1346

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO. 4, NOVEMBER 2003

Switch Placement to Improve System

Reliability for Radial Distribution Systems
with Distributed Generation
Yiming Mao, Student Member, IEEE, and Karen N. Miu, Member, IEEE

AbstractTo improve system reliability for radial distribution

systems with distributed generation (DG) under fault conditions,
switch placement schemes are proposed to form self-supported
areas after fault isolation. Customer priority is also considered
in this problem. The switch placement problem is formulated
as a nondifferentiable, multiobjective optimization problem.
Graph-based algorithms, which incorporate direct load control,
are developed to locate switches. Their results enable DG to support customers continuously in the event of fault. The proposed
algorithms can be applied to unbalanced distribution networks
with single or multiple distributed generators. Simulation results
on a 394-bus distribution system, including priority customers,
and single or multiple distributed generators, were also provided.
Index TermsDistributed generation, fault, radial distribution
systems, switch placement.

I. INTRODUCTION

ISTRIBUTED GENERATION (DG) is expected to play

an increasing role in emerging power systems [1]. Studies
have predicted that DG will be a significant percentage of all
new generation going online. Different resources can be used in
DG, such as diesel and gas. Its impact on distribution systems
may be either positive or negative depending on the systems
operating condition [1], [2], DGs characteristics and location.
Potential positive impacts include
improved system reliability;
loss reduction;
deferment of new generation;
improved power quality.
To achieve the above benefits, DG must be reliable, dispatchable, of appropriate size, and at suitable locations. More important, DGs should be properly coordinated with protection
systems.
In [4], the switch placement problem was formulated to
improve system reliability without considering DG. Now with
DGs, instead of isolating them from the system [1], [3], [5],
[9] in the event of a fault, we can investigate allowing DGs to
support an isolated area by opening switches during upstream
utility outages. As shown in Fig. 1, when a fault occurs, without
Manuscript received January 16, 2003. This work was supported in part by
the National Science Foundation under Contract ECS-9984692 and in part by
the Office of Naval Reaserch under Contract N0014-01-1-0760.
The authors are with the Electrical and Computer Engineering Department, Drexel University, Philadelphia, PA 19104 USA (e-mail:
yiming@io.ece.drexel.edu; miu@ece.drexel.edu).
Digital Object Identifier 10.1109/TPWRS.2003.818609

Fig. 1. Improve system reliability by switch placement and network

reconfiguration.

forming a self-supported area, all of the loads in the area

isolated from the substation will lose power. By allowing DG
to continuously support loads in the inner region, the reliability
of the system would improve. Thus, the main objective of this
paper is to identify corresponding switch locations.
In addition, direct load control [7], [8], has been shown to
be a viable way to control the power system toward desirable
operating conditions. Concurrently, effective load management
could also be employed when determining switching schemes.
We formulate the switch placement problem with DGs as a
nondifferentiable, multiobjective optimization problem subject
to electrical, operational, and network constraints. In addition,
special consideration has been made for customer priority. We
note, standard reliability indices such as SAIDI and SAIFI [10],
are based on steady state information. Therefore, static formulations are developed. Since a multiobjective formulation is selected, tradeoffs between objectives must be made when we design our algorithms.
Graph-based search algorithms have been developed which
incorporate direct load control if available. They employ switch
indices from [6] which yielded high quality service restoration
results using a fixed number of switches. Their results provide
the following critical information:
1) where to install new sectionalizing switches;
2) which existing switch must be opened or closed;
3) which loads should be on and off.

