NEW EVOLUTIONARY PARTICLE SWARM ALGORITHM (EPSO)

APPLIED TO VOLTAGE/VAR CONTROL

Vladimiro Miranda

Nuno Fonseca

[email protected]

[email protected]

INESC Instituto de Engenharia de Sistemas e Computadores do Porto

P. Repblica 93 4050 Porto Portugal Fax 35122084172
FEUP Faculdade de Engenharia da Universidade do Porto, Portugal
Abstract This paper presents a new optimization model
EPSO, Evolutionary Particle Swarm Optimization,
inspired in both Evolutionary Algorithms and in Particle
Swarm Optimization algorithms. The fundamentals of the
method are described, and an application to the problem of
Loss minimization and Voltage control is presented, with
very good results.

EPSO joins together the best of two worlds. It is a

Particle Swarm algorithm, because there is exchange of
information among solutions, when they are successively
moved around in the search space; and it is an
Evolutionary Computation method, because solution
characteristics are mutated and passed to the following
generations by the action of a selection mechanism.

Keywords: Evolutionary Algorithms, Particle Swarm

Optimization, Voltage/Var control

In a classical PSO model, particle movement is

conditioned by three strategic parameters: inertia, memory
and coordination (information exchange).

1. INTRODUCTION
This paper has the main objective of introducing to the
Power System community a new and powerful metaheuristic hybrid variant called EPSO Evolutionary
Particle Swarm Optimization.
EPSO is a general-purpose algorithm and it can thus be
applied to a diversity of problems in any scientific area.
However, in order to illustrate the technique, we have
selected a problem in the Power Systems environment and
will therefore present in the paper, in the application
sections, a solution for the Voltage/Var control problem
obtained by EPSO and the comparison of its performance
with existing methods
The Particle Swarm Optimization is an optimization
algorithm that was introduced in 1995 by Kennedy [1].
We will refer to it as: Classic PSO. Imagine that we have
a population of particles looking around in a given search
space for the global optimum. This particle movement
mimics, in a way, the coordinated movement of flocks of
birds, schools of fish or swarms of insects: this is a good
image of a PSO optimization algorithm.
In Evolutionary Algorithms, there is no coordination in
the movement of individuals within the search space.
However, the powerful selection procedure allows
solutions with superior characteristics to pass these from
generation to generation, while the mutation (and
recombination) schemes produce diversity in the solution
pool.

In previous applications of the Classic PSO and other

variants like CPSO (Cooperative PSO) [2], the strategic
parameters of the algorithms were set to certain values
that had already been used with god results. But there is
no valid explanation to sustain that we should use a
particular value for those parameters if we have a different
problem.
Also, we cant say that a certain value is the best during
all the process of optimization. That had already been
realized for the inertia factor. This parameter is usually
decreased as the number of iteration increases [3].
EPSO [4] defines these parameters as the genotype of a
moving solution. Therefore, they are subject to mutation
and the particles holding them as phenotypes are subject
to selection. This scheme turns out to be a successful selftuning mechanism, a self-adaptive evolutionary process
acting on strategic parameters, to use the language of
the Evolution Strategy community.
As we will show, EPSO has a better behavior than
Classical PSO (namely, it is robust, insensitive in a large
degree to initial values of parameters) and it also has a
better behavior than other meta-heuristics (in this paper, a
comparison will be made with simulated annealing).
The hybrid characteristics of Evolutionary and of
Particle Swarm model give it guaranteed convergence
properties. In terms of efficiency, therefore, lower bounds
are guaranteed, but experience demonstrates that there is
an effective acceleration and a better search for the
optimum than classical approaches.

2. EPSO DESCRIPTION
In EPSO, each particle (solution at a given stage) is
defined by the following characteristics:

position in the search space ( xik ; value of the

coordinate position i, for the k particle)

velocity ( v ik ; value of the coordinate velocity i for

the k particle).

