Energy and Buildings 75 (2014) 239248

Contents lists available at ScienceDirect

Energy and Buildings

journal homepage: www.elsevier.com/locate/enbuild

Scheduling-based real time energy ow control strategy for building

energy management system
Shin Jae Kang a , Jungsung Park b , Ki-Yong Oh c , Jae Gu Noh b , Hyunggon Park d,
a

Wireless Convergence Platform Research Center, Korea Electronics Technology Institute (KETI), Seoul, Republic or Korea
KEPCO Research Institute, Korea Electric Power Corporation (KEPCO), Daejeon, Republic of Korea
Department of Mechanical Engineering, University of Michigan, Ann Arbor, USA
d
Department of Electronics Engineering, Ewha Womans University, Seoul, Republic of Korea
b
c

a r t i c l e

i n f o

Article history:
Received 1 April 2013
Received in revised form 14 January 2014
Accepted 7 February 2014
Keywords:
BEMS
DR
Real-time control
Optimal energy ow
Scheduler

a b s t r a c t
We propose a novel strategy for BEMS (Building Energy Management System), which efciently controls energy ows in a building so as to minimize the total cost of energy for a nite period. We also
consider Demand Response (DR) events during the period. The proposed strategy includes prediction,
long-term scheduling, and real-time control (RTC) of components within a building. During the period,
the process from prediction to RTC is iterated in every time unit when the system status is changed by
a dynamic environment. The scheduler determines the optimal energy ows based on the prediction,
and RTC utilizes the scheduling result so that the energy ow can be adaptively controlled in a dynamic
environment. Finally, the system status change information is fed back for the next iteration. Simulation
results indicate potential cost savings that are approximately 1020% compared to a typical BEMS with
a conventional RTC scheme.
2014 Elsevier B.V. All rights reserved.

1. Introduction
In the power system areas, there have been enormous efforts on
efcient energy management in order to resolve problems incurred
from the energy crises. The concept of the Smart Grid is one of the
solutions to those problems. In Smart Grid systems, all the information of every grid component should be accessible and all the
components composing the power grid should be controlled by
two-way communication [1]. Besides the infrastructures and facilities, there have been other efforts on electricity pricing policies
[2]. Real-time pricing is one of the well-known policies for dealing
with peak energy consumption. In this pricing scheme, the electricity pricing rate is available for customers one hour to one day
in advance. Thus, customers can manage the usage of their electric
devices in order to minimize the total electricity cost. For example, they can control the target temperatures in air conditioners
or electric heating appliances, or brightness of light bulbs based
on the electricity pricing rates [3]. In addition, DR incentive policy has been effectively used as a scheme for reducing peak load
usage. Customers who participate in the DR events may receive
benets and incentives from power utility companies when the

Tel.: +82 2 3277 3896; fax: +82 2 3277 3494.

E-mail address: hyunggon.park@ewha.ac.kr (H. Park).
http://dx.doi.org/10.1016/j.enbuild.2014.02.008
0378-7788/ 2014 Elsevier B.V. All rights reserved.

utilities announce the DR event [4]. While there are various optimization strategies and polices for reducing energy consumption
[57], they mainly focus on controlling passive components such as
load control and shift but not active components such as renewable
energy sources.
In addition to nationwide efforts on energy management, strategies for small size power grids have been actively discussed.
Microgrid is the downscaled cells of the Smart Grid [8]. Microgrid
can be operated as a dispatchable load of the conventional power
system as well as an independent power system for small-scale
areas. By considering distributed energy resources (DER), such as
wind and solar, and energy storage system (ESS) in conjunction
with scheduling and RTC schemes, Microgrid can achieve several
advantages such as enhancement of local reliability, reduction of
feeder losses, and increased efciency whose main purpose is to
minimize either the total electricity cost or CO2 emission from the
designed Microgrid. For load control, Microgrid and Smart Grid are
able to encourage customers to participate in peak power consumption control, which results in load reduction. However, this action
is very limited; it is highly likely for customers not to reduce their
power consumption. On the other hand, the load in buildings can
be controlled directly by BEMS, so that the strategy for the building
is different with that of the Microgrid.
Alternative efforts for efcient energy management have also
been discussed in areas smaller than Microgrid, where the most

240

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

representative example is BEMS [9]. The basic concept of the small

sized Microgrid has been applied to homes, buildings, and industrial plants [10]. Unlike in the past, where passive components
were the dominant energy consumption load, the more recent
end components have become active. This means that they may
have intelligence with regards to energy management, so that the
scheduling and RTC can include renewable energy, ESS, and load
controls with feedback information [11]. One key part of BEMS is
the HVAC (Heating Ventilating Air Conditioning) system, which has
been studied in [12,13]. In HVAC, model predictive control (MPC)
is widely deployed for its simplicity if a system is modeled, the
complexity required for its implementation can be signicantly
reduced [14,15]. Moreover, MPC has been used to control energy
ows in Microgrid [1618] and household environments [19]. A
general MPC scheme is not suitable for event driven situations, i.e.
a target event such as DR event should be accomplished for a short
period of time. This is mainly because of the static time window
size (i.e., the time horizon considered for future energy ow scheduling optimization). In addition, since the general MPC does not
include an RTC scheme, the stable energy ow control may not be
accurately optimized in a dynamically changing environment.
In order to overcome these limitations of prior works, the
proposed framework includes prediction, long-term scheduling
and RTC. The main contributions of this paper are summarized as
follows. The proposed long-term scheduling and RTC of energy
ow (such as energy to/from the grid, energy to/from the ESS,
and load shift) in a building equipped with renewable power generators and ESS: (i) addressing prediction on pricing, renewable
energy, and load pattern to formulate the optimization problem
for long term scheduling, (ii) considering management scheme for
special events such as DR event, and (iii) addressing hour-based
parameter setup of RTC for stable control against to dynamics such
as load and renewable energy uctuation. In our simulations, we
show that the proposed approach can reduce total electricity cost
by approximately 15% compared to conventional RTC strategies.
The remainder of this paper is organized as follows. In Section 2,
the overview of the considered framework is presented. In Section 3, the scheduling algorithm and RTC scheme are proposed. In
Section 4, the performance evaluation of the proposed framework

