See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/224254542

A novel optimization algorithm for optimal

reactive power dispatch: A comparative study

Conference Paper August 2011

DOI: 10.1109/DRPT.2011.5994144 Source: IEEE Xplore

CITATIONS READS

22 98

3 authors:

Worawat Nakawiro I. Erlich

King Mongkut's Institute of Technology Lad University of Duisburg-Essen
18 PUBLICATIONS 137 CITATIONS 150 PUBLICATIONS 2,605 CITATIONS

SEE PROFILE SEE PROFILE

Jos Luis Rueda

Delft University of Technology
78 PUBLICATIONS 300 CITATIONS

SEE PROFILE

Available from: Jos Luis Rueda

Retrieved on: 16 May 2016
 A Novel Optimization Algorithm for Optimal
Reactive Power Dispatch: A Comparative Study
Worawat Nakawiro
Institute of Electric Power Systems (EAN)
University of Duisburg-Essen
Duisburg, Germany
worawat.nakawiro@uni-due.de
AbstractOptimal reactive power dispatch (ORPD) is a key
instrument to achieve secure and economic operation of power
systems. Due to complex characteristics of ORPD, heuristic
optimization has become an effective solver. A novel heuristic
optimization algorithm namely the Mean-Variance Mapping
Optimization (MVMO) is proposed to handle the ORPD
problem. The concept, mechanism and implementation of
MVMO are discussed in this paper. Based on the IEEE 57- and
118- bus systems, MVMO is compared with some basic and
enhanced evolutionary algorithms. Simulation results show that
MVMO is an excellent algorithm for ORPD and should deserve
more attention. The superiority to the other algorithms is very
pronounced in the IEEE 118-test case.
Keywords-component; Reactive power dispatch, Mean-Variance
Mapping Optimization, Heuristic optimization
I. INTRODUCTION
The optimal reactive power dispatch (ORPD) plays an
important role in optimal operation of electric power systems.
It is a complex nonlinear optimization problem with a mixture
of continuous and discrete control variables. ORPD is a sub-
class of the optimal power flow (OPF) problem [1]. Generally,
the control variables of ORPD consist of transformer tap
positions, generator set points (either reactive power injection
or voltage), and reactive power compensations. The objective
of ORPD is to minimize transmission losses or other
concerned objective, such as voltage deviation from the
desired level, etc. Furthermore, physical and operation
constraints must be maintained within the allowable limits.
Several classical gradient-based optimization algorithms
have been applied to solve different ORPD problems such as
[2-4]. These techniques are computationally fast. However,
they have difficulties in handling problems with the non-
convex or discontinuous landscape and discrete variables.
In the past decade, heuristic optimization algorithms, such
as genetic algorithm (GA) [5, 6], particle swarm optimization
(PSO) [7, 8], differential evolution (DE) [9, 10], evolutionary
programming (EP) [11], bacterial foraging optimization (BFO)
[12], etc have been applied in reactive power optimization.
These techniques have shown effectiveness in overcoming the
disadvantages of classical algorithms. Specially, PSO and DE
have received great attention from researchers because of their
novelty and searching power. However, it does not mean that
these techniques do not have any limitations. In solving
978-1-4577-0365-2/11/$26.00 2011 IEEE 1555
Istvn Erlich and Jos Luis Rueda
Institute of Electric Power Systems (EAN)
University of Duisburg-Essen
Duisburg, Germany
istvan.erlich@uni-due.de
complex multimodal problems, these methods may be easily
trapped into a local optimum. Furthermore, their searching
performance depends on the appropriate parameter settings
[13]. Premature convergence and local stagnation are
frequently observed in many applications [14].
Mean-variance mapping optimization (MVMO) originally
proposed in [15] is a novel optimization algorithm. The basic
concept of MVMO shares certain similarities to other heuristic
approaches. However, the novel contribution in MVMO is the
special mapping function. The shape and location of the
mapping curve are adjusted according to the searching
process. The output of this mapping function is always inside
[0,1]. This means that MVMO can guarantee no violation of
the variable limits during the searching process. Moreover,
MVOM updates the candidate solution around the best
solution in every iteration. Thanks to the well-designed
balance between search diversification and intensification,
MVMO can find the optimum quickly with minimum risk to
premature convergence. In power system optimization,
MVMO was successfully applied to solve the optimal reactive
power dispatch of wind farms in [16]. In this paper, the
MVMO algorithm is used to solve the ORPD problem on two
standard test systems.
The remaining of this paper is organized as follows. Section
II presents a mathematical formulation of ORPD. The MVMO
algorithm for solving ORPD is described in details in section
III. In section IV, simulation results of MVMO are compared
with other heuristic methods on the IEEE 57- and 118- bus
systems. Finally, conclusion and future research direction are
drawn in section V.
II. PROBLEM FORMULATION
The goal of ORPD is to determine the optimal settings of
reactive power control variables leading to the minimum
power losses while maintaining several constraints. The
problem can be mathematically described as follows.
Ploss = f ( x, u, d)
Minimize (1)
(
= Gk ( i , j ) U i2 + U 2j 2U iU j cos ij )
k v E
Subject to:
a) Generator bus voltage