At a given moment, there is at least one particle that

holds the best position in the search space. The population
of particles is aware of such position, represented as
( x ibest , value of the coordinate position i, for the best
particle).
Each particle also keeps track of its previous best
position ( x ik , mem , value of coordinate position i,
memorized as its previous best, for the k particle).
The particles will reproduce and evolve along a
number of generations, according to the following steps:
- Replication: each particle is replicated a number r of
times, giving place to identical particles (in this paper
we take r = 1).
- Mutation: the strategic parameters of the replicated
particles undergo mutation according to:

* w ik, j = w ik, j + N ( 0 , 2 )

(1)

where is a learning dispersion parameter and

N(0,1) is a random number following a the
normalized Gaussian distribution with zero mean
and variance 2.
The strategic parameters are randomly set between 0
and 1 at the beginning of the algorithm. In each
iteration, the strategic parameters of the replicated
particles are mutated according to equation (1). In
this equation, j can be the inertia, memory or the
coordination factor.
- Reproduction (movement): each particle generates as
offspring a new particle according to the
transformation process, similar to the Classic PSO
basic equation:

*vik = wik,inertiavik + wik,mem(xik,mem xik ) + wik,coop(xibest* xik ) (2)

*xik = xik +*vik

(3)

The offspring is held separately for the original

particles and for the mutated particles.
Furthermore, instead of defining a crisp best-so-far
point as a target, the particles are attracted to a sort
of foggy best-so-far region (another change
relative to Classic PSO).

This is done by introducing random noise in the

definition of the best-so-far point:

x ibest* = x ibest + ' N (0,1)

(4)

is a noise dispersion parameter, usually small, and

N(0,1) is a random number following a the
normalized Gaussian distribution with zero mean
and variance 1.
- Evaluation: each offspring particle plus the originals
are evaluated according to their current position.
- Selection: among the offspring of a particle, with and
without mutated parameters, a stochastic tournament
is played to select the particle that will survive to the
next generation.
As in many other meta-heuristics, EPSO deals with
inequality constraints through a penalty strategy. In the
case of EPSO, the selective pressure applied helps in
eliminating the individuals or particles with excursions
outside the feasible domain, which receive a penalized
fitness value.
3. TESTING EPSO

In this section we illustrate the superiority of EPSO

regarding the Classic PSO algorithm, in solving classical
difficult test problems.
3.1 Test functions

Schaffers function:

f1(x) = 0.5 +

(sin x2 + y2 )2 0.5

(5)

(1.0 + 0.001(x2 + y2 ))2

Rosenbrock function:
n

f 2 (x ) =

(100 (x i+1 x i 2 ) 2 + (x i 1) 2 )

(6)

i =1

Sphere function:
n

f 3 (x ) =

xi2

(7)

i =1

Alpine function:
f 4 ( x ) = sin( x 1 ) sin( x 2 ) x 1 x 2

(8)

The parameters used in these functions are presented in

Table 1. The threshold used as the stopping criterion is
listed in the Stop column.
Function
f1
f2
f3
f4

n
2
30
30
2

Domain
n
[-50 , 50]
n
[0 , 30]
n
[-50 , 50]
n
[0 , 100]

Stop
1.0E-10
100
0.01
98.9627

Table 1: Parameters used in the test functions.

3.2 Results in the test functions

The following pictures illustrate the typical

convergence in the test functions, for the EPSO and PSO.

S c h a ffe r F6
1 .E+ 0 0
0

50000

100000

150000

200000

1 .E- 0 2
EPS O
Fitness

1 .E- 0 4

PS O

1 .E- 0 6
1 .E- 0 8
1 .E- 1 0
1 .E- 1 2
N u m b e r o f e v a lu a t io n s

Figure 1 - Typical convergence in the Schaffers function.

The results presented in Figures 1, 2, 3 and 4 show a

clear superiority of the EPSO algorithm. The PSO results
could be optimized if weve tuned by hand the strategic
parameters. EPSO was able to provide better results
independently of the strategic parameter initialization.
If we are trying to optimize a different problem (ex:
Optimal Power Flow), where we dont now which are the
better strategic parameters, then EPSO is certainly better
because of the self-tuning mechanism.
To demonstrate the superiority of EPSO over the
Classic PSO we compare the average number of
evaluations that both algorithms need to reach the
stopping criterion. The maximum number of evaluations
was fixed in 200000. Table 2 presents the results of this
test.

Function
R o s e n b ro c k
EPSO
PSO

1 .E + 0 9
1 .E + 0 8
1 .E + 0 7
Fitness

1 .E + 0 6
1 .E + 0 5

EPSO

PSO

f1

11862.2

59547.0

f2

27005.3

180310.8

f3

16421.4

161625.0

f4

78539.8

199190.1

1 .E + 0 4
1 .E + 0 3
1 .E + 0 2
1 .E + 0 1
1 .E + 0 0
0

50000

100000

150000

200000

N u m b e r o f e v a lu a tio n s

Table 2: Comparison of EPSO with the Classical PSO: average

number of evaluations.

Figure 2 - Typical convergence in the Rosenbrock function.