is discussed with a set of actual measured and collected data in

Seoul, Republic of Korea, and the conclusion is drawn in Section 5.
2. Overview of the framework
2.1. Brief overview of BEMS
In this paper, we consider a building energy system shown in
Fig. 1, where it consists of the components such as DERs, ESS and
loads, and these are controlled by the BEMS. The power lines are
connected to a power grid. The loads in the building can be categorized into base type or controllable type. Several xed facilities
and servers in the building may be considered as the base type.
The controllable type may include electric heating appliances and
air conditioning systems, where the loads from the controllable
type can be scheduled and controlled by BEMS with a deployed
energy management strategy. BEMS collects the information from
both inside and outside of the building, which includes the recent
status of all components of the building (from inside of the building) and the current pricing rate, climate status, etc. (from outside
of the building). Such information can be exchanged and shared
based on the network infrastructures (e.g., power line communication, Zigbee, conventional wired/wireless networks, etc.). Based
on the collected information, BEMS performs the management of
energy in the building in order to minimize either running cost or
CO2 emission.
2.2. The framework of the proposed BEMS
The main focus of this paper is on BEMS, in particular a new proposed framework of BEMS, and its components such as prediction,
long-term scheduling and RTC strategies. Fig. 2 shows the proposed
framework of BEMS. The main objective of the proposed BEMS is
the minimization of total cost for electricity which is consumed
for a specic period (in general 115 days). Before the long term
scheduling, future electricity pricing, future renewable power generation, and future load usage pattern should be forecasted for the
next N hours; N is the product of 24 h and the considered days. Based
on the historical data of pricing, renewable power, and load pattern,

Fig. 1. Energy system of buildings.

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

241

[20,21]. In this paper, we deploy the dynamic regression approach

proposed in [21], where price pt at time t is determined by:
pt = c +

K


wld Bl dt +

l=1

K


wl Bl pt + t

l=1

where c is a constant and dt denotes the demand at time t. B is a

backshift operator determined by Bl dt = dt1 , Bl pt = ptl . The selecp
tion of wld and wl is achieved with several iterations in order for
the model not to follow a white noise process. t denotes an error
at time t. More details can be found in [21].

Fig. 2. The proposed framework of BEMS.

hour-based value for each is forecasted for the considered period.

The forecasted values are sent to the long-term scheduler, and
the optimization problem for the long-term scheduling is formulated. The cost function of the optimization problem is to minimize
the total cost for operating all electric facilities in the building for
the considered period. Among the all hour-based results from the
optimization, the result for the next hour is utilized for setup of
parameters of the RTC part. After computing the current parameters required to control all components in the building, the BEMS
performs RTC. The control signals from RTC are sent to the interface
to control the components. At the end of each time slot (in general
one hour), the system updates the accumulated power consumed
at loads and the forecasted value.
With this update, prediction for the next N 1 hours is conducted. The whole process from the prediction to RTC is repeated
until the end of the target period. For example, if the considered
period is a week, the 168 (24 h 7 days) predicted values for pricing,
renewable energy, and load pattern are the output of the prediction
part. These results are used in the long term scheduling for allocating energy ows for the next 168 h. Among the allocated energy
ows, the values only for the next hour are utilized for the parameter setup in the RTC. And then, RTC is conducted for an hour with
the setup, and after one hour the BEMS updates the current status.
Based on this status, prediction for the next 167 h is conducted. The
process from the prediction to RTC is repeated 168 times, which is
the considered period of the BEMS.

2.3. Prediction and forecast mechanisms

2.3.2. Photovoltaic forecasting

For solar energy prediction, proper methodologies are appropriately deployed based on the forecast horizon. For the forecast
for very short time horizon among the methodologies, persistence
forecast is generally considered. It is performed by considering
current or recent PV power plants or radiometer outputs and is
extrapolated to account for changing sun angles. The accuracy of
persistence forecast may decrease sharply with forecast duration,
as cloudiness changes more with longer forecast duration. In addition to the persistence forecast, a prediction method based on total
sky imagery can be used to forecast from real-time (nowcast) up
to 1530 min. Numerical Weather Prediction (NWP) is the best
forecasting technique if the long time horizons are more than 5 h
[22,23].
2.3.3. Wind forecasting
The wind power is generally estimated based on the combination of one or multiple wind power forecast models. One of the
simplest models is the persistence model also used in PV forecasting which is known as a very effective model for very short
forecast horizons (up to 6 h). There are more advanced wind power
forecast models that are referred to wind power forecast systems
for relatively long forecast horizons. Note that they require more
computational complexity in general [22,24].
2.3.4. Load forecasting
Load forecast can be categorized into short-term forecast (usually from 1 h to 1 week), medium forecast (usually from a week
to a year), and long-term forecast (longer than a year). For load
forecasting several factors should be considered, such as time factors, weather data, and possible customers classes. Most of load
pattern forecasting schemes use statistical techniques or articial
intelligence algorithms such as regression, neural networks, fuzzy
logic, and expert systems [25]. In this paper, we adopt the prediction model proposed in [26], where a transfer function model for
load prediction is expressed as:
yt =

m

w (B)
i

i=1

As discussed in Section 2.2, BEMS needs to predict electricity

pricing, the amount of renewable energies, load usage patterns, etc.
For more accurate predictions, we deploy the following prediction
schemes in the proposed algorithm.