UGimin UGi UGimax i = 1, 2,..., nG (2)
b) Susceptance of shunt compensator
Bimin Bi Bimax i = 1, 2,..., nB (3)
c) Transformer tap setting
tapimin tapi tapimax i = 1, 2,..., nT (4)
d) Load bus voltage
ULimin ULi U Limax i = 1, 2,..., nPQ (5)
e) Line flow
SFi SFimax i = 1, 2,..., nL (6)
where Ploss is the total active power losses of the transmission
network. The vector of control variables is u=[UG B tap]Tand
(2)-(4) are the corresponding limits. nG , nB and nT denote the
number of generators, shunt compensators and transformers,
respectively. The vector of state (dependent) variables is
x=[UL SF]T and (5)-(6) are the corresponding limits.
nPQ and nL denote the number of load (PQ) buses and
transmission lines, respectively. The vector d=[PD QD]T
contains active and reactive power demand at all load buses.
In every update of the control vector u, the state vector x;
the load bus voltage both magnitude and angle are computed
by the power flow equations given by:
N
0 = PGi PDi U i U j Yij cos( i j ij ) (7)
j =1
N assigned. For a minimization problem, an individual is better
0 = QGi QDi U i U j Yij sin( i j ij ) (8) if the fitness is smaller. The static penalty scheme is used in
j =1
where PGi and QGi are active and reactive power generation at
bus i, respectively; PDi and QDi are active and reactive power
load at bus i, respectively; Yij is the admittance matrix
corresponding to the ith row jth column and ij is the difference
in the voltage angle between the ith and jth buses; N is the total
number of buses.
III. MEAN-VARIANCE MAPPING OPTIMIZATION ALGORITHM
MVMO operates on a single solution rather than a set of
solutions like in many EAs. The internal searching space of all
variables in MVMO is restricted in [0,1]. Hence, the real
min/max boundaries of variables have to be normalized to 0
and 1. During the iteration it is not possible that any
component of the solution vector will violate the
corresponding boundaries. To achieve this goal, a special
mapping function is developed. The inputs of this function are
mean and variance of the best solutions that MVMO has
discovered so far. The elegant property of MVMO is the
ability to search around the local best-so-far solution with a
small chance of being trapped into one of the local optimums.
This feature is contributed to the strategy for handling the
zero-variance. Further details are referred to our previous
work. The procedure of MVMO for solving the ORPD
problem with D variables can be summarized as follows.
Step 1: Read power system data and set MVMO parameters
Step 2: Initialization: Initialize an initial vector x in [0,1] based
on the uniform random distribution.
Step 3: Fitness evaluation: Denormalize x, run the power flow
and evaluate the fitness function.
Step 4: Termination: Check the termination criteria. If yes,
terminate MVMO. Else, continue to step 5.
Step 5: Solution archive: Store x, fitness value, objective value
and feasibility to the archive if x is better than any of
existing solutions.
Step 6: Based on the archived solutions, compute mean xi and
variance vi for each dimension i.
Step 7: Parent assignment: Assign the best archived solution
xbest as the parent.
Step 8: Variable selection: Select m < D dimensions of x.
Step 9: Mutation: Apply the mapping function to the selected m
dimensions.
Step 10: Crossover: Set the remaining D-m dimensions of x to
the values of xbest
Step 11: Go to step 3.
Fig. 1 MVMO implementation procedure for ORPD
A. Fitness evaluation and constraint handling
For each individual, a power flow calculation is performed,
feasibility of the solution is checked and a fitness value f is
this study to handle constraints. Because the control variables
in x are self-restricted, all dependent variables are constrained
by applying the integrated fitness function as follows:
n
min f = f + i max [ 0, gi ]
(9)
i =1
where f is the original objective function; n is the number of
constraints; is the order of the penalty term (usually 1 or 2);
i is the penalty coefficient of the constraint i and gi is the
inequality constraint i represented by:
g i (x, u) 0 (10)
where x is the vector of state variables and u is the vector
control variables. Other constraint handling techniques are
also applicable to MVMO.
B. Termination criteria
The MVMO search process is terminated based on
completion of the pre-specified number of fitness evaluations
(power flow calculations in ORPD) or no improvement in the
best fitness over the last s fitness evaluations. In this paper, the
first criterion was adopted.
1556
C. Solution archive than strategy 1. However, this observation is still neither
The solution archive (SA) serves as the knowledge base of general nor conclusive. The performance is also problem-
the algorithm for guiding the searching direction. The best n dependent.
individuals that MVMO has found so far are saved in the SA.
Here the data structure of SA is slightly modified from the
original one in [15] in order to handle constraints. Two
information; fitness and feasibility of each individual are
additionally stored as shown in Fig. 2.