0885-8950/03$17.00 2003 IEEE

MAO AND MIU: SWITCH PLACEMENT TO IMPROVE SYSTEM RELIABILITY FOR RADIAL DISTRIBUTION SYSTEMS

Details of the problem formulations and solution algorithms are

stated in the following sections. Detailed simulation results on
a 394-bus system are also presented.
II. PROBLEM FORMULATION
The switch placement problem is formulated as a nondifferentiable, multiobjective optimization problem subject to electrical, operational, and network constraints. Customer priority
is also taken into account in the formulation. Four objectives
considered are:
(O1) minimize the number of new switches to be installed;
(O2) maximize the amount of priority load in the island;
(O3) maximize the amount of total load in the island;
(O4) minimize the number of switch operations.
Here, the treatment of priority customers (O2) can differ.
First, maximizing priority loads can be treated as an objective.
However, in a more stringent manner, one can consider servicing of priority loads as a constraint. With these differences,
two formulations are developed which share the other three
objectives (O1), (O3), and (O4).
Electrical and operational constraints are considered. Electrical constraints are the three-phase power flow equations
whose solution [11] will allow for operational constraint
checking. For ease of fault location, isolation and coordination of protection devices, a radial network structure will be
preserved in the resulting island.
The two formulations are summarized as follows:
A. Maximizing High Priority Loads (Formulation 1)

(1)
(2)

1347

where
:

three-phase power flow equations;

voltage at bus , phase ;
,
: total three-phase real and reactive power generation at bus ;
,
: total prefault three-phase real and reactive power
generation at bus ;
,
: power ratings of the generator at bus ;
:
percentage of quickly adjustable power of the
DG at bus with respect to its nominal output;
:
current flow entering bus , phase ;
represents how much the generator output can vary without
downgrading power quality such as frequency and voltage magnitude. It can be zero for those generators with nonadjustable
output.
:

B. Servicing All High Priority Loads (Formulation 2)

(10)
(11)
(12)
subject to (5)(9) and
(13)
is the set of prefault priority loads in the area isowhere
lated from substation.
It should be mentioned that if a load bus is single or two phase,
the formulations above would be adjusted accordingly.
Despite the difference, the two formulations are similar in
many aspects. They both consider customer priority and electrical, operational, and network constraints. Also, they have the
same search space, which is limited to the branches found in the
area isolated from substation due to faults.

(3)
III. SYSTEM MODELING
(4)
where
: number of new switches to be installed;
number of switch operations;
total load at bus ;
: set of priority loads in the resulting island;
:
set of load buses in the resulting island
subject to the following constraints:
:
:

Before we explain the detailed algorithms, the system modeling used in the three-phase power flow solver will be briefly
reviewed first.
A. Distributed Generator Modeling

(5)
(6)

(7)

If there is only one generator in an isolated area, it will be

modeled as a slack bus. If there is more than one generator,
the one with the largest adjustable real power output will be
modeled as a slack bus. While the power flow solver used in
this work accommodates both PQ and PV bus models, in this
paper, the remaining distributed generators will be modeled as
PV buses. If the output of a certain generator is not adjustable,
the prefault output P is assigned. The remaining load will be
distributed among adjustable generators in such way that the
real power generation ratios among them are the same ratios
obtained from the prefault condition.
B. Other Component Modeling

(8)
(9)

The modeling used for the power flow solvers of [11] and [12]
is invoked and includes three-phase modeling for nine trans-

1348

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO. 4, NOVEMBER 2003

former connection types (both grounded and ungrounded connections), lines, switches, and ZIP loads. The convergence of
these power flow solvers is addressed analytically in [11], with
DGs modeled as PQ buses, and experimentally in [12] which
allow for PV buses.
We note that the ability for DGs to feed unbalanced loads
may be limited to their service connection devices. In addition
to three-phase transformer models available in the literature, a
model for a three-phase power converter for unbalanced radial
systems [13] is available.

IV. SOLUTION ALGORITHM

The proposed graph-based solution algorithms for switch
placement are based on the following assumptions:

the system has radial structure;

prefault system information is known;
fault and DG locations are known;
faults have been isolated.

Given the complexity of this problem and the importance of

computational speed, heuristic, graph-based approaches are employed. They leverage switch indices developed in [6] which
yielded high quality results using a fixed number of switches
for network reconfiguration.
First, common definitions in these two algorithms are
introduced:
Controllable load: The part of a load under direct load control (DLC);
Uncontrollable load: The part of a load energized when the
bus is in service;
Boundary switch: The open switch on the boundary of the
resulting isolated area;
Estimated total load limits

(14)