i (B)

(1 B)di Xi,tbi +

q (B)
t
p (B)

where

q (B) = 1

q


l Bl , p (B) = 1

l=1

2.3.1. Pricing prediction

There are three dominant factors that signicantly inuence
the wholesale power markets in the future: the price of natural
gas, the cost of CO2 production, and renewable generator developments. For the prediction of pricing, Auto Regressive Integrated
Moving Average (ARIMA) models have been widely used. Alternatively, Articial Neural Networks (ANN) techniques that have been
used for load forecasting, are currently used for price prediction


l=1

l Bl , i (B)

l=1

= 1

p


l Bl , wi (B) = 1

s


wl Bl

l=1

In these expressions, yt is a transfer function of tth time series loads

and Xi,tbi is (t bi )th time series of Xi which can be temperature,
economic index, etc. having an inuence on the power consumption. di indicates the continuous difference order of Xi . B, t and m

242

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

Fig. 3. Conceptual path to the optimal point. The proposed strategy begins at the starting point and goes to the optimal point with the best effort based on the current
available data.

denote the backward shift operator, the tth disturbance term and
the number interpretation variables, respectively.
3. Proposed strategies for long term scheduling and
real-time control
In this section, we discuss the proposed long-term scheduling
and RTC strategies. We also provide the corresponding optimization steps in detail. An illustrative example on how the proposed
algorithms work is shown in Fig. 3. With current status information collected, BEMS predicts the hourly based pricing, renewable
energy, and load for specic days (N hours). Based on the prediction results, the optimization problem for long term scheduling is
formulated and solved. The optimal result is depicted as a solid
arrow from the starting point. Although there is a clear optimal
path to the optimal point (depicted by dotted line), the solid arrow
is not oriented the optimal direction. This is because the scheduler
is based on the forecasted values which contain error. The dynamics
of renewable energy and load usage may incur swings in random
directions (depicted with the dash-dot line in Fig. 3). This can be
corrected by RTC and parameter setup of RTC based on the optimal solution from the long-term scheduler. After 1 h, the BEMS
again predicts pricing, renewable energy, and load pattern with the
updated current status information for the remaining hours of the
considered period (i.e., N 1 hours). Subsequently, the optimization problem for scheduling of the next N 1 hours is reformulated.
This process is iterated until the real-time control based on the last
scheduling result is nished.
3.1. Optimization problem for long term scheduling
In this chapter, details of optimization formulation for long
term scheduling are presented. Notations used in the optimization problems for long term scheduling and real-time control are
summarized in Table 1.
The electricity pricing, sales pricing, renewable energy
(including solar and wind power), and loads are denoted
by C  [C1 , . . ., CN ]T , S  [S1 , . . ., SN ]T , R  [R1 , . . ., RN ]T , and
L  [L1 , . . ., LN ]T , respectively. In order to minimize the total cost,
energy ow should be scheduled. Moreover, as discussed, loads can
be either base type or controllable type. If the loads are base type,
they cannot be controlled. However, if the loads are controllable
type, they can be shifted over a period of time. The variables
of energy ow and variable loads considered in this paper are
T

denoted by X  [X1 , . . ., XN ] , Y  [Y1 , . . ., YN ] , B 

s T,
B1s , . . ., BN

T

T

u
ctrl
Bu  B1u , . . ., BN
, and Lctrl  L1ctrl , . . ., LN
representing energy
from the grid, energy to the grid, energy from the ESS, energy to the
ESS, and the controllable load, respectively. The subscript indicates
the number of hours ahead which needs to be determined by the
long-term scheduler. For example, a real value scalar, X1 denotes
the energy from the grid during the rst hour.
The objective function is dened as the total cost for energy that
has been consumed for operating all components in the building
during a certain period of time. In addition, a special event such
as a DR event can happen during the period. Thus, the objective
function can be expressed as:

CT X ST Y C Inct

(1)

where CInct is the incentive for participating in the DR event. The

goal is to minimize the total cost in (1) subject to several constraints,
stated as follows. The total energy ow in the building should be
balanced,
X + R + Bu Y Bs = Lbase +Lctrl .

(2)

Note that the energy ow balance is performed at each time-step

that is one hour in this paper and the system is assumed to be in a
quasi-steady state in each time-step.
In the perspective of BEMS, energy from the power grid, energy
from the renewable power generators, and energy from the ESS
are positive, however, energy delivered to the power grid, ESS,
and consumed by loads are negative. The energy usage shift of the
controllable load and corresponding time should be limited. Thus,
T

T
T base
Lmin Lctrl Lmax , Iidx
Lctrl
d
d = 1 Ld 1 Ld

(3)

where
T

ctrl
ctrl
Lctrl
d = [L24(d1)+1 , . . ., L24(d1)+24 ] ,

Ld = [L24(d1)+1 , . . ., L24(d1)+24 ]T ,
idx
idx
Iidx
d = [I24(d1)+1 , ..., I24(d1)+24 ]

for 1 d N/24, Ididx = 1 if di is within the load shift time and Ididx = 0
i

otherwise. The rst constraint of (3) is the limitation on the amount

of energy which can be shifted by the controlling load and the

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

243

Table 1
Notations of long-term scheduling and RTC parameter setup.
C
R
Cn
S
I
X
Xn
Iidx
Y
Yn
L
Bs
Bns
Lbase
Bu
Bnu
Lctrl
Bint
Bmax
Bmin