Solution
# Objective Fitness Feasibility
x1 x2 xi xD
1
2

n
xi --- --- ---
vi --- --- ---
Fig. 2 Data structure of the solution archive
To avoid losing a good solution in constrained
optimization problems, the following rules are set up to
compare the individual generated at each iteration and existing
archived solutions.
Any feasible solution is preferred to any infeasible
Fig. 3 Variable selection strategies (m=3)
solution;
Between two feasible solutions, the one having better 3) Mutation
For each of the m selected dimension, mutation is used to
objective value is preferred;
assign a new value of that variable. Given a uniform random
Between two infeasible solutions, the one having
number xi [0,1] , the new value of the i component xi is
'
smaller fitness value (smaller constraint violation) is
preferred. determined by:
An update takes place only if the new individual is better
than those in the archive. The archive size is fixed for the entire xi = hx + (1 h1 + h0 ) xi' h0 (11)
process. The archived individuals are dynamically sorted so
that the first ranked individual is always the best. Feasible where hx, h1 and h0 are the outputs of the transformation
solutions are placed in the upper part of the archive. Among mapping function based on different inputs given by:
these solutions, they are sorted based on their original objective hx = h(ui = xi' ), h0 = h(ui = 0), h1 = h(ui =1) (12)
values. Infeasible solutions are sorted according to their fitness
values and then placed on the lower part of the archive. Once Note that the output of (11) is always inside [0,1] for every
the archive is filled up by n feasible solutions, any infeasible generated xi. The mapping function is parameterized as
candidate solution does not have chance to be saved in the follows:
archive.
h( xi , si1 , si 2 , ui ) = xi (1 eui si1 ) + (1 xi ) e(1ui )si 2 (13)
D. Reproduction
Reproduction refers to the process in which an offspring is where si1 and si2 are shape factors allowing asymmetrical
created. In MVMO, it refers to the steps 7-11 of the procedure slopes of the mapping function. Interested readers are referred
shown in Fig. 1. The details of each component are described to our previous work [15] for further details on how to set the
in the following subsections. two different shape factors. The slope is calculated by:
1) Parent assignment si = ln( vi ) fs (14)
The first ranked (best-so-far) solution denoted as xbest is
assigned as the parent. where fs is a MVMO parameter namely the shape scaling
2) Variable selection factor.
MVMO searches around the mean saved in the archive for Fig. 4 shows an example of the mapping function with a
the better solution only in m selected directions. This means uniform shape factor of 10 for two optimization variables x1
that only these m selected dimensions of the offspring will be and x2. The actual range of the two variables is -2.048 x1,x2
updated while the remaining D-m dimensions take the 2.048. Observe that the mean xi is an input of (13). Assume
corresponding values from xbest. Four strategies for selecting that the mean of the two variables is at -1 and 1, respectively.
the variables were implemented in MVMO as shown in Fig. 3. In the normalized scale, these correspond to 0.256 and 0.744.
From our experience, strategies 2 to 4 generally perform better It is clearly shown in the left of Fig. 4 that the curves are flat

