1

A Multi-Objective Genetic Algorithm Approach

to Optimal Allocation of Multi-Type FACTS
Devices for Power Systems Security
D. Radu, and Y. Bsanger, Member, IEEE

Abstract--A multi-objective programming procedure is used

for solving the problem of optimal allocation of Flexible AC
Transmission Systems (FACTS) devices in a power system. The
evolutionary approach consists of a multi-objective genetic
algorithm (MOGA), which is used to characterize the Pareto
Optimal Frontier (non-dominated solutions) and to provide to
Decision Makers and engineers insightful information about the
trade-offs to be made. In this paper, two technical and
economical objective functions are considered: maximization of
system security and minimization of investment cost for FACTS
devices. The optimization process is focused on three parameters:
the location of FACTS in the network, their types and their sizes.
For these proposals, we employed a hybrid software developed in
MatlabTM which uses the EUROSTAGTM software for load flow
calculations. The proposed procedures are successfully tested on
an IEEE 14-bus power system for several numbers of FACTS
devices.
Index Terms--FACTS devices, multi-objective
algorithms, optimal location, Pareto frontier.

genetic

I. NOMENCLATURE
DM: Decision Maker
FACTS: Flexible Alternating Current Transmission System
GA: Genetic Algorithm
MOGA: Multi-Objective Genetic Algorithm
MOP: Multi-Objective Problem
POF: Pareto Optimal Frontier
SVC: Static Var Compensator
TCSC: Thyristor Controlled Series Capacitor

II. INTRODUCTION

HE actual power systems are facing nowadays new

challenges due to deregulation and restructuring of the
electricity markets. The old structures, monopoly based, are
substituted by a competitive marketplace. The new resulting
power systems became more complex structures. All these
structures are confronting with problems raised by the
difficulties in building new transmission lines and the
significant increase in power transactions associated to
competitive electricity markets. Hence, in order to be able to
D. Radu and Y. Bsanger are with LEG (Power Engineering Laboratory of
Grenoble), UMR 5529 INPG/UJF - CNRS, ENSIEG BP 46, 38402 Saint
Martin dHres Cedex, France (e-mail: Daniel.Radu@leg.ensieg.inpg.fr and
Yvon.Besanger@leg.ensieg.inpg.fr).

1-4244-0493-2/06/$20.00 2006 IEEE.

obtain a high operational efficiency and networks security,

large interconnected systems have been built. In this context,
one possible solution to improve the system operation was the
use of FACTS technologies [1]. These are based on the state
of the art concept that incorporates power electronics
technologies, having a high efficiency, from technical point of
view, in transmission grid utilization.
The issue of optimal location of FACTS devices in power
systems has been researched and discussed widely and several
strategies were proposed. But most of those studies have taken
into account only the methods oriented towards technical
criteria (the capability of FACTS to enhance the system
loadability and to maximize the power system security was
shown in [2] and [3] respectively, the ability to control the
power flows is studied in [4] and the benefits to eliminate the
network congestions was presented in [5]) or to economical
approach which has minimized the overall system cost
function [6] and total generation fuel cost [7] or maximized
the return of investment as in [8].
This paper mainly focuses on the determination of optimal
allocation of FACTS devices into power systems, from both
technical and economical point of view, in order to provide a
better security level. To carry out these multi criteria
optimization problems, we employed a MOGA technique.
The outline of the paper is organized as follows: Section III
presents the FACTS allocation problem as a bi-objective
optimization problem, Section IV describes a multi-objective
technique used to solve the MOP, in Section V is shown the
description of the adopted MOGA, and, in Section VI, an
example system is used to validate the proposed approach.
Finally, section VII presents the conclusions of the paper.
III. FACTS ALLOCATION
A. FACTS devices
Todays AC power transmission networks are not designed
for easy voltage and power flow control in a deregulated
market giving rise to steady-state control problems and
dynamic stability problems. For various reasons, electricity
grid upgrades, and especially the construction of new
transmission lines, cannot follow the growing power plant
capacity and energy demand [9]. Due to this situation,
operators are looking for solutions to use the existing power
lines more efficiently. Therefore, the development of FACTS

devices responds to all grid operators needs, presented above.

They can be used for load flow control, voltage control and
stability improvement in transmission systems, as well as for
additional special applications [10].
In the present study, we have considered two classes of
FACTS such as shunt controllers and series controllers. From
the first category we studied the SVC, and, from the second
category, the TCSC. The SVC has already been in use for
three decades with excellent operating experiences. The
demand for SVC has increased continuously, as systems
became heavy loaded and problems arose regarding voltage
control. The TCSC has been used thanks to its applications in
the load flow control and the improvement of stability
conditions in the system. The developed method is based on
steady-state analyses, therefore, concerning the modeling of
FACTS devices into a power system, we used the steady-state
model of FACTS which is implemented in EUROSTAGTM
software. SVC and TCSC models are presented in [11].