We also compare the average results of both algorithms

for a fixed number of evaluations. So, considering a
number of evaluations of 200000 (20 particles over 5000
iterations in the EPSO algorithm and 20 particles over
10000 iterations in the Classical PSO) we obtained the
following results:

Sph e re
1 .E+ 0 5
1 .E+ 0 4

EPS O

1 .E+ 0 3

PS O

Fitness

1 .E+ 0 2
1 .E+ 0 1
1 .E+ 0 0
1 .E- 0 1 0

50000

100000

150000

200000

1 .E- 0 2
1 .E- 0 3
1 .E- 0 4
N u m b e r o f e v a lu a t io n s

Figure 3 - Typical convergence in the Sphere function.

Function

EPSO

PSO

f1

2.15E-13

5.45E-11

f2

33.8828

114.443

f3

7.81E-04

1.91E-02

f4

98.9627

86.1071

Table 3: Comparison of EPSO with the Classical PSO: average

results.

A lp in e
PS O
100

The results presented in Table 2 and Table 3 were

obtained with a population of 20 particles, over 500
simulations.

EPS O

95
90
Fitness

85
80
75
70
65
60
55
50
0

100

200

300

N u m b e r o f e v a lu a t io n s

Figure 4 - Typical convergence in the Alpine function.

400

Notice that the position and the strategic parameters

(inertia, coordination factor and memory) were always
initialized randomly for each particle. The results of the
PSO would have perhaps a margin for improvement if the
strategic parameters were even better tuned by hand, but
this would involve a tedious work of experimentation case

by case. The EPSO algorithm was able to provide

immediately good results independently of such
initialization.
This is a very important improvement in the algorithm,
because the results of Classic PSO are reported to be very
dependent of the strategic parameter initialization [3], and
this has been confirmed by our experience. This statement
does not mean that one could not find solutions with less
good initial PSO parameters but the performance of the
PSO algorithm, in our experiments, never reached the
quality of EPSO, did not display on average the same
quality of results and, most important, did not display the
same robustness, which is vital for a practical application
the users must trust the algorithm, must believe it gives
reliable and consistent results, must be confident that, if
they run it a number of times, they will get the same kind
of answer.
4. APPLICATION OF EPSO TO VOLTAGE/VAR
CONTROL

All these equations, written in a general form, must be

understood as representing a full AC model, with losses
evaluated, for instance, with a Newton-Raphson
algorithm. Because these are well known equations, we
felt we could take the liberty of adopting the above
representation.
We may have also other objectives, such as the
preference for keeping control margins, i.e., searching for
solutions that do not require the set points of controllers to
be at their maxima or minima. This means that one is
facing a multicriteria problem, with two objectives:
Minimize losses
Minimize distance of control variables from
nominal set points (usually, the center on the
intervals defining their range of variation).
In fact, this is achieved in practice by applying a
penalty factor to the fitness function, such as depicted in
Figure 5.

4.1 Loss reduction in distribution systems

The application of PSO-like algorithms to the

Voltage/Var control problem was pioneered by authors in
Japan and reported in [6][7][8]. Their models included a
form of blending evolutionary concepts with the PSO
algorithm, with positive results. However, their valuable
work remained one step away from a true self-adaptive
approach, which is what this paper now presents.
We illustrate EPSO in a loss reduction-Voltage/Var
control problem for a didactic example with the IEEE 24
nodes/36 branches network defined in [9]. This network
also includes 31 transmission lines, 5 transformers, 11
capacitor banks and 9 synchronous generators. The size of
a problem of this nature, however, is not related really
with the size of the network but with the number of
controllers available.
For the sake of a comparison with a competing
algorithm, based on Simulated Annealing, we took as
control variables only the set point of transformers and
capacitor banks.
This Simulated Annealing [10] [11], algorithm is a well
tested application developed by INESC Porto and
included in a commercial DMS, used by a number of
utilities. The problem of Voltage/Var control can be
formulated as follows:
Minimize
Subject

(u, x)

(9)

(u, x, p) = 0

(10)

(u, x, p) 0

(11)

Equation (9) is the objective function of this problem

and, in general, represents the active losses. The
constraints of this problem, (10) and (11), are respectively
the power flow equations and operation limits, namely
bands of admissible voltage values at nodes.

