Electrical Power and Energy Systems 33 (2011) 14571461

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

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

Optimal power system operation using parallel processing system

and PSO algorithm
Jong-Yul Kim a,, Kyeong-Jun Mun b, Hyung-Su Kim c, June Ho Park a
a

Pusan National University, Geumjeong-gu, Busan, South Korea

Korea Atomic Energy Research Institute, Daejeon, South Korea
c
Namhae College, Gyeongnam, South Korea
b

a r t i c l e

i n f o

Article history:
Received 2 December 2008
Received in revised form 15 February 2011
Accepted 1 June 2011
Available online 2 July 2011
Keywords:
Particle swarm optimization
Heuristic
Optimal power ow
PC cluster system
Parallel processing

a b s t r a c t
In recent studies, PSO algorithm is applied to solve OPF problem. However, population based optimization method requires higher computing time to nd optimal point. This shortcoming is overcome by a
straightforward parallelization of PSO algorithm. The developed parallel PSO algorithm is implemented
on a PC-cluster system with 8 Intel Pentium IV 2 GHz processors. The proposed approach has been tested
on the test systems. The results showed that computing time of parallelized PSO algorithm can be
reduced by parallel processing without losing the quality of solution.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
Optimal Power Flow (OPF) is a useful tool in planning and operation of a power system. The OPF problem can be described as the
optimal allocation of power system controls to satisfy the specic
objective function such as fuel cost, power loss, and bus voltage
deviation. The control variables include the generator real powers,
the generator bus voltages, the tap ratios of transformer and the
reactive power generations of VAR sources. Therefore, the OPF
problem is a large-scale highly constrained nonlinear non-convex
optimization problem [1]. Recently, many heuristic optimization
methods in [24] to overcome the limitations of the mathematical
programming approaches have been investigated. Particle Swarm
Optimization (PSO) is a newly proposed population based heuristic
optimization algorithm [5]. Compared with other heuristic optimization methods, PSO has comparable or even superior search performance for some hard optimization problems in real power
systems [68]. However, population based optimal research methods such as GA, EP and PSO require relatively higher computing
time than conventional optimization techniques. In parallel processing, problems are divided into several sub problems, and allocated to each processor. This can reduce computing time and
enhance computation efciency [9]. In this paper, parallel PSO
algorithm is proposed to improve the computing time and also
Corresponding author. Tel.: +82 55 280 1336; fax: +82 55 280 1339.
E-mail address: jykim@keri.re.kr (J.-Y. Kim).
0142-0615/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2011.06.026

PC-cluster system is developed to implement parallel PSO algorithm. To verify the usefulness of the proposed algorithm, parallel
PSO algorithm has been tested and compared with standard PSO
algorithm having with single processor. The standard IEEE 30 and
118-bus power systems have been employed to carry out the simulation study.
2. Optimal power ow problem formulation
The OPF problem can be formulated as a constrained optimization problem as follows:

Minimize f x; u
subject to gx; u 0
hx; u 6 0

1
2
3

where x is a set of state variables, and u is a set of controllable

variables.
In this paper, the objective function of OPF is minimization of
fuel cost for all generators which can be formulated as follows:

Min f Pgi

Ng 

X
ai bi P gi ci P2gi

i1

where f(Pgi) is the total fuel cost ($/h) of all generators; Pgi is the
active power output generated by the ith generator; ai, bi, ci are fuel
cost coefcients; and Ng is the total number of generators. The

1458

J.-Y. Kim et al. / Electrical Power and Energy Systems 33 (2011) 14571461

equality constraints g(x, u) are the nonlinear power ow equations

which are formulated as follows:

Pgi  Pdi  V i

Nb
X

V j Y ij coshi  hj  uij 0 i 1; . . . ; Ng

j1

Q gi  Q di  V i

Nb
X

V j Y ij sinhi  hj  uij 0 i 1; . . . ; Ng

j1

where Pgi and Qgi are the active and reactive power generations at
bus i; Pdi and Qdi are the active and reactive power demands at
bus i; Vi and Vj are the voltage magnitude at bus i and j respectively;
hi and hj are the voltage angles at buses i and j respectively; uij is the
admittance angle; Yij is the admittance magnitude; and Nb is the total number of buses.
The OPF inequality constraints, h(x, u), represent limits of control variables and state variables. The system operation constraints
consist of the transmission line loadings, load bus voltages, reactive power generations of generator, and active power generation
of slack generator. These variables should be within the set lower
and upper limits.