branch loading. The coefficients q and r are used to penalize

more or less the overloads and voltage variations respectively.
For the presented study they are considered to be equal to 2.
As mentioned previously, it is important to take into
account the economical aspects of the FACTS devices
presence in the power systems due to high investment and
operating costs. Hence, the economical objective function
presented in (1) is represented by the total investment costs of
SVC, cinv SVC, and TCSC devices, cinv TCSC:
Fe = cinv SVC (rre SVC ) + cinv TCSC (rre TCSC )
(3)

B. Problem formulation
The presence of FACTS devices into a power system
brings many benefits, as quoted above. Nevertheless, in order
to reach a defined goal, it is important to choose the suitable
types of FACTS, an appropriate location and the rates of
these.
As indicated above, the goal of optimization was the
determination of optimal allocation of FACTS devices into a
power system in order to enhance the systems security level,
keeping in the same time a low investment cost in the new
equipments. Therefore, the presented problem becomes a
multi-objective optimization problem, with two different
criteria to be optimized and this can be expressed, in equation
form, as:

The determination of suitable location and parameters of

FACTS devices must be made by respecting the power flow
balance and the bounds of FACTS devices parameters.
Therefore, taking into account these aspects and considering
the equations (2)-(5), the formulation of the FACTS devices
optimal allocation problem can be expressed as follows:

Min F ( x) = [ Ft ( x), Fe ( x)]

Subject to x
c j ( x) = 0

(1)

j = 1K n

hk ( x) 0 k = 1K p
where F is known as the objectives vector, x represents a
decision vector, is the solution domain and cj(x) and hk(x)
are the equality and inequality problem constraints
respectively.
In this MOP, Ft(x) and Fe(x) represents technical and
economical criterion to be optimized respectively. Therefore,
from the technical point of view, the FACTS devices are
located in order to remove the overloads, to distribute
uniformly the load flows and also to prevent the under or over
voltages at networks busses. Consequently, we consider the
following technical objective function [12]:
a

2q

2r

b
Vm ref Vm
Sl

+
(2)
Ft =
wl
wm

S
V
m ref

l =1 l max
m =1
where Sl and Sl max are the apparent power in line l and the
apparent power rate of line l respectively. Vm is the voltage
magnitude at bus m and Vm ref is the bus m nominal voltage.
The weights wl and wm are determined in order to have the
same index value for 10% voltage difference and for 100%

where rre SVC and rre TCSC are the operating rate of the FACTS
devices in MVar.
According to [6], the investment costs shown above, given
in US$/kVar, are determined by the following relations:
cinv SVC ( s ) = 0.0003rre2 SVC 0.3051rre SVC + 127.38

(4)

cinv TCSC ( s ) = 0.0015rre2 TCSC 0.713rre TCSC + 153.75

(5)

Min F ( x) = [ Ft ( x), Fe ( x)]

(6)

Subject to x
E( f , g ) = 0
B( f ) b1 ; B( f ) b2
where f is a vector that represents the variables of FACTS
devices, g represents the power systems operating state, E(f,g)
represents the active and reactive power balance equations
corresponding to equality constraints and B(f) is the inequality
constraint concerning the FACTS devices bounds limits
(lower and upper limits represented by b1 and b2 respectively).
C. Possible approaches
Seeing that the optimization process was oriented towards
three parameters: FACTS location, their types and their rates,
which can take discrete and continues values, the case
discussed above becomes a combinatorial optimization
problem. The use of classical optimization methods for such
kind of problems often fails because of the exponentially
growing computational effort. There are several possible
solutions which can be used to solve these types of problems
such as: linear optimization, integer optimization, polynomial
time approximation and local search. Among these methods,
we chose to use local search (or heuristic methods), which is a
robust way to obtain good solutions to real problems in a
reasonable time [13]. More particularly, from the class of the
heuristic methods, we used the GAs [14] which are stochastic
search techniques based on the mechanics of natural selection
and natural genetics. They search for a solution inside a
subspace of the total search space, being able to give a good
solution in an acceptable computation time.
Furthermore to the aspects presented above, the problem

multi-objective character imposes the consideration of suitable

solving methods which are able to provide acceptable
solutions.
IV. MULTI OBJECTIVE TECHNIQUE
A. Genetics Algorithms
GAs start with an initial set of random solutions called
population. A population of candidate solutions, or
individuals, is maintained, and individuals made to compete
with each other for survival. Once evaluated, through the
fitness function calculation, stronger individuals have a greater
chance to contribute to the production of new individuals (the
offspring) than weaker ones, which may not even contribute at
all (selection procedure). Offspring are produced through
recombination, whereby they inherit features from each of the
parents, and through mutation, which can confer some truly
innovative features as well. In the next selection step (next
generation), offspring are made to compete with each other,
and possibly also with their parents. Improvement in the
population arises as a consequence of the repeated selection of
the best parents, which are in turn more likely to produce good
offspring, and the consequent elimination of low-performers
[15]. After several generations, the algorithm converges to the
best individual, which hopefully represents the optimal
solution to the problem [16].