Penalty

min

Nominal
setpoint

Max

Control
variable

Figure 5 Example of penalty function to be added to the loss

function (per control variable, scaled by a weighting factor) to
favor solutions that do not push controls to their limits

The Voltage/Var control problem in distribution

systems is usually a problem of minimizing losses and
controlling voltage levels, by acting on transformer taps
and on capacitor bank taps. It is rare to find synchronous
generators directly connected to the network where one
could act on their excitation. However, EPSO can deal
with these variables as well, with excellent results.
4.2 Results of the loss reduction problem

In order to compare EPSO results with those obtained

with the Simulated Annealing (SA) application we needed
to establish the same stopping criterion. As the Simulated
Annealing already had this criterion fixed as a certain
number of iterations without improvement in the best-sofar solution, we used the same criterion. For this particular
exercise, the maximum number of iterations allowed
without improvement in the solution was fixed in 270.
In this particular application, all the variables of
control are discrete (set point of transformers and
capacitor banks). There is a version of Discrete PSO [12],
but as for now, the EPSO only deals with continuous
variables.

Weve used probabilistic rounding to solve this

problem. Instead of using simple rounding, i.e., consider
the nearest value, weve considered that the probability of
rounding to the nearest discrete value increases as the
distance decreases. On average, the value of the variable
is probabilistically rounded to the nearest discrete value,
but there is always the possibility that it is not, at any
given point.
This scheme avoids trapping in local discrete values,
and has all the flavor of the techniques used in
evolutionary computing.
In terms of convergence comparison between both
algorithms, we can immediately reveal that:
- EPSO finds its best solution in less iterations.
- the initial solution is better for the EPSO, because it
has a population of particles, while the Simulated
Annealing only starts with one initial solution.
- there is an extra computing effort in applying EPSO,
when compared to the Simulated Annealing option
(measured in the number of load flows run);
- EPSO consistently discovers better solutions than the
Simulated Annealing algorithm.
A typical convergence pattern observed for both
algorithms can be observed in Figure 5, where EPSO (as
usual) found a better solution than SA.
We tested EPSO with different population sizes. The
objective of this test was to find out the influence of
population size in the results. As it can be seen in Table 4,
the quality of the solution improves if we increase the
number of particles. Of course, there is a price to pay in
terms of computing effort.
As for now, we were not able to establish a secure rule
for defining the optimal number of particles. Our
assumption is that this will depend on the complexity of
the problem, and with increasing complexity it will be
necessary to increase the number of particles.

Average
losses
(MW)

Std. Deviation
(kW) to the
best solution

EPSO 2 particles

61.7947

3.10334

EPSO 5 particles

61.7912

2.64314

EPSO 10 particles

61.7889

1.96607

EPSO 20 particles

61.7880

1.48473

Simulated Annealing

61.7921

9.81175

Table 4: Comparison of EPSO with Simulated Annealing

As we can see in Table 4, the EPSO reveals superiority

in terms of the solution found (both in the best solution
discovered and in the average optimum obtained in 1000
runs, as seen in the Table) and in terms of its robustness
(evaluated as the root of the mean square error, or
standard deviation, relative to the best solution found).
In particular, EPSO gives consistently a near-optimum
result, while the Simulated Annealing model failed many
times to reach a solution as good (and thats why the
dispersion of results in this case is much larger than with
EPSO).
Therefore, EPSO is a much more reliable algorithm for
practical applications.
4.3 Voltage control

For this test weve increase the reactive load in bus 8 of

the same IEEE 24-bus system. The voltage at this bus
became very low and we run the EPSO algorithm to redispatch the reactive power in order to set the voltage
back inside the limits.
As it can be seen in Figure 6, the algorithm was able to
find a new set point, to both transformers and capacitor
banks, which forced the voltage into the acceptable limit
(0.9 1.1 p.u.).
The Simulated Annealing algorithm failed to obtain a
feasible solution for this case.

Simulated Anneling vs Evolutionary PSO

Voltage control

74000

EPSO 20 particles

Figure 5 - Comparison in convergence between EPSO

and Simulated Annealing

46

43

40

37

34

0,884
31

2000

28

1500

25

1000

Number of iterations

22

500

19

0,888

16

60000

13

62000

0,892

64000

0,896

10

66000

68000

0,900

70000

Voltage in bar 8 (p.u.)

Active Losses (kW)

0,904

Simulated Anneling

72000

Number of iterations

Figure 6 - Voltage Control with EPSO evolution of the

controlled voltage along the iterations of an EPSO algorithm.