Si 6 Si 6 Smax
i

i 1; 2; . . . ; Nl

V min
6 V i 6 V max
i
i

i 1; 2; . . . ; Nb

max
Q min
gi 6 Q gi 6 Q gi

i 1; 2; . . . ; Ng

max
Pmin
gs 6 P gs 6 P gs

7
8
9
10

Concerning control variables, active power output and voltage

of generators, transformers tap ratio, and shunt capacitors are restricted by lower and upper limits as follows:

Pmin
gi

6 Pgi 6

Pmax
gi

i 1; 2; . . . ; Ng  1

max
V min
gi 6 V gi 6 V gi

t min
6 t i 6 t max
i
i
min
shi

6 shi 6

i 1; 2; . . . ; Ng

i 1; 2; . . . ; Nt

max
shi

i 1; 2; . . . ; Nsh

11
12
13
14

3. Parallel computation of PSO algorithm using PC clustering

3.1. PC cluster system
After mid 1980, high performance computers have been needed
according to the development of large scale science and engineering. Since supercomputers are expensive, cluster systems replaced
supercomputers because it has the availability of inexpensive high
performance PCs, and high speed networks, and development of
integrated circuits. PC cluster system provides higher availability
as well as greater performance by lower cost with interconnecting
several PCs or workstations. PC cluster system is very competitive
with parallel machine in terms of a ratio of cost to performance because clustering is one of the types of parallel or distributed processing system, which is composed of a collection of
interconnected low cost PCs working together as single and integrated computing resources. Also, it is easy to add nodes that construct the PC cluster. A basic construction diagram for PC cluster is
shown in Fig. 1.
The performance of the PC cluster system depends on the quality of message passing system, libraries, and compilers for parallel
programming and performance of individual nodes. Therefore, it is

important to select each component described above properly to

obtain better performance. The PC cluster system implemented
in this paper is composed of eight nodes based on fast Ethernet
with Ethernet switch. For operating system, master node uses
Windows 2000 server, and slave nodes use Windows 2000 pro.
To connect each node, fast Ethernet card and switching hub were
used. In data communication, MPI library was used, which is effective for parallel application by using message-passing method
through TCP/IP over Internet. Symantec PC anywhere was used
for remote control of each node, and MS visual C++ 6.0 was used
for compilers of parallel programming. Table 1 shows the picture
and the specication of the PC cluster system developed in this
paper.
3.2. Parallel computing of PSO algorithm
The PSO is basically developed through the simulation of bird
ocking in two-dimensional space. In PSO, each particle i
(i = 1, . . . , N) in the population is characterized by three vectors
(xi, vi, pi) which represent their temporal position, velocity, and
the best position. The tness of each particle is given by the function value f(xi). Each particle stores its best position pi called personal best, p-best, which gives the best tness in memory. They
can also consult their neighbors best position. Most simply, the
neighbor is the whole population (fully connected topology), and
therefore, the neighbors best is the best position among personal
bests of the whole population. Hence, the position pg is called global best. Now each particle i moves around the search space, and
renews its velocity using its past experience (personal best) and
the populations experience (global best) as follows:

v i xv i c1 r1 pi  xi c2 r2 pg  xi

15

The parameter c1 and c2 are the acceleration constant, r1 and r2

are the uniform random numbers within the range [0, 1]. If vi is larger than a predened velocity vmax called maximum velocity, it is
set to vmax. Similarly, if it is smaller than vmax, it is xed to vmax.
The parameter x is called inertia weight [10], which controls the
exploration (global search)exploitation (local search) tradeoff.