Minimize

Economical objective

B. Multi-Objective Optimization
The problem described in Section III is, like mentioned
above, a multi-objective combinatorial optimization problem,
and thus it was necessary to use a multi-objective technique
for solving it. The use of multi-objective techniques gives
information on the consequences of the decision with respect
to all the defined objective functions. While traditional
optimization procedures result in one solution point only, the
MOP usually has no unique, perfect (or utopian) solution, but
a set of non-dominated, alternative solutions, known as the
Pareto-optimal set which define the POF. The principle of
dominance can be defined in the following form: a solution is
clearly better than (dominating) another solution, if it is better
or equal in all objectives, but at least better in one objective.
Using this principle, the set of best compromise solutions
results by removing all solutions that are dominated by at least
one other solution. The remaining solutions are all of equal
quality (indifferent). A mutual comparison of always two
solutions shows that each one is always better and worse in at
least one objective. This set of indifferent solutions is referred
as the POF (Fig. 1).
Feasible Search Space

POF
Technical objective
Minimize

Fig. 1. POF representation.

The POF offers complete information about the optimal

solutions of the problem and becomes an important
knowledge for the DM [17].
In this paper, the aim of the optimization is to determine the
POF of the problem described in the section above. The
choice of the optimal solution among the POF points remained
to DM. For these types of problems the Evolutionary
Algorithms (EAs) represent a standard tool. EAs can exploit
the population-based feature and converge in parallel to the
Pareto front. While optimizing, different solutions in the
population converge to different areas of the Pareto front, and
thus an approximation of the Pareto front can be obtained in a
single optimization run (trade-off surface). There are many
EAs described in the literature, reviews of this can be found in
[18]. For the present work, among the EAs, we chose to
employ MOGA technique.
C. Multi-Objective Genetic Algorithms
The MOGA technique was firstly proposed by Fonseca and
Fleming in 1993 [15]. It is an extension of the classical GAs.
The main difference between a conventional GAs and a
MOGA resides in the assignment of fitness. Once fitness has
been assigned to individuals, selection can be performed and
genetic operators applied as usual. The MOGA proposed a
Pareto-based ranking procedure of the individuals, where each
individual is assigned a rank which counts how many
individuals in the current population dominate them. In this
way, non-dominated individuals are always assigned the same
rank, independently of the shape of the trade-off surface.
The assignment of fitness according to rank, for the
MOGA, which was used in this paper by modifying the
traditional GAs, may be extended as follows:
For the each t generation, sort population in descendent
order according to rank rank(xi,t) of the individual xi. The
individuals rank is given by:
rank ( xi , t ) = 1 + pi(t )

(7)

pi(t)

where
are the number of individuals, in current
population, which dominate the individual xi. All nondominated individuals are assigned rank 1, see Fig. 2.
f2

3
1

5
2

POF

f1

Fig. 2. Multi-objective Pareto ranking.

Assign fitness to each individual by interpolating from the

best (rank 1) to the worst, applying the linear ranking fitness
scaling technique [16], [19].
Average the fitnesses of individuals with the same rank, so
that all of them will be sampled at the same rate.
Apply the fitness sharing [20] in the objective space to
maintain a population diversity.
Select the individuals for reproduction, by applying a
Stochastic Universal Sampling or Roulette Wheel Selection
techniques [14]-[16].

V. DESCRIPTION OF THE USED MOGA

To reach the aim of the optimization problem described in
the Section III, we developed a hybrid MOGA, which we
describe below.
A. Individual representation
MOGA require the parameter set of the optimization
problem to be coded as a finite-length string over some finite
alphabet. Since the goal of optimization was to allocate the
FACTS devices, taking into account three parameters
(location, type and rate), an individual is represented by three
strings of length nFACTS, where nFACTS is the number of FACTS
devices which want to be optimally allocated [1], [6], [12].
In the first individual string, one can find information
about the location of the FACTS devices in the power system.
It contains the integer numbers from 1 to nB, where nB is the
total number of branch system. In this string, a number can
occur only once, which means that only a FACTS device is
allowed into a branch. The SVC device is allocated to the
middle of the branch, by the introduction of a new bus.
The second string represents the coding for the types of
FACTS and can be assigned to discrete values 1 or 2, the first
corresponding to a SVC and the second to a TCSC device,
respectively.
The last individual string corresponds to the rate of FACTS
devices. Here, a special encoding was made to take into
account the FACTS devices bounds constraints presented in
(6) [1]. The encoded rates in this string, renc, take continue
values between 0 and 1, corresponding to the minimal and the
maximal rate that FACTS devices can take respectively.
In the Fig. 3 is shown, only for the exemplification purpose,
an example of individual representation composed of four
FACTS devices on a 14-branch network.
G3

G1
L1

- SVC

- TCSC

11

10
L2
L4

20
L3

L14

110
3

101

L5

L8

13
L11

L9

L13

L6
L7

L10

0.54 0.23

L12

120

2
2

1
4

1
9

0.85 Encoded rate

Type
Location

nFACTS

14

12

G2

G4

a)

b)

Fig. 3. Individual representation. (a) Power system. (b) Individual.