CInct
Rn
Robs
Imax
Imin
Sn
XRTC
Iidx
d
YRTC
Ln
s
BRTC
Lbase
d
u
BRTC
Lctrl
d
ctrl
Lobs
Lmax
Lmin
N

Grid energy pricing

Energy from renewables
Grid energy pricing at n hours
Sales pricing to grid
DR incentive rate
Energy from grid
Energy from grid at n hours
Index for load shift
Energy to grid
Energy to grid at n hours
Total load
Energy to ESS
Energy to ESS at n hours
Base load
Energy from ESS
Energy from ESS at n hours
Controllable load
Initial energy in ESS
Maximum capacity of ESS
Minimum capacity of ESS

second constraint of (3) states that load shift cannot be conducted

over a day.1
The energy from and to the power grid where the maximum
value is the capacity of the power line connected to the building is
expressed as:
0 X Xmax , 0 Y min{U, R + Bu },

(4)

last three constraints of (7). Thus, the optimization problem for

long term scheduling is correspondingly expressed as the following
linear programming problem:
minimize CT X ST Y C Inct
subject to X + R + Bu Y Bs Lbase +Lctrl ,

where U denotes the capacity of a transformer in the power distribution system.

DR incentive that may be expressed differently depending on a
DR incentive policy can be expressed as:
C Inct = max{0, I(Bint + 1T (Bs Bu ))}
I min I(Bint + 1T (Bs Bu )) I max

0 Bis Bmax Bint

0 Biu Bint +

i1

j=1

Bjs +

j=1
i1

Bjs

i1


X + R + Bu Y Bs Lbase +Lctrl ,
T

T
T base , Iidx Lctrl 1T L 1T Lbase ,
Lctrl
Iidx
d
d
d 1 Ld 1 Ld
d
d
d

Lmin Lctrl Lmax , 0 X Xmax , 0 Y U, 0 Y R+Bu ,

(5)
0 Bis Bmax Bint

where the incentive for participating in the DR event is provided

when energy reduction during the DR event is larger than a threshold level. Otherwise, the incentive would be zero even though BEMS
would save a certain amount of energy during the event. In addition, the maximum incentive rate is limited, which means excessive
energy savings during the event is not protable. Accordingly, intelligent energy savings in the ESS, which will be used during the DR
event, provides the best economic benet. Finally, the energy to the
ESS and energy from the ESS are expressed as:
i1


Variable for calculating DR incentive

Energy from renewables in n hours
Real time observed renewables
Maximum DR incentive rate
Minimum DR incentive rate
Sales pricing to grid at n hours
RTC parameters for energy from grid
Index for load shift on the dth day
RTC parameters for energy to grid
Total load at n hours
RTC parameters for energy from ESS
Base load on the dth day
RTC Parameter for energy from ESS
Controllable load on the dth day
Real time controllable load usage
Maximum controllable load
Minimum controllable load
N = hd (total hours)
Variable for RTC parameter setup
Days

Bjs +

j=1
i1

Bjs

i1


Bju , 1 i N,

j=1

Bju , 1 i N,

j=1
j=1
C Inct 0, C Inct I(Bint + 1T (Bs Bu )),

I min I(Bint + 1T (Bs Bu )) I max

(7)
where
T

ctrl
ctrl
Lctrl
d = [L24(d1)+1 , . . ., L24(d1)+24 ] ,

Bju , 1 i N

j=1

0 Biu Bint +

i1


i1


(6)

Bju , 1 i N

Ld = [L24(d1)+1 , . . ., L24(d1)+24 ]T ,
idx
idx
Iidx
d = [I24(d1)+1 , . . ., I24(d1)+24 ]

j=1

where the energy to the ESS should not exceed the maximum
capacity of the ESS and the energy from the ESS should not exceed
the remaining energy in ESS at every time horizon.
In order to efciently nd the optimal solution from the optimization problem formulated above, several techniques from a
linear programming can be used [27], i.e., the equality constraints
in (2) and (3) can equivalently be expressed as the rst four
constraints of (7) which are all inequality forms. Moreover, the
non-linear function, max function, in (5), can be expressed as the

1
Load shift means the amount of shifted power by changing the operation time of
loads, while load shift time means acceptable time line in which loads can be shifted
without incurring signicant inconvenience to residences in a building.

for 1 d N/24, Ididx = 1 if di is within the load shift time and Ididx = 0
otherwise.

3.2. Parameter setup for RTC

In the optimal solution for all time horizons from the scheduler,
the values for the next hour are used to set the parameters of the
RTC. The fundamental role of the RTC is to compute the real-time
energy ow control parameters for real time operations (in general, less than several seconds). The parameters forwarded from
the long-term scheduler provide the information about hour-based
energy ows for a predetermined duration (e.g., a week, a month,
etc.) to the RTC, which determine the optimal direction of energy
ows in the long-term perspective. Based on the optimal direction

244

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

Table 2
Error values (%) for prediction.
Day ahead

Price error (%)

PV error (%)

Wind error (%)

Load error (%)

1
2
3
4
5
6
7

2.6
3.3
2.7
1.9
2.5
3.7
4.0

16
17
20
22
23
27
30

8
20
30
32
34
36
40

8
13
14
16
17
18
19

4. Simulation results

Fig. 4. Proposed RTC structure.