x xini  xfin  MAX iteration  Iteration=MAX iteration

xfin

16

Then the particle changes its position by the equation of

motion:

xi xi v i

17

The population size is one of the key factors that will affect the
search performance of the PSO algorithm for seeking the optimal
solution. The larger population size can guarantee the higher
chance of obtaining the optimal solution. However, it is obvious
that more computing time is needed. To reduce the computing
time with same quality of solution, parallel PSO algorithm is proposed and paralleled by the PC cluster system. The most important
issue of parallelizing PSO algorithm is exchange model of evolution
information. Different ways will result in different performances.
The proposed conguration is a kind of parallel algorithm based
on coarse grain model, in which the population is divided into
some sub-populations evolving independently.
Each sub-population exchanges require information only between two neighboring sub-populations connected by arrowed
lines as shown in Fig. 2. Each sub-population is allocated in each
processor that involves in parallel computing. With each processor
that can communicate with the neighboring sub-populations, the
best solution of each processor is transferred to the neighboring
processors by migration operation every generation. The owchart

J.-Y. Kim et al. / Electrical Power and Energy Systems 33 (2011) 14571461

1459

Fig. 1. Conguration of parallel PSO algorithm with the ring structure.

Table 1
Specication of PC cluster system.
Item

Specication

CPU
Mother board
Chipset
RAM
HDD
NIC
Network switch
Operating system
MPI library
Compiler

Intel 2.0 GHz

LeoTech P4XFA
VIA P4X266A
DDR SD RAM 256 MB
Samsung 40 GB 5600 rpm
3Com 3CSOHO 100-TX
3Com 3C16465C Switch
Window 2000 Server/Window 2000 Pro
MPICH 1.2.5
Visual C++ 6.0

Fig. 2. Structure of population in parallel PSO algorithm.

Fig. 3. Flow chart of parallel PSO algorithm.

for searching optimal solution using the proposed parallel PSO is

presented in Fig. 3.

3.3. PSO algorithm for the optimal power ow problem

In optimal power ow problem to allocate power system controls optimally, we should determine the unit active power outputs, generator-bus voltage magnitudes, transformer tap ratios,
shunt capacitor capacities. To get the effective solution of optimal
power ow by the proposed parallel PSO, we should design PSO
appropriately for the optimal power ow problem. To solve the
OPF problem by PSO, we select unit active power outputs, generator-bus voltage magnitudes, transformer tap ratios, and shunt
capacitors as control variables in PSO position vector as follows:

S1 Pg1 ; . . . ; Pgn ; V g1 ; . . . ; V gn ; t 1 ; . . . ; t n ; sh1 ; . . . ; shn

S2 Pg1 ; . . . ; Pgn ; V g1 ; . . . ; V gn ; t 1 ; . . . ; t n ; sh1 ; . . . ; shn
..
.
Sp Pg1 ; . . . ; Pgn ; V g1 ; . . . ; V gn ; t 1 ; . . . ; t n ; sh1 ; . . . ; shn

where Pgi is the active power output of the ith generator, Vgi the
voltage magnitude of the ith generator bus, ti the transformer tap
ratios of the ith transformer, shi the no. of bank of the ith shunt
capacitor and p is the no. of position vector.
In the evaluation procedures of PSO, tness value can be obtained by the following equations. As shown in Eq. (18), tness is
composed of fuel cost of the generator and several constraints for
the power system operations.