(15)

is the set of distributed generators and is the

where
percentage of losses on the branches with respect to total
power generation.
These limits are defined by scaling distributed generator ratings
to avoid overloading the DGs. The estimated total load limits
are used as a guide for sizing the area to be serviced by the
DGs and to minimize the number of computationally intensive
three-phase power flow runs.
Note, the limits in (14) and (15) are calculated by subtracting
estimated losses from generator power limits. will be updated
and the estimated total load limits will be recalculated after each

Fig. 2. Flowchart of the proposed solution algorithms.

power flow run. Then, the limits guide the switch placement and
load control schemes.
A common procedure is shared by the two algorithms due to
the similarities of their formulations. For each area isolated from
the substation, seven steps may be taken to place switches.
Step 1) Find the separated area formed by fault isolation.
The open tie switches connected to this area are not
to be included in the area.
Step 2) Turn off all nonpriority controllable loads and open
all the closed switches in the area.
Step 3) Build network , which includes the buses with DGs
and the priority loads in the area, and the buses and
branches between them. All of the open network
switches inside will be closed.
Step 4) Check whether total load is within the estimated total
load limits.
a) if the upper limits are not satisfied, shed loads;
b) otherwise, add loads.
Step 5) Run power flow to check whether there is a constraint violation in the resulting isolated area.
Step 6) If there is a constraint violation, start the constraint
handling process and then go to Step 5. If there is no
violation, go to Step 7.
Step 7) Output results.
A general outline for the two algorithms is shown in Fig. 2.
Details of each step are outlined in the following subsections
where two closely related algorithms emerge to address our previous problem formulations.

MAO AND MIU: SWITCH PLACEMENT TO IMPROVE SYSTEM RELIABILITY FOR RADIAL DISTRIBUTION SYSTEMS

A. Maximizing Priority Loads (Algorithm 1)

1) Adding Loads: Uncontrollable loads will be added into
the system first by using existing network switches. The process
has four steps.
Step 1) Build the boundary switch list of .
Step 2) Try to expand by closing these boundary switches
one by one without exceeding the upper limits
of estimated total load power limits. A boundary
switch index [6] was built so that the boundary
switch with the largest downstream real power loads
will be closed first.
Step 3) If after one open switch is closed, the total load is still
within the estimated total load power limits, update
and the boundary switch list. If the upper limits of
the power limits cannot be satisfied, the switch will
not be closed and will be removed from the list.
Step 4) If the boundary switch list is not empty, go to Step
2. If it is empty, stop adding uncontrollable loads.
Controllable loads will be added if there is still an estimated
margin for the upper limit of DG capacities (14), (15). The load
with the highest priority and the largest amount of controllable,
three-phase real power within the upper power limits will be
turned on first. The adding process will end when all controllable loads have been attempted.
If lower limits are violated, new switches must be added into
the island. The details of adding new switches are found in
IV-A-3. After adding new switches, we will reattempt to add
controllable loads to maximize the total load serviced in the resulting isolated area.
2) Shedding Loads: If there are controllable loads in ,
they will be shed first. The load with the least priority and the
smallest amount of controllable, three-phase real power will be
shed first. The process will stop when the total load is within
the estimated total load limits.
If after all the controllable loads have been shed, the total load
is still larger than the estimated upper load limit in (14) and/or
(15), we will try to shed uncontrollable loads from the system.
The procedure is as follows.
Step 1) Estimate the real power load limits (14).
Step 2) Find all the closed switches of and order them according to their downstream loads. Open the fewest
number of switches where the remaining area would
be within the limits. In this step, no priority load will
be shed.
Step 3) If a candidate solution cannot be found in Step 2,
add new switches (IV-A-3) and then go to Step 4. If
a candidate is found in Step 2, go to Step 6.
Step 4) Store the result from Step 3 as result1. Reclose any
switches opened in Step 2 and Step 3. Search for a
candidate where priority loads can be shed.
Step 5) Open switches to shed the least amount of priority
load and meet the guidelines of (14) and (15). If a
candidate is found without adding a new switch,
go to Step 6. If a candidate is found with new
switches added, compare the result with result1 and
identify the better solution according to the order of
the objectives from (1)(4).

