8ol. 0<& No. //& pp. /=C4/?/<& /<C/.
). M. $ay& ')pectral (stimation&* in "dvanced Topics in %ignal rocessing& M. ). Lim
and !. 8. @ppenheim& (ds& "rentice 1all& (nglewood -liffs& NM& /<CC.
). M. $ay& ?undamentals of %tatistical %ignal rocessing2 Volume ID (stimation Theory&
"rentice 1all& (nglewood -liffs& NM& /<<=.
). M. $ay& ?undamentals of %tatistical %ignal rocessing2 Volume IID 1etection Theory&
"rentice 1all& ;pper )addle ,iver& NM& /<<C.
D. $endall& '"ower @utages during Mar%et #eregulation&* I((( 3ontrol %ystems
.aga,ine& pp. == J =<& #ecember 344/.
>. :. $ennedy& ower !uality rimer& McDraw-1ill& New Kor%& 3444.
I. !. $han& '!utomotive (lectrical )ystems7 !rchitecture and -omponents&*
roceedings of the (ighteenth 1igital "vionics %ystems 3onference& 8ol. 3& pp. C.-.B-/ J
C.-.B-/4& /<<<.
". $oponen& ,. )eesvuori and ,. >ostman& '!dding "ower Xuality Monitoring to a
)mart %:h Meter&* I(( ower (ngineering 0ournal& pp. /B< J /0=& !ugust /<<0.
T. I. Laa%so& 8. 8alima%i& M. $ar5alainen and ;. $. Laine& ')plitting the ;nit #elay&*
I((( )ignal "rocessing Maga2ine& 8ol. /=& No. /& pp. =4 J 04& Manuary /<<0.
-. :. Lander& ower (lectronics& )econd (dition& McDraw-1ill& London& /<CA.
-. Laughman& $. #. Lee& ,. -o.& ). )haw& ). >. Leeb& L. $. Norford and ". !rmstrong&
'!dvanced Nonintrusive Monitoring of (lectric Loads&* accepted by I((( ower and
(nergy .aga,ine (to appear on MarchE!pril 344=).
:. M. Lee and -. 1. Lin& ';tility #eregulation and Its impact on Industrial "ower
)ystems&* I((( Industry "pplications .aga,ine& pp. ?4 J ?0& May /<<C.
). >. Leeb& " 3onjoint attern Recognition "pproach to NonIntrusive *oad .onitoring&
"h. #. Thesis& MIT& #epartment of (lectrical (ngineering and -omputer )cience& /<<=.
). >. Leeb& ). ,. )haw and M. L. $irtley& 'Transient (vent #etection in )pectral (nvelope
(stimates for Nonintrusive Load Monitoring&* I((( Transactions on ower 1elivery&
8ol. /4& No. =& pp. /344 J /3/4& Muly /<<B.
/<0
). >. Leeb& >. -. Lesieutre and ). ,. )haw& '#etermination of Load -omposition using
)pectral (nvelope (stimates&* roceedings of the CE
th
"nnual North "merican ower
%ymposium& I((( "ower (ngineering )ociety& >o2eman& Montana& @ctober 3 J =& /<<B.
-. Li& :. Nu and T. Tay5asanant& 'Interharmonics7 >asic -oncepts and Techniues for
Their #etection and Measurement&* to appear on 0ournal of (lectric ower %ystems
Research& ("aper No.7 /CB<)& 344=.
M. ). Lim and !. 8. @ppenheim& '!ll-"ole Modeling of #egraded )peech&* I(((
Transactions on "coustics2 %peech and %ignal rocessing& 8ol. 30& No. =& pp. /<A 3/4&
/<AC.
M. ). Lim& !. 8. @ppenheim and L. #. >raida& '(valuation of an !daptive -omb +iltering
Method for (nhancing )peech #egraded by :hite Noise !ddition&* I((( Transactions
on "coustics2 %peech and %ignal rocessing& 8ol. 30& No. ?& pp. =B?=BC& /<AC.
K.-M. Lin& 1. !. Latchman& M. Lee and ). $atar& '! "ower Line -ommunication
Networ% Infrastructure for the )mart 1ome&* I((( 4ireless 3ommunications& pp. /4? J
///& #ecember 3443.
#. Luo& 1etection and 1iagnosis of ?aults and (nergy .onitoring of $V"3 %ystems
with *east)Intrusive ower "nalysis& "h. #. Thesis& MIT& #epartment of !rchitecture&
344/.
#. Luo& L. $. Norford& ). ,. )haw& and ). >. Leeb& 'Monitoring 18!- (uipment from
a -entrali2ed Location- Methods and +ield Test ,esults*& "%$R"( Transactions 8ol.
/4C& No. /& 3443.