J.-Y. Kim et al. / Electrical Power and Energy Systems 33 (2011) 14571461

Fitness PNg

i1 F i P gi

PN c

j1

18

xj  Penj

where Fi(Pgi) is the fuel cost function of the ith generator, xj the
coefcient of the jth constraint, Penj the penalty function of the
jth constraint, Ng the no. of generators, Nc the no. of constraints
and a is the constant value.
4. Test results
Two test systems, IEEE 30 and 118-bus systems are used to verify the proposed algorithm with emphasis on the validity of the
calculation results and the effectiveness of the proposed algorithm.
The simulation parameters of PSO algorithm are listed in Table 2.
4.1. IEEE 30-bus system

Table 3
Minimum solution found by Parallel PSO-OPF in IEEE 30-bus system.
Parallel PSO-OPF solution
P1
P2
P5
P8
P11
P13
t412
t69
sh10
sh12
sh15
sh20

177.13
48.82
21.40
21.30
11.82
12.00
1.096
0.908
0.034
0.0302
0.0293
0.0455

V1
V2
V5
V8
V11
V13
t610
t2827
sh21
sh23
sh24

1.08
1.06
1.03
1.03
1.07
1.05
0.984
0.982
0.0484
0.0370
0.05

Table 4
Summary of searching performance in IEEE 30-bus system.

The IEEE 30-bus system has a total of 24 control variables as follows: ve unit active power outputs, six generator-bus voltage
magnitudes, four transformer-tap settings, and nine shunt capacitors. Transformers are in-phase transformers with assumed tapping range of 0.91.1 pu and shunt capacitors are in the range of
00.05 pu. The lower voltage magnitude limits at all buses are
0.95 pu, and the upper limits are 1.1 pu for generator buses and
1.05 pu for the remaining buses including the reference bus.
The best cost of PSO algorithm with single processor gives
800.68 $/h, and proposed parallel PSO algorithm with eight processors gives 800.64 $/h. Both of algorithms show almost same quality
solution which is less than the 804.8 $/h reported in [11]. OPF solution found by parallel PSO and corresponding control variables setting are described in Table 3. More detail searching performance of
parallel PSO is presented in Table 4. The computing times for standard and parallel PSO algorithm with eight processors are 8.03 s
and 1.73 s respectively.
To show the effects of the parallel operation by the PC clustering, speedup is evaluated. Speedup is described below:

Method

Processor number

Cost ($/h)

Computation time (s)

Ref. [10]
PSO
Parallel PSO

1
1
8

804.8
800.68
800.64

8.03
1.73

5
4

Speed up

1460

3
2
1
0

No. of processors
Fig. 4. Speed up according to the processor number in IEEE 30-bus system.

 speedup (Sp)

Sp

T
Tp

19

where T is run time on one processor and Tp is run time on p

processors.
Fig. 4 shows the speedup as the number of processors increases.
From Fig. 4, it is found that speedup increased as the number of
processors increased almost linearly, but somewhat lowered
because there exists overhead when communication executed
between processors.
4.2. IEEE 118-bus system

using single processor is 17560.4 $/h, and proposed parallel PSO

algorithm with 8 processors also gives the similar result of
17554.0 $/h. These results are less than the 17860.09 $/h reported
in [12]. Minimum solution found by PSO algorithm in 118 bus system and summary of searching performance are presented in Tables 5 and 6. The computing time is obviously reduced from
1362.26 s to 169.0 s by the parallel PSO algorithm with 8 processors. Fig. 5 shows the speed up as the number of processors
increases. From Fig. 5, it is found that speedup increased as the
number of processors increased almost linearly, but somewhat

Table 5
Minimum solution found by PSO-OPF in IEEE 118-bus system.

The IEEE 118-bus system has 118-bus, 14-generator,

9-transformer, 179-branch. It also has a total of 27 control
variables. The lower voltage magnitude limits at all buses are same
with IEEE 30-bus system. The best cost of standard PSO algorithm

Table 2
Simulation parameters.
Parameter

Value

Max iteration
Population
C1
C2
w

50
60
2.0
2.0
0.90.4

Parallel PSO-OPF solution

P1
P10
P12
P25
P26
P49
P59
P61
P65
P66
P80
P89
P100
P103

290.00
328.88
210.00
241.60
241.61
237.97
195.00
210.01
345.77
315.00
336.16
315.00
230.34
265.00