4.1. System description

for next hours (i.e., the parameters forwarded from the long-term
scheduler), RTC can adaptively control the short-term energy ows.
In addition, the RTC plays a key role in the proposed algorithm
by appropriately and iteratively dealing with the gaps between
predicted and actually measured data. Recall that the long-term
scheduling is optimized based on the predicted pricing, amount of
renewable energy, and load usage patterns. Thus, it is inevitable
to have (potentially small) gaps between the optimally obtained
and actually measured prices, the amount of generated renewable energy, and the load usage patterns due to the environment
dynamics.
In order both to remain the optimal direction and to deal with
s , Bu , are
the gaps, optimal real-time energy ows, XRTC , YRTC , BRTC
RTC
proportionally determined as X1 , Y1 , B1u , and B1s , respectively. The
real-time energy balance condition is thus represented by
ctrl
X1 + Robs + B1u Y1 B1s = L1base + Lobs

(8)

which means that is computed as

ctrl
= (L1base + Lobs
Robs )/(X1 + B1u Y1 B1s )

(9)

base and R
with the observed Lobs
obs . Thus, can be considered as a
real-time correction term for the gaps. With , the parameters for
s , and Bu
are nally determined by
XRTC , YRTC , BRTC
RTC
u
s
XRTC = X1 , YRTC = Y1 , BRTC
= B1u , BRTC
= B1s .

(10)

The updated parameters enable the RTC to more accurately control the energy ows in the presence of gaps from the long-term
scheduler. The RTC is depicted in Fig. 4.
The control logic can be implemented based on the various control algorithms such as PID (Proportional Integral Derivative) [28],
GMV (Generalized Minimum Variance) [29], and GPC (Generalized
Predictive Controller) [30]. Algorithm 1 summarizes the proposed
BEMS strategy including the long-term scheduling algorithm and
the RTC algorithm.
Algorithm 1 (Proposed BEMS strategy).
Given: the considered period N : =24 d
Repeat:
predict C, S, L, R of the next N hours with observed values
1:
obtain X1 , Y1 , B1u , B1s , L1ctrl with (7)
2:
3:
obtain parameters with (10)
4:
control energy ow with the parameters
at the end of each hour
5:
if Cfrcst =Cob , Sfrcst =Sob ,Rfrcst =Rob , and Lfrcst =Lob
6:
N : = N 1 and do Step 3 with X2 , Y2 , B2u , B2s , L2ctrl
else
N : = N 1 and do Step 1
7:
Until: N=0

In this section, we evaluate the effectiveness of the proposed

framework and strategies in terms of the cost reduction using actual
data. The data are collected for every hour from 12 AM of January
1st, 2010 to 11 PM of December 31st, 2011. For load usage history,
we gather the observed data from a commercial building (contract
power level is 6250 kW) in Seoul, Republic of Korea. The amount
of generated solar and wind energies are computed based on the
observed solar radiation and wind speed/direction data in Seoul. In
order to compute the amount of energy from the PV (photovoltaic)
generator, an array of TBD 125125-96-P PV panel (maximum
power of 240 W) is used to determine the observed solar radiation
rate (MJ/m2 ). The amount of energy from wind power generators
is computed based on the hourly observed wind speed (m/s) and
direction (degree)) with a series of 5 wind turbines of Enercon E33
300 (maximum power is 300 kW). In this paper, the assumed total
maximum power generated from the renewable power generator
is about 3000 kW for fully analyzing the effect of scheduling. This
assumption is reasonable because a series of several wind turbines
were previously installed at the Bahrain World Trade Center. The
maximum capacity of ESS is assumed to be 6250 kW as in SMES.
We deploy the following existing models: auto regression (AR)
and auto regression integrated moving average (ARIMA) [26,31],
numerical weather prediction [22], and dynamic regression [21] for
the prediction of loads, amount of renewable energies, and electricity pricing, respectively, which is discussed in Section 2.3. In order
to make our simulation environments be more realistic, we intentionally corrupt the data actually collected from the commercial
building based on statistical noise (statistically random uctuations). More specically, Gaussian random errors are generated
with mean shown in Table 2 (which is obtained from [21,22,26,31])
and unit variance. The errors are injected in to the actually collected
data, eventually generating prediction results for electricity pricing,
PV and wind power generation, and load usage patterns.
In our simulations, we assume that BEMS is kept current with
the status of all equipment and is able to control them. Moreover,
the EMS is able to access the data such as RTP (Real-Time Pricing)
released by power utilities and additional information to predict
pricing, solar power, wind power, and loads. The time horizon for
load shifting ranges from 6 AM to 6 PM during the weekdays of
the considered commercial building. It is also assumed that 60% of
the total loads are not controllable (i.e., base loads), while the rest
of the loads are controlled by the proposed scheduling strategies
[32]. It is also assumed that the initial stored energy in the ESS is
0 kW. Finally, when DR Incentive policy is active, incentives can
be provided in proportion to the total energy saved during the DR
event period (about 2 h). The maximum incentive is determined
based on the contract power level of the building (in our paper,
6250 (kW) Incentive rate (dollars/kW)).
In this simulation, we consider the following service scenarios.
Based on an announced DR event from a power utility company, the
BEMS can deploy scheduling and RTC algorithms a week in advance

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

Table 3
Total cost of operating a building in a year based on the proposed strategy.

a)
0.16

0.14

Price (Dollars)

0.12

Season

Proposed (baseline)

Only with scheduling

Conventional RTC

Winter
Summer
Spring
Autumn

0% ($21,398)
0% ($23,670)
0% ($17,114)
0% ($18,204)

+4.72% ($22,408)
+4.93% ($24,837)
+4.23% ($17,838)
+4.37% ($19,000)

+11.17% ($23,788)
+11.56% ($26,406)
+7.21% ($18,348)
+7.17% ($19,509)