1349

Step 6) Update according to switch status. Return to main

algorithm.
3) Adding New Switches: If adding loads or shedding loads
using existing switches fails to find a feasible solution, new
switches will be placed to form the isolated area. These new network switches will be opened to form the area serviced by the
DGs. Note, if adding loads (IV-A-1) fails, the boundary switch
with the largest real power downstream loads will be closed
to form and then new switches are added. If shedding loads
(IV-A-2) fails, we will directly go to add new switches.
The process of adding new switches is similar to shedding
uncontrollable loads. The difference between them is the search
space. Here, the search space will be all the branches of instead of all closed switches. The procedure is as follows:
Step 1) Estimate the real power load limits (14).
Step 2) Add the fewest number of switches, whose operation would release downstream loads to insure the
resulting area would be within the limits. No priority load will be shed in this step. Store the result
as result1.
Step 3) If more than one new switch is added in result1, redo
Step 2 but this time priority loads are allowed to be
shed. Store the result as result2.
Step 4) Compare result1 with result2. The one with fewer
new switches will be chosen as the candidate solution. If they have the same number of new switches,
result1 will be chosen as the candidate.
4) Constraint Handling: If power limits of distributed generators are violated, then shedding process/adding process will
be used to remove the violations. If voltage violations exist, then
the following two techniques will be used to remove voltage violations:
use voltage regulation on DG;
add/shed loads.
In this paper, the voltage regulation will be tried first. If that
alone cannot eliminate voltage violation, adding or shedding
load processes will be invoked.
If branch constraint violations occur, the shedding load
process will start in the area that includes all the loads downstream of overloaded branches. The process will stop when
the sum of the load currents in the area is less than the upper
limit of the overloaded branch. Finally, the adding process will
be used where its search space is limited to the area without
overloaded branches and their downstream branches.
B. Servicing All Priority Loads (Algorithm 2)
For our second algorithm, servicing all the priority loads becomes a constraint. Due to the similarities between the two problems, the resulting solution algorithms have a lot in common.
Therefore, instead of stating Algorithm 2 in detail, only the differences between them will be highlighted.
For the adding process, the only difference between Algorithm 1 (IV-A) and Algorithm 2 is in adding new switches. No
branches with priority loads downstream will be considered. For
the shedding process, the process will stop at Step 4 so that all
priority loads will be included.

1350

Fig. 3.

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO. 4, NOVEMBER 2003

One-line diagram of the 394-bus test system.

The above solution algorithms integrated with a three-phase

power flow solver were coded in Matlab. the test results are
stated in the next section.

TABLE I
BASIC INFORMATION OF TEST CASES

V. NUMERICAL RESULTS
A 394-bus radial distribution system is used in all simulation tests in this section. The amount of system load is 25 260
kW and 5746.7 kVar. Detailed information about the number of
components follows:

A one line diagram of the system is shown in Fig. 3. When

faults are on the lines with switches, they are isolated by those
switches. Faults on lines without switches are isolated by
opening proper downstream and upstream switches.
Cases with different fault locations, different number of DGs,
and different number and locations of priority loads have been
tested. In order to show the effectiveness of the proposed algorithms, the fault location remained fixed; while the number of
DGs and priority loads will be varied.
The basic information of the system for all five cases to be
presented is found in Table I. The results of the five cases are
listed in Tables IIVI.

TABLE II
CASE 1 RESULTS WITH DG AT BUS 59 AND 67% CONTROLLABLE LOAD

We assume all DGs have the same percent of adjustable

output . All nonpriority loads have the same percentage
of controllable load. Priority loads are considered to be uncontrollable. It should be mentioned that the percentage of