1557
around the two mean values. This means that most of the
samples drawn from these functions will center around -1 and
1 in the actual scale. However, notice that the range of these
functions span over the entire range between 0 and 1. This
characteristic gives chances for random samplings to cover the
entire search space by getting away of the mean values. The
right of Fig. 4 shows 10000 samples randomly generated from
the mapping function in the left.
Fig. 4 Mapping function (left) and corresponding two dimensional
searching space (right)
4) Crossover
For the remaining D-m dimensions, the values of xbest are
inherited. This process is equivalent to evolutionary processes
where an offspring is a combination of two parents.
IV. SIMULATION RESULTS
The IEEE 57- and 118- bus systems are used as the test
cases and named as case 57 and case 118, respectively to
examine the performance of MVMO and compare it with
other heuristic algorithms. The system data of the two cases
are taken from [17]. For both test systems, the lower and upper
limits of load bus voltages are 0.95 p.u. and 1.05 p.u.,
respectively. Generator voltages at the high voltage terminal
are defined as continuous variables. The lower and upper
limits are set to 0.94 p.u. and 1.06 p.u., respectively. Discrete
control variables consist of transformer tap positions and the
susceptance of shut compensators. All under-load tap
changing (ULTC) transformers are assumed to have 21
discrete taps within 10% of the nominal voltage (1% for each
tap). Each transformer tap is defined by an integer between -
10 to 10. These ULTC data are fictitious values. The number
of taps and the voltage range in practical cases can be
different. All shunt compensators have 11 discrete steps of
different ratings (defined by an integer between 0 to 10).
For the two test cases, the performance of MVMO is
compared with the following algorithms.
1. PSO: a standard PSO version 2007 [18];
2. DE: a basic DE namely DE/current-to-best/1 [19,
20],
3. JADE: an adaptive DE algorithm [13] and
4. JADE-vPS: a modified JADE algorithm [21].
The difference between the latter two algorithms is the
adaptive scheme for the population size in JADE-vPS. In PSO,
1558
the learning rate and inertia weight are internally specified
according Clercs stagnation analysis. The crossover rate CR
and the scaling factor F of DE are specified by 0.2 and 0.6,
respectively. The parameters CR and F are self-adapted in
JADE whereby the population size PS must be specified by
the user. In JADE-vPS, the parameter PS is dynamically
adjusted during the searching process. The internal parameters
of MVMO are set as follows: fs = 1; AF = 2.5 and sd= 25. The
variable selection is selected to the strategy 3 (see Fig. 3) and
the number of random variable m is chosen to 6 at the
beginning and progressively declined to 2 in every 1000 FEs.
The size of solution archive is fixed to 2.
The termination criterion is the maximum number of
function (power flow) evaluations (FEs). All programs were
implemented in MATLAB R2010a. The computing
environment is Intel Core 2 Quad CPU 2.83 GHz and 3.25 GB
RAM memory. The software package namely PAST is used.
Because the most time-consuming parts in these methods are
the repeated power flow calculations and the number of such
calculations is fixed, the computational time of all algorithms
is not significantly different. The comparison in this paper will
be based on quality of the final results.
A. Simulation results- case 57
The IEEE 57-bus system consists of seven generators, 80
lines where 15 of which are equipped with ULTC
transformers. Shunt reactive power compensators are
connected to buses 18, 25 and 53. The limit of these
susceptances is [0,0.2], [0,0.18] and [0,0.18], respectively.
Therefore, the ORPD search space has 25 dimensions. The
population size PS of PSO, DE and JADE and the initial value
of PS in JADE-vPS is set to 50.
Table I Statistical results case 57
MVMO PSO DE JADE JADE-vPS
Minimum 24.8512 24.8479 24.8360 24.8493 24.8451
Average 24.9917 24.9336 24.8701 24.9494 24.9565
Maximum 25.2608 25.1642 25.0307 25.2044 25.3768
Standard deviation 0.1029 0.0671 0.0352 0.0666 0.0896
36
MVMO
PSO
34 DE
JADE
JADE-vPS
Power losses (MW)
32
30
28
26
24
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
No. of function evaluations 4
x 10
Fig. 5 Average convergence characteristics - case 57
 To fairly compare the proposed MVMO method with
8
others, every algorithm is independently run for 50 times. Tap21-20 Tap34-32 Tap14-46 Tap9-55
Then statistical values consisting of minimum, average, 6
maximum and standard deviation of active power losses are
4
computed as listed in Table I. The average convergence of