H. Luo& H. Nu& K. Hheng and N. Lu& '#+T and #)"-based (lectric (nergy Measurement
!lgorithm of 1armonic )ource Load&* roceedings on the International 3onference on
ower %ystem Technology Fower3onG& 8ol. ?& pp. 3?CA J 3?<4& 3443.
). L. Marple& '! New !utoregressive )pectrum !nalysis !lgorithm&* I((( Transactions
on "coustics2 %peech and %ignal rocessing& 8ol. 3C& No. ?& pp. ??/ ?B?& /<C4.
). L. Marple& 1igital %pectral "nalysis& "rentice 1all& (nglewood -liffs& NM& /<CA.
". Martin& '-omparison of "itch #etection by -epstrum and )pectral -omb !nalysis&*
roceedings of the International 3onference on "coustics2 %peech and %ignal rocessing
(I-!))")& "aris& +rance& pp. /C4 /C=& May /<C3.
,. M. Mc!ulay and T. +. Xuatieri& '"itch (stimation and 8oicing #etection based on a
)inusoidal )peech Model&* roceedings of the International 3onference on "coustics2
%peech and %ignal rocessing (I-!))")& 8ol. /& pp. 3?< 3B3& !pril /<<4.
,. 1. Mc-lanahan& '(lectric #eregulation&* I((( Industry "pplications .aga,ine& pp.
// J /C& !pril 3443.
M. 1. Mc-lellan& '"arametric )ignal Modeling&* in "dvanced Topics in %ignal
rocessing& M. ). Lim and !. 8. @ppenheim& (ds& "rentice 1all& (nglewood -liffs& NM&
/<CC.
:. McNamara& The 3alifornia (nergy 3risisD *essons for a 1eregulating Industry& "enn
:ell& Tulsa& @%lahoma& 3443.
M. M. Miller& 'Multiple 8oltage (lectrical "ower #istribution )ystem for !utomotive
!pplications&* roceedings of the HA
st
Intersociety (nergy 3onversion (ngineering
3onference (I(-(-)& 8ol. =& pp. /<=4 J /<=A& !ugust /<<0.
M. M. Moreno-(guila2& M. "eracaula and !. (suivel& 'Neural Networ% based !pproach
for the -omputation of 1armonic "ower in a ,eal-time Microprocessor-based 8ector
-ontrol for an Induction Motor #rive&* roceedings of the I((( International
%ymposium on Industrial (lectronics& 8ol. /& pp. 3AA J 3C/& 3444.
#. ". Morgan& (. >. Deorge& L. T. Lee and ). M. $ay& '-ochannel )pea%er )eparation by
1armonic (nhancement and )uppression&* I((( Transactions on %peech and "udio
rocessing& 8ol. B& No. B& pp. ?4A?3?& /<<A.
,. Nai%& N. Mohan& M. ,ogers and !. >ulaw%a& '! Novel Drid Interface& @ptimi2ed for
;tility-scale !pplications of "hotovoltaic& :ind-electric and +uel-cell )ystems&* I(((
Transactions on ower 1elivery& 8ol. /4& No. ?& pp. /<34 J /<30& @ctober /<<B.
M. !. Naylor and ). +. >oll& 'Techniues for )uppression of an Interfering Tal%er in -o-
channel )peech&* roceedings of the International 3onference on "coustics2 %peech and
%ignal rocessing (I-!))")& #allas& TN& pp. 34B 34C& !pril /<CA.
M. Naylor and M. "orter& '!n (ffective )peech )eparation )ystem which ,euires No "
riori Information&* roceedings of the International 3onference on "coustics2 %peech
and %ignal rocessing (I-!))")& Toronto& @ntario& -anada& pp. <=A <?4& May /<</.
M. Neubert& '"owering up&* I(( Review& pp. 3/ J 3B& )eptember 3444.
!. Ni5enhuis and 1. ). :ilf& 3ombinatorial "lgorithms& !cademic "ress& NK& /<AB.
Nils M. Nilsson& "rtificial Intelligence& Morgan $aufmann& )an +rancisco& -!& /<<C.
L. $. Norford and ). >. Leeb& 'Non-intrusive (lectrical Load Monitoring in -ommercial
>uildings based on )teady-state and Transient Load-detection !lgorithms&* (nergy and
=uildings& 8ol. 3?& No. /& pp. B/ J 0?& /<<0.
M. !. @liver& ,. Lawrence and >. >. >aner5ee& '"ower Xuality7 1ow to )pecify "ower-
Xuality-Tolerant "rocess (uipment&* I((( Industry "pplications .aga,ine& pp. 3/ J
=4& )eptember 3443.
!. 8. @ppenheim& !. ). :ills%y and ). 1. Nawab& %ignals and %ystems& second edition&
"rentice-1all& (nglewood -liffs& NM& /<<C.