The real rated value of each FACTS device can be obtained

according to the different FACTS models under the following
criteria:
- SVC
As mentioned above, the EUROSTAGTM software was used
for load flow calculations and the FACTS models
implemented in it. Thus, the SVC is implemented in
EUROSTAGTM load flow module like a power injection in a
bus, declaring it as a PV node. Therefore, the SVC occurs in
the load flow by the control of the voltage at the bus where it
operates. Consequently, the rate value of the SVC, rSVC, in the
GAs encoding, is represented by the bus voltage reference, the
real rate, rre SVC, being obtained from the load flow results.
Furthermore, we considered a PV node with reactive power

limits. For this study, we considered that the SVC has an

operating range between -200 MVar and 200 MVar and a
reference voltage between 0.95 p.u. and 1.05 p.u. Hence, rSVC
is given by the following expression:
rSVC = 0.95 + 0.10 renc
(8)
- TCSC
Similarly, we used a TCSC device model from
EUROSTAGTM software. There, the TCSC is modelled, for
the load flow computation, like a controllable reactance
inserted in the system branch, which can increase or decrease
the line reactance, Xline. Therefore, the rate value of the TCSC,
rTCSC, in the GAs encoding, is represented by a reactance
which depends on the line reactance where TCSC operates.
The real rate, rre TCSC, is obtained after the application of TCSC
device sizing and rating algorithm described in [21] and [22].
In this paper, we considered that the TCSC had a working
range between 0.8Xline and +0.2Xline [1]. Thus, the rTCSC
value can be obtained by using the following formula:
rTCSC = 0.8 X line + X line renc
(9)
B. Initial population
A MOGA starts with an initial population of individuals,
which is randomly generated. This population is built from a
set of initial parameters such as: nFACTS the number of FACTS
devices which want to place optimally, which types of FACTS
devices wants to locate, nind the number of individuals in the
population and the data file of the analyzed power system.
Firstly, one assigns to the branches of the network numbers
from 1 to nB, after an individual is created. For this, among the
nB numbers, different nFACTS numbers are randomly chosen,
and are placed in the individual first string. Thereafter, the
second individual string is created by randomly drawn among
the types of FACTS devices selected to be optimal located
and, for the last string, random numbers from the range 0 to 1,
respectively, are generated. These operations are repeated nind
times to obtain the initial population.
C. Fitness evaluation
The next step in the MOGA, after the initial population
generation, is represented by the evaluation of the objective
function of the problem through the fitness function. In
general, the fitness coincides with the objective function, if it
is an unconstrained problem or an adaptation of this for a
constraints problem. For the MOP presented in the Section III,
the constraints are eliminated in the following manner: the
first constraints presented in (6) are implicitly respected
through the load flow computation and the second set of
constraints is eliminated by encoding in the first string of
individual representation. Hence, the objective function form
(6) can becomes a fitness for the MOGA. More particularity,
knowing that the GAs are designed for the search of a
maximization problem solutions, in this work we employed as
GAs fitness the inverse of the objective function (6).
Seeing that it is studied a bi-objective MOP, the fitness
evaluation was done in two steps. In the first step was
evaluated the technical objective through EUROSTAGTM
program. To this end, one decoded an individual and, from it,
a new data file was generated, which serve to compute the
load flow of power system with EUROSTAGTM. From the

load flow results, one can obtain all the information needed to
compute the technical objective function, Ft. In the second
step was read the load flow results, extracting from them the
rate values for the FACTS devices, rre SVC or rre TCSC. With
these values was evaluated the economical objective function,
Fe. These procedures are performed for all the individuals in
population.

good characteristics have a higher probability of contributing

one or more offspring in the next generation. The
reproduction operator may be implemented in GA in a number
of ways [14]. In this work, we considered Stochastic
Universal Sampling (SUS) and Roulette Wheel Selection
(RWS) methods [14]-[16]. After applying the reproduction
operator, the matting pool of the next generation is obtained.