0.1

4.2. Overall performance of proposed BEMS strategy

0.08
Winter
Summer
Spring
Fall

0.06

0.04

0.02

10

15

20

Time (0h to 23h)

b).
0.2
Price from Grid
Price to Grid
Incentive Rate

0.18

Price (Dollars)

245

0.16

0.14

0.12

0.1

0.08

10

15

20

Time (0h to 23h)

Recall that the priorities of the energy ows are determined

based on the price as the objective function of the long-term scheduling algorithm is to minimize the total cost of electricity for
demand (discussed in Section 3.1). Hence, the proposed strategy
can decrease the total cost further if there is a larger difference
between minimum and maximum prices, as well as a high average
price level. The overall performance comparison of several BEMS
strategies is shown in Table 3.
The results, the total cost for electricity in a building, decrease
with the proposed algorithm by introducing the proposed strategy
to BEMS. For the strategy where only scheduling is implemented,
the cost decreased about 45%. For the strategy using conventional RTC , the cost decreased approximately 712%. As can been
seen from Table 3, the performance improvement across seasons
between the proposed strategy and that of using the conventional
RTC scheme is apparent. During the summer and winter seasons,
the average of the pricing rate is higher than that of the spring
and fall seasons. In addition, the pricing rate during the summer
and winter is more volatile than during the spring and fall seasons (Fig. 5(a)). This means that when the pricing rate is uctuated,
scheduling does contribute signicantly to the decrease of total
running cost. The relation between performance and uctuation is
analyzed in detail in Section 4.4. When comparing the BEMS strategy to purely scheduling strategy across seasons, there is about 5%
cost savings in total running cost. When the BEMS scheme is implemented with a long-term scheduler, the volatility of pricing is no
longer a crucial factor on decreasing the running cost. The over cost
compared to the proposed strategy is from the unexpected load use
or dynamics of solar and wind power because of the absence of RTC.

Fig. 5. Comparison of various price rates.

4.3. Performance of the proposed BEMS

of the DR event. BEMS tries to optimally schedule and control
energy ows during a week such that economical prot becomes
maximized by participating in the DR event. The performances
of the proposed strategy are compared with the other strategies,
i.e., only with scheduling (the proposed long-term scheduling) and
conventional RTC.2
We assume the seasonal average pricing rates and the comparison of pricing rates between buying, selling, and incentive as shown
in Fig. 5. Electricity pricing rates during the summer and winter
seasons are more expensive and have higher uctuations than during the spring and fall. This is because the demand for electricity
is higher during the summer and winter seasons for cooling and
heating the building. This can be observed from Fig. 5(a). Fig. 5(b)
depicts the pricing rates for buying, selling, and incentive over a
one day period. For the participation of the DR event, the incentive
rate is set higher than the electric pricing rate.

2
In this paper, conventional RTC means the control of the current energy ow
without considering the prediction of the future power usage. For performance
comparison, we use the approach discussed in [33] as a conventional RTC.

In this scenario, the BEMS performance is analyzed and quantied with the use of DR events. We compare the energy ows
and total cost between other mechanisms. As previously discussed,
because this is a commercial building, we assume that the loads
can be shifted anywhere between 6 AM to 6 PM. In addition, the
maximum capacity of the ESS is determined by the contract power
level of the building. In this simulation, it is set to 6250 kW. For
optimized performance, we can intuitively expect that energy consumption by the load should be minimized during high pricing rate
periods, and maximized during low pricing rate periods. As shown
in Fig. 6, the pricing rate in the morning is lower than in the daytime. Thus, the proposed strategy for BEMS makes decisions such
that the usage of the controllable load is shifted from daytime to
morning as depicted in Fig. 6.
Fig. 7 shows the amount of corresponding energy ows in the
building. As shown, for example, in the period of 40 h and 43 h of
Fig. 7, we can conclude that there is small amount of energy ow
from the grid (Fig. 7(a)) and also a small amount of energy stored
in the ESS (Fig. 7(d)) when the price is high, while there is large
amount of energy ows to grid (Fig. 7(b)) and from the ESS (Fig. 7(c))
to provide energy for satisfying the demand from the loads.

246

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

4

2500

x 10

Applied DR by Scheduler

2.5

Without Strategies

Cost by
Cost by
Cost by
Optimal

Total Cost (Dollars)

1500

1000

1.5

500

0.5

20

40
60
80
100
120
140
Time, From Monday 0h to Sunday 23h (Total 168h)

160

20

Fig. 6. Comparison of load consumption. The dotted line represents an applied

DR strategy with proposed scheduling, while the solid line represents no applied
strategy.

Fig. 8 shows the resulting total cost when different strategies

are deployed. The proposed strategy can reduce the total cost more
than the other two strategies, i.e., the proposed strategy reduces
total cost by approximately 5% more than the scheduling-only
strategy, and approximately 13% more than with the conventional

100

120

140

160

6000
200
4000
150
2000
100

Energy from ESS by Scheduler

6000

Energy from ESS (kW)

Energy from Grid (kW)

80

RTC strategy. When the proposed scheme is compared to the optimal cost (the case that all predicted values are exactly same as real
and observed values), the cost savings is about 5% less due to the
error and dynamics between an optimal case and a general case.

c).
250

60

Fig. 8. Comparison of cost trajectories.