!. 8. @ppenheim& ,. :. )chafer and M. ,. >uc%& 1iscrete)time %ignal rocessing&
second edition& "rentice-1all& ;pper )addle ,iver& NM& /<<<.
). @sows%i& 'Neural Networ% for (stimation of 1armonic -omponents in a "ower
)ystem&* I(( roceedings)3& 8ol. /=<& No. 3& pp. /3< J /=B& March /<<3.
). @sows%i& ')8# Techniue for (stimation of 1armonic -omponents in a "ower
)ystem7 ! )tatistical !pproach&* I(( roceedings on 6eneration2 Transmission and
1istribution& 8ol. /?/& No. B& pp. ?A= J ?A<& )eptember /<<?.
,. "alomera-!rias and L. $. Norford& "pplications of Non)intrusive (lectrical *oad
.onitoring to the 1etection and 1iagnosis of ?aults "ssociated with the 'n)off
/<C
%witching of $V"3 (quipment& Technical ,eport to 1oneywell& Inc.& MIT #epartment of
!rchitecture& !ugust <& 344/.
!. "apoulis& ). ;. "illai& robability2 Random Variables and %tochastic rocesses&
+ourth (dition& McDraw-1ill& NK& 3443.
T. :. "arsons& ')eparation of )peech from Interfering )peech by Means of 1armonic
)election&* 0ournal of the "coustical %ociety of "merica& 8ol. 04& No. ?& pp. <// </C&
/<A0.
(. "ar2en& ')ome ,ecent !dvances in Time )eries Modeling&* I((( Transactions on
"utomatic 3ontrol& 8ol. /<& No. 0& pp. A3= A=4& /<A?.
N. "avlidou& !. M. 1an 8inc%& M. Ka2dani and >. 1onary& '"ower Line -ommunications7
)tate of the !rt and +uture Trends&* I((( 3ommunications .aga,ine& pp. =? J ?4& !pril
344=.
L. "hilipson and 1. L. :illis& 9nderstanding (lectric 9tilities and 1e)Regulation&
Marcel #e%%er& NK& /<<C.
8. +. "isaren%o& 'The ,etrieval of 1armonics from a -ovariance +unction&* 6eophysical
0ournal of Royal "stronomical %ociety& 8ol. ==& pp. =?A =00& /<A=.
!. "rice& ). >artley& ). Male and D. -ooley& '! Novel !pproach to ;tility )cale (nergy
)torage&* I(( ower (ngineering 0ournal& pp. /33 J /3<& Mune /<<<.
M. D. "roa%is& -. M. ,ader& +. Ling& -. L. Ni%ias& M. Moonen and I. $. "roudler&
"lgorithms for %tatistical %ignal rocessing& "rentice 1all& ;pper )addle ,iver& NM.
3443.
T. +. Xuatieri and ,. D. #anisewic2& '!n !pproach to -o-channel Tal%er Interference
)uppression using a )inusoidal Model for )peech&* I((( Transactions on "coustics2
%peech and %ignal rocessing& 8ol. =C& No. /& pp. B0 0<& /<<4.
). ,ech%a& (. Ngandui& M. Nu and ". )icard& '! -omparative )tudy of 1armonic
#etection !lgorithms for !ctive +ilters and 1ybrid !ctive +ilters&* The HH
rd
"nnual
I((( ower (lectronics %pecialists 3onference F%(3G& 8ol. /& pp. =BA J =0=& 3443.
(. M. ,eingold& M. Nievergelt and N. #eo& 3ombinatorial "lgorithms& "rentice-1all&
(nglewood -liffs& NM& /<AA.
M. ,issanen& 'Modeling by )hortest #ata #escription*& "utomatica& 8ol. /?& pp. ?0B-?A/&
/<AC.
M. ,issanen& '! ;niversal "rior for Integers and (stimation by Minimum #escription
Length&* "nnals of %tatistics& 8ol. //& No. 3& pp. ?/0 ?=/& /<C=.
,. ,oy and T. $ailath& '()",IT- (stimation of )ignal "arameters via ,otational
Invariance Techniues&* I((( Transactions on "coustics2 %peech and %ignal rocessing&
8ol. =A& No. A& pp. <C? <<B& /<C<.
1. )aari& ". $oponen& (. Tahvanainen and T. Lindholm& ',emote ,eading and #ata
Management )ystem for %:h Meters with "ower Xuality Monitoring&* I(( ower
(ngineering 0ournal& pp. /0? J /0C& !ugust /<<0.
). ,. )haw& -. >. !bler& ,. +. Lepard& #. Luo& ). >. Leeb and L. $. Norford&
'Instrumentation for 1igh "erformance Nonintrusive (lectrical Load Monitoring&* "%.(
0ournal of %olar (nergy (ngineering& 8ol. /34& No. =& !ugust /<<C.