D. Specifics MOGA operators

After the evaluation of both problem objectives, one can
apply the specific MOGA operators [15], which have been
defined in the Section IV-C. Hence, a rank of each individual
is determined according with (7), then a Pareto ranking
procedure is applied for all the individuals and a first trade-off
surface is drawn. After this, the population is sorted in
descendent order according to the rank and one assigns the
scaled fitness to each individual according with its position in
ordered population, but not with its original fitness. This is
done in order to obtain a population which was assigned a
single fitness raw, suitable for the application of classical GAs
operators. As shown above, we used a linear ranking fitness
scaling technique. The linear ranking is as follows:
x 1
fitness ( xind ) = 2 MAX + 2( MAX 1) ind
(10)
nind 1
where fitness(xind) is the scaled fitness of the individual ind,
xind is the position in the ordered population of individual ind
and the parameter MAX is the selective pressure, towards the
most fit individuals. In this paper, it was used a selective
pressure equal to 2, which means that the best individual will
have a scaled fitness equal to 2, the worst individual equal to 0
and other ones between these values.
After that, a single value of fitness is derived for each group
of individuals in a population with the same rank, through
averaging, so that all of them will be sampled at the same rate,
during the selection. Although all preferred individuals (nondominated) are assigned the same fitness, their actual number
of offspring may differ. The unbalance can easily accumulate
with the generations and results in a population drifting
towards an arbitrary region of the trade-off surface, a
phenomenon known as genetic drift [23]. To counteract the
genetic drift, the fitness sharing was applied. This technique
aims to promote the formulation and maintenance of stable
subpopulations (niches). It is based on the idea that
individuals in a particular niche have to share the available
resources. The more individuals are located in the
neighbourhood of a certain individual, the more its fitness
value is degraded. The neighbourhood is defined in terms of
an Euclidean distance measure between the individuals and
specified by the so-called niche radius. The fitness sharing
procedure is described in [20]. The use of fitness sharing is
restricted by the difficulty to determine the appropriate value
for niche radius. In the present paper, we used a solution given
in [24] for a bi-objective problem. Therefore, the application
of all procedures mentioned above leads to a population,
special prepared, which is ready to be reproduced.

F. Crossover and mutation

The main aim of crossover is to reorganize the information
of two different individuals and produce a new one in order to
promote the exploration of new regions in the search space.
Therefore, with the matting pool obtained from the process of
reproduction, crossover proceeds in two steps: in the first step,
individuals are mated randomly and in the second step
crossover is applied to the mated individual couples, using
coin tosses to select the crossing sites. There are several
procedures to perform crossover, in this paper we applied the
simple crossover and two-points crossover [14]. For the first
case, one randomly selects one cut site along two mated
individuals and exchanges the right parts of their, to generate
offspring (Fig. 4,a). For the second case, two crossing sites are
picked up uniformly at random, along two mated individuals.
After this, between the selected mated individuals, only the
strings parts which are between the crossing sites are
exchanged, elements outside these being kept to be part of the
offspring [1] (Fig. 4,b). The crossover may occur with a
probability pc which is generally close to 1.
The crossover operator uses can lead to a situation where,
in the first individual string, we have the same location for
several FACTS devices. As mentioned above, only one
FACTS device can be located on a transmission line. Hence, if
this situation occurs, an "arrangement of the FACTS
locations" process has to be applied [1], [12].

E. Reproduction
Reproduction is a process in which individuals are copied
according to their fitness, which means that individuals with

1 0.85

0.54 0.23

1
2
1
0.44
0.12
4
90.5 0.81
6
2
1
2
1

2
1

a)

10

0.54 0.23

0.54 0.23 0.5 0.81

2
2
1
0.44
0.12 3 1 0.85
6 10
2
1
1
1

2
2
1

1
1

10

0.54 0.12 0.5 0.85

0.85

2
1
1
0.44
0.12
90.5 0.81
6
4

2
2

2
1
2
0.44
0.23 9 1 0.81
5 10

1
4

b)

Fig. 4. Crossover operator. (a) Simple crossover. (b) Two-points crossover.

Mutation is used to introduce some sort of artificial

diversification in the population to avoid premature
convergence to a local optimum [6]. Mutation alters one or
more individual strings elements with a specified probability
pm. Every element of the individual strings has an equal
chance to be mutated. The mutation is made in all three strings
of the individual at the same time, by random generation of
new location of FACTS in the network (the random process is
done until the new location is different from all other locations
which are presents in the individual), new type and new
FACTS rate. Mutation occurs with a lower probability at the
initial generation and it is increased to the end of generations
in order to diversify the population.
G. Iterative process
Following reproduction, crossover and mutation, the new
population is ready to be tested. For this, we decode the new
individuals created by the MOGA and calculate the fitness

function like mentioned above. Hence, the operation of fitness

evaluation,
non-dominated
solutions
determination,
reproduction, crossover and mutation are repeated until the
maximal number of generation, ngen, is reached, this
representing the stop criterion of MOGA.
H. POF
Following the initial generation, repeated selection of nondominated individuals and production of offspring soon
produces a reasonable description of the trade-off surface of
the problem, at each generation the MOGA bringing the
solutions of the problem closer to the POF. At the end of the
iterative process, when the maximal number of generations is
reached, the trade-off surface is very close to or identical with
POF. Thus, the POF gives the Pareto optimal solutions of the
problem which are used by the DM. The proposed approach
made also at the last generation, a verification of individuals
from POF to see that the buses voltage and branch power
flows limits are respected. In this manner, it can be provided
to the DM information about out of bounds individuals.
The proposed MOGA is summarized in Fig. 5.
START
Input nFACTS and
MOGA parameters
Read the network data
Generate initial population
Evaluate the technical objective function
with EUROSTAG
Read load flow results

Create new population by:

selection/reproduction,
crossover and mutation

TCSC rating and sizing

G
2

G
1
4

G
G

11

10

12
13

14

Fig. 6. IEEE 14-bus power system diagram.