MAO AND MIU: SWITCH PLACEMENT TO IMPROVE SYSTEM RELIABILITY FOR RADIAL DISTRIBUTION SYSTEMS

1351

TABLE III
CASE 2 RESULTS WITH DG AT BUS 59 AND NO CONTROLLABLE LOAD

TABLE V
CASE 4 RESULTS WITH DG AT BUSES 70, 92, AND NO CONTROLLABLE LOAD

TABLE IV
CASE 3 RESULTS WITH DG AT BUSES 70, 92, AND 67% CONTROLLABLE LOAD

TABLE VI
CASE 5 RESULTS WITH DG AT BUSES 70, 92, AND NO CONTROLLABLE LOAD
AND FIVE PRIORITY LOADS

uncontrollable load can be set in the range from 0 to 1, and can

be set individually to each nonpriority load. Load is represented
in per unit. The per-phase power base is 33 333 kVA.
A. Single DG at Bus 59
The DG at bus 59 had the capacity of 4000 kVA and had
3500-kW real power generation before fault. Two cases are
studied where the amount of controllable load is varied.
Case 1: (Nonpriority loads 67% controllable);
Case 2: (Nonpriority loads 100% uncontrollable).
Remarks:
In Case 1, both algorithms end with the resulting island
without losing any node. In Case 2, both algorithms need
to install a new normally closed switch which must open to
form the resulting island and end with the same resulting
island.
Note that both Algorithm 1 and Algorithm 2 yield the same
results in both cases. This is because, the solution of Algorithm 1 included all priority loads in the area isolated by
fault, which satisfied the constraints of Algorithm 2.
The difference in total load serviced between Case 1 and 2
suggests that more controllable loads may result in a larger
area supported by distributed generators and may avoid the
installation of a new switch.
As such, different case studies can be devised whose results provide information on whether implementing direct
load control (DLC), adding new network switches or both

should be chosen to increase the amount of loads serviced

by DGs after fault isolation.
B. Multiple DGs at Bus 70 and Bus 92
With the same fault location, two DGs at bus 70 and 92 are assigned. They all had the capacity of 2000 kVA and had 1750-kW
real power generation before fault.
Case 3: (Nonpriority loads 67% controllable);
Case 4: (Nonpriority loads 100% uncontrollable).
Remarks:
The proposed algorithms can handle the cases with multiple DGs inside the area isolated by fault.
The results of Case 3 and 4 with two DGs show the same
trend as in Case 1 and 2. If the Algorithm 1 can find a
solution without losing any priority load, Algorithm 2 will
have the same result as Algorithm 1.
Again, the results of Case 3 and 4 are consistent with the
results of Case 1 and 2. They also demonstrate the benefits
of DLC when DGs are employed.
C. Two DGs With Five Priority Loads
In the last case, the two DGs are sized and located as in Case
3 and 4. However, five different priority loads were assigned.
Case 5: (Nonpriority loads 100% uncontrollable)
Remarks:
Since Algorithm 1 has the freedom to shed priority loads,
no new switch is added; however, the DG cannot service
all priority loads.

1352

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 18, NO. 4, NOVEMBER 2003

Since Algorithm 2 is constrained to service all priority

loads, a new switch must be added in order to service them.
In this case, more total load was also served.
These solutions provide two options to meet different
needs. You can choose to install new switches or lose
some priority loads.
General comments are now noted. First, the results of these
two algorithms can also lead to the application of this work
to the sizing and location of DGs. Second, it is acknowledged
that the implementation of islanding systems would prove
challenging for large-scale terrestrial distribution systems in
the near term. However, the proposed work could assist in
rigorously evaluating the tangible and intangible benefits of
islanding to justify potential costs. Finally, we note that the cost
and construction of small isolated power distribution systems
and the required communication systems which would allow
for islanding, such as naval shipboard systems, is realizable in
the near future.
VI. CONCLUSION
In this paper, the switch placement problem to improve
system reliability for radial distribution systems with DGs after
a fault was formulated. A multiobjective optimization problem
was presented. In order to improve reliability, the problem
included maximizing the amount of load to be continuously
supported by the DG in isolation from the substation. In order
to consider costs, the problem included minimizing the number
of new switches to be placed. Customer priority and constraints
are also considered in the formulation.
Graph-based solution algorithms were designed to evaluate
new switches, operate existing switches, and invoke direct load
control in order to form the service areas for the DGs. The algorithms are flexible and accommodate changes in the treatment
of priority loads. The proposed algorithms can be used in planning and/or online application problems for radial distribution
systems with DG.
Extensive testing on an unbalanced 394-bus radial distribution system was performed. Highlights in this paper include test
cases with different numbers and sizes of DG, different priority
loads, and different amounts of direct load control. The simulation results demonstrate the effectiveness of the proposed algorithms to service priority customers and to avoid unnecessary
new switch placements.
Several potential applications of this work are noted. First,
tradeoff studies for implementing direct load control versus new
switch expenditures can be studied. Also, a potential exists to
apply the methods toward the sizing and location of DG. Finally, the proposed work can be adapted to assist in rigorous
cost/benefit analysis of islanding in distribution systems.