). ,. )haw and ). >. Leeb& 'Identification of Induction Motor "arameters from Transient
)tator -urrent Measurements&* I((( Transactions on Industrial (lectronics& 8ol. ?0&
No. /& pp. /=< J /?<& +ebruary /<<<.
). ,. )haw& %ystem Identification Techniques and .odeling for Nonintrusive *oad
1iagnostics& "h. #. Thesis& MIT& #epartment of (lectrical (ngineering and -omputer
)cience& 3444.
). ,. )haw& #. Luo& L. $. Norford and ). >. Leeb& '#etection and #iagnosis of 18!-
+aults via (lectrical Load Monitoring&* International 0ournal of $V"3IR Research&
8ol. C& No. /& pp. /= J ?4& 3443.
". )toica and !. Nehorai& 'M;)I-& Ma.imum Li%elihood and -ramer-,ao >ound&*
I((( Transactions on "coustics2 %peech and %ignal rocessing& 8ol. =A& No. B& pp. A34
A?/& /<C<.
M. )tones and !. -ollinson& '"ower Xuality&* I(( ower (ngineering 0ournal& pp. B0 J
0?& !pril 344/.
,. M. )tubbs and X. )ummerfield& '(valuation of Two 8oice-separation !lgorithms using
Normal-hearing and 1earing-impaired Listeners&* 0ournal of the "coustical %ociety of
"merica& 8ol. C?& No. ?& pp. /3=0 /3?<& /<CC.
+. )ultanem& ';sing !ppliance )ignatures for Monitoring ,esidential Loads at Meter
"anel Level&* I((( Transactions on ower 1elivery& 8ol. 0& No. ?& pp. /=C4 J /=CB&
@ctober /<</.
M. L. )weeney& The 3alifornia (lectricity 3risis& 1oover Institution "ress& )tanford&
-alifornia& 3443.
M. D. Tru.ell& "utomatic ?eedbac# 3ontrol %ystem %ynthesis& McDraw-1ill& NK& /<BB.
T. M. ;lrych and ,. :. -layton& 'Time )eries Modeling and Ma.imum (ntropy&* hysics
of the (arth and lanetary Interiors& 8ol. /3& pp. /CC 344& /<A0.
T. L. 8incent and :. M. Drantham& Nonlinear and 'ptimal 3ontrol %ystems& Mohn :iley
] )ons& NK& /<<A.
,. ". :al%er& ':hy in the :orld :ould an (lectric ;tility be Interested in ,enewable
(nergy\* I((( ower (ngineering Review& pp. < J /3& May /<<C.
M. 8. :i5aya%ulasooriya& D. !. "utrus and ". #. Minns& '(lectric "ower Xuality
#isturbance -lassification using )elf-adapting !rtificial Neural Networ%s&* I((
roceedings on 6eneration2 Transmission and 1istribution& 8ol. /?<& No. /& pp. <C J
/4/& Manuary 3443.
M. #. :ise& M. ,. -aprio and T. :. "ar%s& 'Ma.imum Li%elihood "itch (stimation&*
I((( Transactions on "coustics2 %peech and %ignal rocessing& 8ol. 3?& No. B& pp.
?/C?3=& /<A0.
344
1. :old& " %tudy in the "nalysis of %tationary Time %eries& )econd (dition& !lmvist ]
:i%sell& )toc%holm& )weden& /<B?.
,. 1. :ol%& '+uel -ells for 1omes and 1ospitals&* I((( %pectrum& pp. ?B J B3& May
/<<<.
>. +. :ollenberg& 'The "rice of -hange&* I((( otentials& pp. /? J /0& Manuary /<<C.
:. Nu& 1. :. #ommel& M. >. 1ughes& D. :. $. -hang& L. Tang& 'Modeling of
!d5ustable )peed #rives for "ower )ystem 1armonic !nalysis*& I((( Transactions on
ower 1elivery& 8ol. /?& No. 3& pp. B<B J 04/& !pril /<<<.
:. Nu& ')tatus and +uture #irections of "ower )ystem 1armonic !nalysis&* to appear on
I((( "ower (ngineering )ociety Deneral Meeting& Toronto& -anada& Muly /= J /A& 344=.
,. Kacamini& '"ower )ystem 1armonics& "art /7 1armonic )ources&* I(( ower
(ngineering 0ournal& pp. /<= /<C& !ugust /<<?.
,. Kacamini& '"ower )ystem 1armonics& "art ?7 Interharmonics&* I(( ower
(ngineering 0ournal& pp. /CB J /<=& !ugust /<<0.
D. Haccour& (d.& 1eregulation of (lectric 9tilities& $luwer& Norwell& M!& /<<C.