5. CONCLUSIONS
This paper reports two important results:

A new optimization technique, with roots both in

Evolutionary Computing and in Particle Swarm
algorithms.
A new model for loss minimization and voltage
control

First of all, there is a new successful meta-heuristic

tool, available for optimization of complex problems with
multiple local optima EPSO, the Evolutionary Particle
Swarm Optimization method.
EPSO joins together the characteristics of Evolutionary
Algorithms and of Particle Swarm Algorithms.
From an Evolutionary Computing point of view, there
is another operator introduced, side by side with
recombination and mutation, which generates new (and
promising) solutions in the search space it is
Reproduction in the form of Particle Movement.
From a Particle Swarm point of view, there is a selfadaptive tuning of the algorithm by evolutionary
adjustment of the parameters controlling particle
movement.
Both points of view are legitimate and justify the
remarkable convergence characteristics of the method.
The second important result is that EPSO proves very
successful in solving a Power System optimization
problem minimizing losses in a transmission system. In
fact, EPSO performed better than a Simulated Annealing
model that has been used by utilities, both in the quality of
the solution discovered and in the robustness of the result
(dispersion around the best result, found in a number of
repeated runs).
In the tests done, the Simulated Annealing algorithm
demanded a somewhat smaller computer effort (measured
in the number of load flows required) but failed
completely to discover the best solutions, while EPSO
was able to converge to them in all cases.
Furthermore, EPSO was again successful in a Voltage
Control problem, easily discovering a solution for a
difficult problem where other techniques experiment
difficulties in converging.

[2] F. van den Bergh, A.P. Engelbrecht, Training Product

Unit Networks using Cooperative Particle Swarm
Optimization, International Joint Conference on
Neural Networks (IJCNN), Washington D.C., 2001.
[3] Yuhui Shi, Russell C. Eberhart, Parameter Selection
in Particle Swarm Optimization, Proceedings of the
Seventh Annual Conference on Evolutionary
Programming, 1998.
[4] Vladimiro Miranda, Nuno Fonseca, EPSOEvolutionary
self-adapting
Particle
Swarm
optimization, internal report INESC Porto, July 2001
(obtainable from the authors by request).
[5] F. van den Bergh, A.P. Engelbrecht, Effects of
Swarm Size on Cooperative Particle Swarm
Optimizers, Proceedings of the Genetic Evolutionary
Computation Conference (GECCO), , 2001.
[6] Yoshida, H., Fukuyama, Y., Takayama, S. and
Nakanishi, Y., A particle swarm optimization for
reactive power and voltage control in electric power
systems considering voltage security assessment,
IEEE Proc. of SMC '99 , Vol. 6, pp.497 -502, 1999.
[7] H. Yoshida, K. Kawata, Y. Fukuyama, S. Takayama,
and Y. Nakanishi, A Particle Swarm Optimization for
Reactive Power and Voltage Control Considering
Voltage Security Assessment, IEEE Trans. on Power
Systems, vol. 15, no. 4, pp.1232-1239, Nov. 2000.
[8] Fukuyama, Y., Yoshida, H., A particle swarm
optimization for reactive power and voltage control in
electric power systems, IEEE Proc. of Evolutionary
Computation 2001 , Vol.1 , pp. 87 -93, 2001.
[9] Reliability Test System Task Force of the Application
of Probability Methods Subcommittee, IEEE
Reliability Test System , IEEE Trans. On Power
Apparatus and Systems, vol. PAS-98, no. 6, Nov./Dec.
1979.
[10]Jorge Pereira, J. Tom Saraiva, Maria Teresa Ponce
de Leo, "Identification of Operation Strategies of
Distribution Networks Using a Simulated Annealing
Approach", Proceedings of IEEE Budapest Power
Tech'99, paper BPT99-357-17, August 1999.

REFERENCES

[11]Manuel Matos, Maria Teresa Ponce de Leo, J. Tom

Saraiva, J. N. Fidalgo, et al., "Meta-heuristics Applied
to Power Systems", Proceedings of MIC2001 - 4th
Metaheuristics International Conference, Porto,
Portugal, vol.2, pp.483-488, July, 2001.

[1] Kennedy, J., R.C. Eberhart, Particle Swarm

Optimization, IEEE International Conference on
Neural Networks, Pert, Australia, IEEE Service
Center, Piscataway, NJ., 1995

[12]Kennedy, J. and Eberhart, R. C., A discrete binary

version of the particle swarm algorithm, Proc. Conf.
on Systems, Man, and Cybernetics, 41044109.
Piscataway, NJ: IEEE Service Center, 1997.

One expects that EPSO may be applied with equal

success to other problems in Power Systems.