Energy from Grid by Scheduler

8000

Electricity Price from Grid (Dollars)

a)

40

Time, From Monday 0h to Sunday 23h (Total 168h)

250

200

4000

150
2000

Electricity Price from Grid (Dollars)

Total Load Consumption (kW)

2000

proposed st rategy
scheduling-only
conventional RTC
Cost

100

20

40

60

80

100

120

140

160
0

Time, From Monday 0h to Sunday 23h (Total 168h)

20

40

60

80

100

120

140

160

Time, From Monday 0h to Sunday 23h (Total 168h)

d).
250

200

4000

150
2000

100

6000

Electricity Price from Grid (Dollars)

Energy to Grid (kW)

6000

Energy to ESS by Scheduler

250

200

4000

150
2000

100
0

20

40

60

80

100

120

140

Time, From Monday 0h to Sunday 23h (Total 168h)

160

20

40

60

Fig. 7. The result of energy ow amount by the proposed algorithm.

80

100

120

140

160

Electricity Price from Grid (Dollars)

Energy to Grid by Scheduler

Energy to ESS (kW)

b).

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

Table 4
Comparison of total cost with different time window sizes.

5. Conclusion

Day ahead

Proposed (baseline)

Only with scheduling

Conventional RTC

1
2
3
4
5
6
7

0%($3,682)
0%($6,528)
0%($9,847)
0%($13,705)
0%($16,026)
0%($18,408)
0%($21,398)

+10.57%($4,072)
+6.29%($6,938)
+7.09%($10,545)
+7.04%($14,670)
+5.73%($16,944)
+5.01%($19,330)
+4.72%($22,408)

+11.14%($4,092)
+14.37%($7,466)
+15.55%($11,378)
+16.73%($15,998)
+14.06%($18,279)
+11.00%($20,433)
+11.17%($23,788)

0.24
0.22
0.2

Price (Dollars)

247

0.18
0.16
0.14

In this paper, we propose a novel strategy that can optimally

control the energy ows throughout various components of a
building, thereby minimizing the total cost. The proposed strategy focuses on optimal energy allocations and ow controls in
particular when there is a target DR event. The proposed strategy
consists of three modules that are prediction, long-term scheduling
and RTC. The long-term scheduler can make optimal decisions (i.e.,
optimal energy ows) based on several predicted pricing, available
solar and wind powers, and load consumptions. These decisions
are then used by RTC for more detailed control of energy ows
for unpredicted load usages and the amount of generated renewable energy. In order to nd the optimal strategy, we formulate
the energy ow control in a building as an optimization problem.
The simulation results conrm that the proposed strategy leads to
total cost reduction. The proposed strategy can reduce the total
cost of energy required for building operation by approximately
12% when compared to a conventional RTC scheme. Based on the
analysis of the simulation results, the improvement achieved by
the proposed strategy depends on the price uctuation. In addition, simulation results show that the selection of the time window
size for long-term scheduling is signicantly dependent on the
price uctuations. Adaptive time window size selection followed
by monitoring the uctuation leads the more cost savings.

0.12

Acknowledgments
0.1
0.08

20

40

60

80

100

120

140

160

Time, From Monday 0h to Sunday 23h (Total 168h)

Fig. 9. Price for comparing the proposed algorithm performance.

This research was supported in part by Basic Science Research

Program through the National Research Foundation of Korea (NRF)
funded by the Ministry of Education (No. NRF-2010-0009717) and
in part by the MSIP (Ministry of Science, ICT and Future Planning),
Korea, under the ITRC (Information Technology Research Center)
support program (NIPA-2013-H0301-13-1002) supervised by the
NIPA (National IT Industry Promotion Agency). The material in this
paper was presented in part at the Global Communications Conference (Globecom 2013) Atlanta, GA, USA 2013.

4.4. Performance of BEMS with different time window size

References
The required number of days for long term scheduling is analyzed with regards to optimal BEMS performance. As shown in
Table 4, the proposed strategy performance does not highly depend
on the number of days in which the BEMS should consider in
the long term scheduler when compared to the conventional RTC
scheme. Fig. 5 provides a hint for the result in Table 4. If the price
uctuation is higher, the proposed strategy saves more energy;
for the rst four days, the cost reduction compared to with the
conventional RTC scheme continues increasing due to large uctuations. This is especially true for day 2, where the uctuation
is very large and thus prominently increases cost reduction. The
price uctuations for day 3 and day 4 remain high, so the cost
reduction continues increasing. During day 5 and day 6, the price
change plateaus and the cost reduction decreases accordingly. In
the last day, the price uctuation is higher than the two previous
days, which results again in the increase of the cost reduction.
Table 4 and Fig. 9 clearly support the following opinion that
the series of days having the highest price uctuation should be
the highest priority to be considered for setup of the time window
size. When compared to the results from the scheduling-only strategy, the larger time window of the proposed strategy provided less
improvement. The only difference between the proposed strategy
and scheduling-only strategy is the proposed RTC support to deal
with the unexpected load usage and dynamics of solar and wind
energy. For short time window sizes, these dynamics are dominant, but become less dominant as time window size increases.
And therefore, the accumulated decrease rate drops.