Transformer tap position

active power losses found by each algorithm is plotted after 2
the first 100 FEs as shown in Fig. 5. It is clearly shown that
the convergence of MVMO is the fastest. In this test case, the 0

statistical results of MVMO in Table I are not outstanding the -2

other algorithms. However, the MVMO results are on average
very close to the other techniques. An interesting observation -4

made from Fig. 5 is that MVMO is very fast in the global -6

search capability because the lowest power loss has been
found after the first 100 FEs. -8
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
As mentioned in [22] that there are five buses (buses 25, 30, No. of function evalutations x 10
4

31, 32 and 33) in this network that the voltages are outside the Fig. 8 Convergence of selected transformer taps case 57
limits. After the ORPD result given by each method, power The convergence of optimized control variables are shown
flow is calculated to determine bus voltages as shown in Fig. in Fig. 7-Fig. 9. From these figures, the control variables
6. It is shown that all bus voltages can be maintained within change abruptly at the early searching stage. Then, they settle
the limits. These voltage profiles confirm the merits of ORPD to a steady state at the later stage. At this phase, an optimum
in achieving both reduced power losses and voltage security. has been discovered. The CPU time of all methods is
approximately 5 minutes.
1.06
10
1.05 B18 B25 B53
9
1.04
8
1.03
Voltage (p.u.)

Shunt compensator

7
1.02
6
1.01
5
1

MVMO 4
0.99
PSO
DE 3
0.98 JADE
JADE-vPS 2
0.97
0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60
Bus number 1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
No. of function evalutations 4
Fig. 6 Load bus voltage profiles case 57 x 10

Fig. 9 Convergence of shunt compensator case 57

1.06 1.06
B. Simulation results- case 118
1.05 To further evaluate the performance of MVMO, the IEEE
1.05
118-bus system is employed. In this test case, the search space
1.04
of this problem has 77 dimensions. The continuous variables
Generator bus voltage (p.u.)

Generator bus voltage (p.u.)

1.04
1.03 consist of 54 generator bus voltages. There are 23 discrete
variables; 9 transformer taps and 14 shunt compensators. All
1.03 1.02
shunt compensators are defined by susceptances with the
1.01 lower limit of 0 p.u. and the upper limit of 0.2 p.u. The
1.02
population size PS of PSO, DE and JADE and the initial value
U
U
1
U
G6
of PS in JADE-vPS is set to 100.
1.01
U
G1

G2
0.99 U
G8

G9
The convergence of active power losses averaged from 50
U
G3
U
G12
independent trials of different algorithms is shown in Fig. 10.
1
0 0.5 1 1.5 2
0.98
0 0.5 1 1.5 2 Similar to the case 57, the first plotting starts from 100 FEs.
No.of function evaluations
x 10
4 No.of function evaluations
x 10
4
Statistical results are shown in Table II. In this test case,
Fig. 7 Convergence of generator voltages case 57 minimum, average, maximum and standard deviation of power
losses from MVMO are the lowest among all methods. In
other words, MVMO perform the best in this problem. In