REFERENCES
[1] P. P. Barker and R. W. De Mello, Determining the impact of distributed
generation on power systems. I. Radial distribution systems, in
IEEE/Power Eng. Soc. Summer Meeting, vol. 3, 2000, pp. 16451656.
[2] R. E. Brown and L. A. A. Freeman, Analyzing the reliability impact
of distributed generation, in Power Eng. Soc. Summer Meeting, vol. 2,
2001, pp. 10131018.
[3] L. J. Powell, An industrial view of utility cogeneration protection requirements, IEEE Trans. Ind. Applicat., pt. 1, vol. 24, no. 1, pp. 7581,
Jan./Feb. 1988.
[4] R. Billinton and S. Jonnavithula, Optimal switching device placement
in radial distribution systems, IEEE Trans. Power Delivery, vol. 11, pp.
16461651, July 1996.
[5] M. A. Redfern, J. Barrett, and O. Usta, A new microprocessor based islanding protection algorithm for dispersed storage and generation units,
IEEE Trans. Power Delivery, vol. 10, pp. 12491254, July 1995.
[6] K. N. Miu, H.-D. Chiang, B.Bentao Yuan, and G. Darling, Fast service
restoration for large-scale distribution systems with priority customers
and constraints, IEEE Trans. Power Syst., vol. 13, pp. 789795, Aug.
1998.
[7] A. Molina, A. Gabaldon, J. A. Fuentes, and F. J. Canovas, Approach to
multivariable predictive control applications in residential hVAC direct
load control, in IEEE Power Eng. Soc. Summer Meeting, vol. 3, 2000,
pp. 18111816.
[8] K.-H. Ng and G. B. Sheble, Direct load control-a profit-based load
management using linear programming, IEEE Trans. Power Syst., vol.
13, pp. 688694, May 1998.
[9] D. Novosel, K. T.Khoi T. Vu, D. Hart, and E. Udren, Practical protection and control strategies during large power-system disturbances, in
Proc. IEEE Transm. Dist. Conf., 1996, pp. 560565.
[10] R. Billinton, Distribution system reliability performance and evaluation, Elect. Power Energy Syst., vol. 10, no. 3, pp. 190200, July 1988.
[11] K. N. Miu and H.-D. Chiang, Existence, uniqueness, and monotonic
properties of the feasible power flow solution for radial three-phase distribution networks, IEEE Trans. Circuits Syst. I, vol. 47, no. 10, pp.
15021514, Oct. 2000.
[12] R. D. Zimmerman, Comprehensive Distribution Power Flow: Modeling, Formulation, Solution Algorithm and Analysis, Ph.D. dissertation, Cornell University, Ithaca, NY, Jan. 1995.
[13] R. Stoicescu, K. Miu, C. O. Nwankpa, D. Niebur, and X.Xiaoguang
Yang, Three-phase converter models for unbalanced radial power-flow
studies, IEEE Trans. Power Syst., vol. 17, pp. 10161021, Nov. 2002.

Yiming Mao (S99) received the B.S. and M.S. degrees in electrical engineering
from Southeast University, Nanjing, China. He is currently pursuing the Ph.D.
degree at Drexel University, Philadelphia, PA.
His areas of interest include power distribution system analysis and distribution protection systems.

Karen N. Miu (M98) received the B.S., M.S., and Ph.D. degrees in electrical
engineering from Cornell University, Ithaca, NY, in 1992, 1995, and 1998, respectively.
Currently, she is an Assistant Professor in the Electrical and Computer Engineering Department at Drexel University, Philadelphia, PA. Her research interests include distribution system analysis, distribution automation, and optimization techniques applied to power systems.
Dr. Miu received the 2000 National Science Foundation (NSF) Career Award
and the 2001 Office of Naval Research (ONR) Young Investigator Award.