V1
V10
V12
V25
V26
V49
V59
V61
V65
V66
V80
V89
V100
V103

1.09
1.09
1.07
1.10
1.10
1.07
1.06
1.08
1.01
1.10
1.08
1.10
1.10
1.09

J.-Y. Kim et al. / Electrical Power and Energy Systems 33 (2011) 14571461
Table 6
Summary of searching performance in IEEE 118-bus system.
Method

Processor number

Cost ($/h)

Computation time (s)

Ref. [11]
PSO
Parallel PSO

1
1
8

17860.09
17560.4
17554.0

1362.26
169.0

Speed up

developed. To verify the performance of the proposed method, parallel PSO algorithm is tested on an IEEE 30 and 118-bus systems.
For repeated evaluating tness function during evolution process,
it needs lots of computing cost by calculating load ow. Therefore,
proposed parallel PSO algorithm can divide the population into
several sub populations to share the burden of calculating the load
ow. As a result, computing time of parallel PSO algorithm can be
further improved.
References

10

6
4
2
0

1461

No. of processors
Fig. 5. Speed up according to the processor number in IEEE 118-bus system.

lowered because there exists overhead when communication executed between processors.
5. Conclusions
In these days, many heuristic optimization methods such as GA,
EP and PSO are developed and applied to OPF problem. However,
heuristic optimization methods require relatively higher computing time which is one of the major obstacles on dealing with the
on-line OPF. In this paper, parallel PSO algorithm based on PCcluster system is proposed and applied to the OPF problem. For
parallel computing, a PC cluster system consisting of 8 PCs is also

[1] Momoh JA, Zhu JZ. Improved interior point method for OPF problems. IEEE
Trans Power Syst 1999;14(3):111420.
[2] Kumari M Sailaja, Maheswarapu Sydulu. Enhanced genetic algorithm based
computation technique for multi-objective optimal power ow solution. Electr
Power Energy Syst 2010;32(6):73642.
[3] Ketabi Abbas, Alibabaee Ahmad, Feuillet R. Application of the ant colony search
algorithm to reactive power pricing in an open electricity market. Electr Power
Energy Syst 2010;32(6):6228.
[4] Bakirtzis AG, Biskas PN, Zoumas CE, Pehdis V. Optimal power ow by enhanced
genetic algorithm. IEEE Trans Power Syst 2002;17(2):22936.
[5] Kennedy J, Eberhart RC. Particle swarm optimization. In: Proceedings of IEEE
international conference on neural network; 1995.
[6] Arya LD, Titare LS, Kothari DP. Improved particle swarm optimization applied
to reactive power reserve maximization. Electr Power Energy Syst
2010;32(5):36874.
[7] Yang B, Chen Y, Zhao Z, Han Q. Solving optimal power ow problems with
improved particle swarm optimization. In: Proceedings of the 6th world
congress on intelligent control and automation, Dalian; 2006.
[8] Lu Haiyan, Sriyanyong Pichet, Song Yong Hua, Dillon Tharam. Experimental
study of a new hybrid PSO with mutation for economic dispatch with nonsmooth cost function. Electr Power Energy Syst 2010;32(9):92135.
[9] Chung SH, Ryu KR, OS C, Park TW. Parallel processing system for high speed
information retrieval. Parallel Process Syst Newslett 1996;7(2):319.
[10] Iwamatsu M. Locating all the global minimum using multi-species particle
swarm optimizer: the inertia weight and the constriction factor variants. In:
Proceedings of the evolutionary computation conference. Vancouver; 2006.
[11] Lee KY, Park YM, Ortiz JL. A united approach to optimal real and reactive power
dispatch. IEEE Trans Power Appart Syst 1985;104(5):114753.
[12] Todorovski M, Rajicic D. An initialization procedure in solving optimal power
ow by genetic algorithm. IEEE Trans Power Syst 2006;21(2):4807.