1559
terms of the convergence characteristic, MVMO is the fastest
among all methods. After the first 100 FEs, MVMO is able to
locate the optimal area in the search space. Moreover, it can
continue exploring that area for any better solution without
being trapped in one of local optimums. In contrast to other
greedy search algorithm such as DE/best/1 in DE, MVMO
overcomes the well-know premature convergence and local
stagnation. In other words, MVMO is an enhanced greedy
search algorithm.
200
190
180
Power losses (MW)
V. CONCLUSION
MVMO is a novel heuristic algorithm that has recently
shown great promises in dealing real-world complex
optimization problems. Besides its capability, the algorithm is
also simple to be implemented. In this paper, MVMO was
applied to solve the optimal reactive power dispatch problem.
The performance of MVMO was examined and compared
with other heuristic algorithms. The IEEE 57-bus and 118-bus
systems were used as the test cases. The simulation results
show that MVMO obviously has the better convergence speed
MVMO than the other algorithms. The final result is nearly the same
PSO for all algorithms. . MVMO performs outstandingly in the
DE 118-bus case both in terms of convergence and the minimum
JADE
JADE-vPS reached. . The CPU times used in both cases are relatively

170
high. The source of this occurrence is the power flow program
160 used in this paper. This issue will be corrected in our future
150
work. In summary, MVMO has a huge potential in dealing
with large-scale OPF problems.
140

130
REFERENCES

120
[1] H. W. Dommel and W. F. Tinney, "Optimal Power Flow
110
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Solutions," IEEE Trans. Power App. Syst., vol. PAS-87, no. 10,
No. of function evaluations x 10
4 pp. 1866-1876, Oct. 1968.
[2] J. A. Momoh, M. E. El-Hawary, and R. Adapa, "A review of
Fig. 10 Average convergence characteristics - case 118 selected optimal power flow literature to 1993 part I & II,"
Table II Statistical results case 118 IEEE Trans. Power Syst., vol. 14, no. 1, pp. 96-111, Feb. 1999.
[3] J. L. M. Ramos, A. G. Exposito, and V. H. Quintana,
MVMO PSO DE JADE JADE-vPS "Transmission power loss reduction by interior-point methods:
implementation issues and practical experience," IEE Proc.
Minimum 115.7932 120.8967 125.0250 119.1614 119.2006 Gener. Trans. Distrib., vol. 152, no. 1, pp. 90-98, Jan. 2005.
Average 116.8202 125.4868 127.4415 122.6774 125.5745 [4] V. H. Quintana and M. Santos-Nieto, "Reactive power-dispatch
by sucessive quadratic programming," IEEE Trans. Energy
Maximum 119.3584 135.3720 130.2292 132.0349 133.3443 Convers., vol. 4, no. 3, pp. 425-435, 1989.
Standard [5] M. Todorovski and D. Rajicic, "An initialization procedure in
deviation
0.7682 3.1634 1.1337 3.0467 3.8943 solving optimal power flow by genetic algorithm," IEEE Trans.
Power Syst., vol. 21, no. 2, pp. 480-487, May 2006.
[6] Q. H. Wu, Y. J. Cao, and J. Y. Wen, "Optimal reactive power
1.05 dispatch using an adaptive genetic algorithms," Int. J. Electr.
1.04
Power Energy Syst., vol. 20, no. 8, pp. 563-569, 1998.
[7] H. Yoshida, K. Kawata, Y. Fukuyama, S. Takamura, and Y.
1.03
Nakanishi, "A Particle Swarm Optimization for Reactive Power
1.02 and Voltage Control Considering Voltage Security
Assessment," IEEE Trans. Power Syst., vol. 15, no. 4, pp.
Voltage (p.u.)