The simulations were oriented to three directions, the first

was made for method validation and to proof the robustness of
the MOGA. In the second, one studied the effects of the
number of FACTS devices to the power system security level
and for the third, one analyzed a power system contingence
regime.
A. MOGA robustness
For the considered power system, the MOGA was applied
considering several sets of parameters in order to prove its
capability to provide acceptable trade-offs close to the POF.
Thus, a first simulation was done, with an initial population
having 30 individuals and with a maximal number of
generations equal to 200 (Fig. 7). For the second simulation
we used a 50 individuals population, MOGA runs 300
generations (Fig. 8). For these simulations two FACTS
devices were optimal allocated, the types chosen being SVC
and TCSC.

Evaluate the economical objective function

Determine and save the trade-off surface
Apply the specifics MOGA operators
Stop criterion
reached?

No

Yes
Print the all MOP solutions and individuals from POF
END

Fig. 5. Flow chart of the proposed MOGA.

I. Developed software
The MOGA proposed above was implemented into a
software package developed in the MatlabTM language.
For the load flow computation, one has employed the
EUROSTAGTM software version 4.3.

a)

VI. SIMULATION RESULTS

For the validation of the proposed approach, we analyzed
the FACTS allocation problems applied to a test power
system. The simulations are done for IEEE 14-bus power
system shown in Fig. 6. Here, we represent the two parallel
lines that connected bus 1 with bus 2 as only one line. The
voltage magnitude limits of all buses are set to 0.95 for lower
bound and to 1.05 p.u. for upper bound.
The simulations are done in order to characterize the Pareto
Optimal Front and provide decision support in the MOP.

b)
Fig. 7. Trade - off representation for 2 FACTS (nind=30 and ngen=200).
(a) First generation. (b) Last generation (POF representation).

the problems being minimized. Furthermore, the number of

non-dominated solutions increases from a generation to other,
which shows the convergence of trade-off surface to POF. For
the presented simulations, in the last generation, one has 17
non-dominated solutions for the first case and 30 for the
second case, these constituting the Pareto optimal solutions of
our problem. Corresponding individuals from obtained POF
are presented in Table I (each individual is identifiable by a
number, present in the table first column and into the figures).
Hence, the DM has useful information about the optimal
allocation of FACTS devices in the studied network.
It can be seen that for both simulations we obtained almost
the same trade-off surface, which can well characterize the
POF of problem. Hence, we showed the capability of MOGA
to converge to the same POF, with different parameters sets.

Fig. 8. POF representation for 2 FACTS (nind=50 and ngen=300).

TABLE I
OPTIMAL ALLOCATION OF TWO FACTS DEVICES

Ind.
1
2
4
5
6
7
8
9
10

Ind.
1
3
6
7
9
10
12
13
16
17
18
21
22
24
25

Line

Type

9-14
7-9
9-14
9-10
12-13
9-10
9-14
9-10
9-14
9-10
9-14
6-12
9-14
3-4
7-9
9-10
9-14
9-10

TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
TCSC
TCSC
TCSC
TCSC
SVC
TCSC
SVC

Line

Type

9-14
12-13
9-14
10-11
9-14
7-9
9-14
9-10
9-14
12-13
6-12
6-13
9-14
7-9
6-12
10-11
9-14
9-10
6-12
7-9
9-14
6-11
9-14
9-10
9-14
12-13
9-14
12-13
6-12
9-10

TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
TCSC
TCSC
SVC
TCSC
SVC
TCSC
TCSC
TCSC
SVC

nind=30 and ngen=200

Rate
Ind.
Line
(MVar)
-0.94
13.37
-0.93
9.89
-0.006
9.87
-0.998
27.42
-0.998
27.42
-0.86
-0.221
-0.87
-1.908
-4.022
4.89
-0.933
9.89

11
15
20
22
23
26
29
30

12-13
9-14
9-14
9-10
10-11
9-14
7-9
6-12
9-14
9-10
9-14
7-9
12-13
9-10
9-14
7-9

nind=50 and ngen=300

Rate
Ind.
Line
(MVar)
-0.67
17.35
-0.91
2.08
-1.099
33.73
-0.766
6.58
-0.674
17.06
-0.389
8.52
-1.09
33.73
-0.419
2.13
-0.766
6.58
-0.376
33.49
-0.776
-0.242
-1.056
30.82
-0.673
17.35
-0.921
-0.005
-0.371
30.68