[1] A. Ipakchi, F. Albuyeh, Grid of the future, IEEE Power and Energy Magazine 8
(2009) 5262.
[2] S. Borenstein, M. Jaske, A. Rosenfeld, Dynamic pricing, advanced metering and
demand response in electricity markets, White Paper (2009).
[3] R.Z. Freirea, G.H.C. Oliveirab, N. Mendesc, Predictive controllers for thermal
comfort optimization and energy savings, Energy and Buildings 40 (2008)
13531365.
[4] K. Spees, L.B. Lave, Demand response and electricity market efciency, Electricity Journal 20 (2007) 6985.
[5] J.A. Wright, H.A. Loosemore, R. Farmani, Optimization of building thermal
design and control by multi-criterion genetic algorithm, Energy and Buildings
34 (2002) 959972.
[6] D. Bourgeoisa, C. Reinharta, I. Macdonaldb, Adding advanced behavioural models in whole building energy simulation: a study on the total energy impact
of manual and automated lighting control, Energy and Buildings 38 (2006)
814823.
[7] B. Ramanathan, V. Vittal, A framework for evaluation of advanced direct load
control with minimum disruption, IEEE Transactions on Power Systems 23
(2008) 16811688.
[8] M. Manfren, P. Caputo, G. Costa, Paradigm shift in urban energy systems
through distributed generation: methods and models, Applied Energy 88
(2011) 10321048.
[9] J. Figueiredoa, J.S. da Costab, A scada system for energy management in intelligent buildings, Energy and Buildings 49 (2012) 8598.
[10] X. Guan, Z. Xu, Q.S. Jia, Energy-efcient buildings facilitated by microgrid, IEEE
Transactions on Smart Grid 1 (2010) 243252.
[11] T. Chowa, G. Zhangb, Z. Lina, C. Songa, Global optimization of absorption chiller
system by genetic algorithm and neural network, Energy and Buildings 34
(2002) 103109.
[12] L. Prez-Lombarda, J. Ortizb, J.F. Coronela, I.R. Maestrec, A review of hvac systems
requirements in building energy regulations, Energy and Buildings 43 (2011)
255268.
[13] Y. Bichioua, M. Krartib, Optimization of envelope and hvac systems selection
for residential buildings, Energy and Buildings 43 (2011) 33733382.

248

S.J. Kang et al. / Energy and Buildings 75 (2014) 239248

[14] F. Oldewurtela, A. Parisiob, C.N. Jonesc, D. Gyalistrasa, M. Gwerderd, V. Stauche,

B. Lehmannf, M. Moraria, Use of model predictive control and weather forecasts
for energy efcient building climate control, Energy and Buildings 45 (2012)
1527.
[15] S. Qin, T. Badgwell, A survey of industrial model predictive control technology,
Control Engineering Practice 11 (2003) 733764.
[16] A. Parisio, L. Glielmo, Energy efcient microgrid management using
model predictive control, in: 50th IEEE Conference on Decision and Control and European Control Conference (CDC-ECC), Orlando, USA, 2011,
pp. 54495454.
[17] A. Hooshmanda, H. Malkib, J. Mohammadpourc, Power ow management of
microgrid networks using model predictive control, Computers and Mathematics with Applications 64 (2012) 869876.
[18] P.O. Kriett, M. Salani, Optimal control of a residential microgrid, Energy 42
(2012) 321330.
[19] R. Negenborn, M. Houwing, J.D. Schutter, J. Hellendoorn, Model predictive control for residential energy resources using a mixed-logical dynamic model, in:
IEEE International Conference on Networking, Sensing and Control (ICNSC09),
Okayama, Japan, 2009, pp. 702707.
[20] Northwest Power and Conservation Council, Sixth Northwest Conservation and
Electric Power Plan, Technical Report, 2010.
[21] F.J. Nogales, J. Contreras, A.J. Conejo, R. Espinola, Forecasting next-day electricity prices by time series models, IEEE Transactions on Power Systems 17 (2002)
342348.
[22] W. Glassley, J. Kleissl, C. van Dam, H. Shiu, J. Huang, G. Braun, R. Holland,
California Renewable Energy Forecasting, Resource Data and Mapping, Technical Report, Regents of the University of California, 2010.
[23] S. Pelland, G. Galanis, G. Kallos, Solar and photovoltaic forecasting through
post-processing of the global environmental multiscale numerical weather

[24]
[25]
[26]

[27]
[28]

[29]

[30]
[31]

[32]

[33]

prediction model, Progress in Photovoltaics: Research and Applications 1180

(2011) 342348.
R. Kavasseri, K. Seetharaman, Day ahead wind speed forecasting using f-ARIMA
models, Renewable Energy 34 (2009) 13881393.
E.A. Feinberg, D. Genethliou, Applied Mathematics for Restructured Electric
Power Systems, Springer, USA, 2005.
M.Y. Cho, J.C. Hwang, C.S. Chen, Customer short-term load forecasting by using
ARIMA transfer function model, in: International Conference on Energy Management and Power Delivery (EMPD95), 1995, pp. 317322.
S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University Press,
New York, 2004.
Y. Li, K. Ang, G. Chong, Pid control system analysis and design problems,
remedies, and future directions, IEEE Control Systems Magazine 26 (2006)
3241.
S. Bittanti, L. Piroddi, Neural implementation of GMV control schemes based
on afne input/output models, Control Theory and Applications. 144 (1997)
521530.
D. Clarke, C. Mohtadi, P. Tuffs, Generalized predictive control Part 1: The basic
algorithm, Automatica 23 (1997) 137148.
K. Penya, C.E. Borges, D. Agote, I. Fernandez, Short-term load forecasting in airconditioned non-residential buildings, in: IEEE International Symposium on
Industrial Electronics (ISIE11), 2011, pp. 13591364.
R. Lau, S. Ayyorgun, S.C.M.S. Eswaran, A. Misra, S. Bushby, D. Holmberg, Strategy
and modeling for building DR optimization, in: IEEE SmartGridComm, Brussels,
Belgium, 2011, pp. 381386.
H. Dagdouguia, R. Minciardia, A. Ouammia, M. Robbaa, R. Sacilea, Modelling
and control of a hybrid renewable energy system to supply demand of a
green-building, in: International Environmental Modelling and Software Society (iEMSs), Ottawa, Canada, 2010, pp. 18.