1.01
1232-1239, Nov. 2000.
1 [8] A. A. A. Esmin, G. Lambert-Torres, and A. C. Z. d. Souza, "A
0.99 hybrid particle swarm optimization applied to loss power
minimization," IEEE Trans. Power Syst., vol. 20, no. 2, pp.
0.98
MVMO 859-866, May 2005.
PSO
0.97
DE
[9] C. H. Liang, C. Y. Chung, K. P. Wong, X. Z. Duan, and C. T.
0.96 JADE Tse, "Study of differential evolution for optimal reactive power
JADE-vPS flow," IEE Proc. Gener. Trans. Distrib., vol. 1, no. 2, pp. 253-
0.95
0 20 40 60 80 100 120 260, 2007.
Bus number [10] M. Varadarajan and K. S. Swarup, "Network loss minimization
Fig. 11 Load bus voltage profiles case 118 with voltage security using differential evolution," Elec. Power
Syst. Research, vol. 78, pp. 815-823, 2008.
Fig. 11 shows the voltage profiles at load buses resulting [11] Q. H. Wu and J. T. Ma, "Power system optimal reactive power
from all methods. Again, all optimization algorithms can dispatch using evolutionary programming," IEEE Trans. Power
maintain all bus voltages within the limits. The CPU time of Syst., vol. 10, no. 3, pp. 1243-1249, 1995.
all methods is approximately 15 minutes. [12] M. Tripathy and S. Mishra, "Bacteria Foraging-Based Solution
to Optimize Both Real Power Loss and Voltage Stability

1560
 Limit," IEEE Trans. Power Syst., vol. 22, no. 1, pp. 240 - 248
Feb. 2007.
[13] J. Zhang and A. C. Sanderson, "JADE: Adaptive Differential
Evolution with Optional External Archive," IEEE Trans. Evol.
Compt., vol. 13, no. 5, pp. 945-958, Oct. 2009.
[14] J. Lampinen and I. Zelinka, "On stagnation of the differential
evolution algorithms," in 6th Int. Conf. Soft Computing, 2000,
pp. 76-83.
[15] I. Erlich, G. K. Venayagamoorthy, and W. Nakawiro, "A mean-
variance optimization algorithm," in WCCI IEEE World
Congress on Computational Intelligence Barcelona, Spain,
2010.
[16] I. Erlich, W. Nakawiro, and M. Martinez, "Optimal Dispatch of
Reactive Sources in Wind Farms," in IEEE PES General
Meeting Detroit, USA, 2011.
[17] "Power Systems Test Case Archive," [Online]
http://www.ee.washington.edu/research/pstca/.
[18] "Standard PSO 2007 (SPSO-2007) on the Particle Swarm
Central," [Online] http://www.particleswarm.info, 2007.
[19] R. Storn and K. Price, "Differential evolution - a simple and
efficient heuristic for global optimization over continuous
spaces," Journal of Global Optimization vol. 11, pp. 341359,
1997.
[20] K. Price, R. Storn, and J. Lampinen, Differential Evolution - A
Practical Approach to Global Optimization: Springer Verlag,
2005.
[21] W. Nakawiro, "Voltage Stability Assessment and Control of
Power Systems using Computational Intelligence," Ph.D.
Thesis, University of Duisburg-Essen, 2011
1561
[22] C. Dai, W. Chen, Y. Zhu, and X. Zhang;, "Seeker Optimization
Algorithm for Optimal Reactive Power Dispatch," IEEE Trans.
Power Syst., vol. 24, no. 3, pp. 1218 - 1231, Aug. 2009
BIOGRAPHIES
Worawat Nakawiro received received his Ph.D. degree in electic
power systems from the University of Duisburg-Essen, Germany
where he is currently working as a postdoctoral fellow. His research
interests include computational intelligence with emphasis on
heuristic optimization and its applications in power systems and
voltage stability problems.
Istvan Erlich received received his PhD degree from the University
of Dresden/Germany in 1983. Since 1998, he is Professor and head of
the Institute of Electrical Power Systems at the University of
Duisburg-Essen/Germany. His major scientific interest is focused on
power system stability and control, modelling and simulation of
power system dynamics including intelligent system applications. He
is a member of VDE and IEEE.
Jos L. Rueda received the Ph.D. degree in electrical engineering
from the Universidad Nacional de San Juan, San Juan, Argentina, in
2009. Currently, he is a research fellow at the Institute of Electrical
Power Systems at the University of Duisburg-Essen. His current
research interests include power system stability and control, system
identification, power system planning, smart grids and wind power.