27
28
30
32
33
34
35
37
38
39
40
42
44
49
50

9-14
6-13
9-14
9-10
9-14
6-13
6-12
7-9
6-12
9-10
9-14
13-14
9-14
7-9
9-14
6-11
6-12
9-10
9-14
6-13
9-14
9-10
9-14
12-13
9-14
9-10
9-14
9-10
6-12
9-10

Type

Rate
(MVar)

TCSC
TCSC
TCSC
SVC
SVC
TCSC
TCSC
TCSC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC

-0.005
-0.898
-0.933
9.89
9.8
-0.892
-0.271
-0.248
-0.933
9.89
-0.983
22.23
-0.006
9.87
-0.996
22.91

Type

Rate
(MVar)

TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
TCSC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC
TCSC
SVC

-0.622
8.2
-0.805
19.61
-0.622
8.2
-0.376
33.49
-0.406
6.52
-0.745
9.31
-1.099
33.73
-0.776
-0.242
-0.387
19.52
-0.777
8.15
-0.952
6.57
-0.673
17.35
-0.952
6.57
-0.841
30.81
-0.371
30.68

In Fig. 7 and Fig. 8 are shown the trade-off surfaces

obtained for the first simulation and the second simulation,
respectively. In Fig. 7,a are presented the solutions for the first
generation and in Fig. 7,b and Fig. 8 the solutions for the last
generation. It can be seen that problem solutions converge to
POF from the first to the last generation, the two objectives of

B. Influence of nFACTS
The purpose of this simulation was to show that the optimal
allocation of FACTS devices allows to increase the power
system security level and that it depends on the number of
FACTS devices located. For this aim, a simulation with three
FACTS was made, by using the same MOGA parameters as in
the case presented above, in Fig. 7. The POF obtained is
shown in Fig. 9. It can be seen that when nFACTS increases, the
network security level is improved, too (this achievement is
seen into technical objective function decrease, comparatively
with the case presented in Fig. 7,b).

Fig. 9. POF representation for 3 FACTS (nind=30 and ngen=200).

C. Contingences analyses
Benefits of FACTS optimal allocation in power systems are
emphasized in this case study. Thus, a contingence regime is
considered, when the branch 2-3 is out off the service. In this
case, the branch 2-5 is overloaded at 102% of the branch
rating. After optimal allocation of two FACTS devices among
both of them, MOGA having the parameters nind=50 and
ngen=300, the power system recovers its secure state. Hence,
for the individual corresponding to the lower value of the
technical objective function, we obtained a reduction of the
load of the branch 2-5 about 86% of the branch rating. For
this case, the TCSC is located in branch 2-4 having a rate
equal to -11.51 MVar and the SVC is located in branch 9-10
having a rate equal to 27.39 MVar. The TCSC is located in
order to control the power flow through the lines and the SVC
to control the buses voltage level. The obtained POF is
presented in Fig. 10.

8
[8]

[9]
[10]
[11]
[12]
[13]
[14]
Fig. 10. POF representation for 2 FACTS (nind=50 and ngen=300).

[15]

VII. CONCLUSIONS
So far, the methods implemented for the resolution of the
FACTS allocation problem have been oriented to economical
or technical aspects separately. The present paper makes use
of recent advances in multi-objective evolutionary algorithms
to develop a method for the combinatorial optimal allocation
of FACTS into power systems. Optimizations were performed
on three parameters: the locations of FACTS devices, their
types, and their rates. It was considered as optimization
criteria the maximization of the power system security and the
minimization of total investment cost of FACTS devices.
Implementation of the proposed MOGA has performed well
when it was used to characterize POF of the FACTS optimal
location problem.
Three study cases are conducted using the IEEE 14-bus
system. The results show that the proposed MOGA can
produce good solutions. It is shown that FACTS can enhance
the power system security through optimal allocation and also
when their number is increasing, but with a corresponding
increase of total investment costs. All these results illustrate
the effectiveness and robustness of the proposed approach.
Further developments are preconized in order to apply the
proposed method to a real power system.
VIII. REFERENCES
[1]

[2]
[3]
[4]
[5]
[6]

[7]

Y. Bsanger, J. C. Passelergue, N. Hadj-Said, and R. Feuillet,

"Improvement of power system performance by inserting FACTS
devices," in Proc. 1996 IEE AC and DC Power Transmission Conf., pp.
263-268, London, UK.
S. Gerbex, R. Cherkaoui, and A. J. Germond, "Optimal location of multitype FACTS devices in a power system by means of genetic algorithms,"
IEEE Trans. Power Systems, vol. 16, no. 3, Aug. 2001.
D. Radu, and Y. Bsanger, "Blackout prevention by optimal insertion of
FACTS devices in power systems," in Proc. 2005 IEEE Future Power
Systems Conf., Amsterdam , NL.
P. Paterni, S. Vitet, M. Bena, and A. Yokoyama, "Optimal location of
phase shifters in the French network by genetic algorithm," IEEE Trans.
Power Systems, vol. 14, no. 1, pp. 37-42, Feb. 1999.
Y. Lu, and A. Abur, "Static security enhancement via optimal utilization
of Thyristor-Controlled Series Capacitors," IEEE Trans. Power Systems,
vol. 17, no. 2, pp. 324-329, May 2002.
L. J. Cai, I. Erlich, and G. Stamtsis, "Optimal choice and allocation of
FACTS devices in deregulated electricity market using genetic
algorithms," in Proc. 2004 IEEE Power Engineering Society Power
Systems Conference and Exposition, pp. 201-207.
T. S. Chung, and Y. Z. Li, "A hybrid GA approach for OPF with
consideration of FACTS devices," IEEE Power Engineering Review, pp.
47-50, Feb. 2001.

[16]
[17]
[18]
[19]
[20]
[21]

[22]

[23]
[24]

Y. Matsuo, and A. Yokoyama, "Optimization of installation of FACTS

device in power system planning by both tabu search and nonlinear
programming methods," in Proc. 1999 Intelligent System Application to
Power System Conf., pp. 250-254, Rio de Janeiro, Brazil, Apr. 1999.
R. Grnbaum, M. Noroozian, and B. Thorvaldsson, "FACTS powerful
systems for flexible power transmission," ABB Review, no. 5, 1999.
D. Povh, "Applications of FACTS to system," EPSOM98 Conf., Zurich,
Swiss, Sep. 1998.
EUROSTAG 4.3. Operation Manual, TRACTEBEL Inc. / Electricit de
France Inc., 2004.
S. Gerbex, "Metaheuristics applied to optimal location of FACTS
devices into a power system," Ph.D. dissertation, Dept. Electrical Eng.,
EPFL - Swiss Federal Institute of Technology Lausanne, 2003.
K. I. Aardal, S. van Hoesel, J. K. Lenstra, and L. Stougie, "A decade of
combinatorial optimization," in CWI Tracts (122, pp. 5-14), 1997.
D. E. Goldberg, Genetic Algorithms in search, optimization, and
machine learning, Addison-Wesley Publishing Company, Inc., 1989.
C. M. Fonseca, "Multiobjective Genetic Algorithms with application to
control engineering problems," Ph.D. dissertation, Dept. of Automatic
Control and Systems Eng., The University of Sheffield, 1995.
M. Gen, and R. Cheng, Genetic Algorithms & engineering design, John
Wiley & Sons, Inc., 1997.
D. Bche, "Multi-Objective Evolutionary Optimization of Gas Turbine
Components," Ph.D. dissertation, Swiss federal institute of technology,
Zurich, 2003.
C. M. Fonseca, and P. J. Fleming, "An overview of Evolutionary
Algorithms in multiobjective optimization," Evolutionary Computation
Review, vol. 3, no. 1, pp. 1-16, 1995.
J. E. Baker, "Adaptive selection methods for Genetic Algorithms", Proc.
ICGA 1, pp. 101-111, 1985.
E. Zitzler, "Evolutionary Algorithms for multiobjective optimization:
methods and applications," Ph.D. dissertation, Swiss federal institute of
technology Zurich, 1999.
D. N. Kosterev, W. A. Mittelstadt, R. R. Mohler, and W. J. Kolodziej,
"An application study for sizing and rating controlled and conventional
series compensation," IEEE Trans. Power Delivery, vol. 11, no. 2, pp.
1105-1111, Apr. 1996.
H. G. Sarmiento, E. Estrada, J. Naude, C. Tovar, M. A. Avila, and C.
Fuentes, "An example in sizing a series compensation scheme with fixed
and variable capacitors," in Proc. 2001 IEEE AC-DC Power
Transmission Conf., pp. 72-77.
A. M. S. Zalzala, and P. J. Fleming, Genetic algorithms in engineering
systems, The Institution of Electrical Engineering, Control Engineering
Series 55, 1997.
J. Horn, and N. Nafpliotis, "Multiobjective Optimization using the
Niched Pareto Genetic Algorithm," University of Illinois at UrbanaChampaign, Illinois, USA, Tech. Rep IlliGAl Report 93005, Jul. 1993.

IX. BIOGRAPHIES
Daniel Radu was born in 1975 in Bucharest,
Romania. He obtained a Ph.D. degree in Electrical
Engineering, in 2004, from the University
POLITEHNICA of Bucharest, Romania. From
1998 to 2004 he has been assistant professor of
Electrical Engineering at the same University. He is
currently post doc researcher at the Power
Engineering Laboratory of Grenoble LEG, France.
His research interests include FACTS devices and
power systems analyses and control.
Yvon Bsanger (M01) was born in 1967 in Brive,
France. He received the Ph.D in Electrical
Engineering from the National Polytechnic Institute
of Grenoble (INPG) in 1996. He is currently
Associate Professor at the National Superior
Electrical Engineering School of Grenoble
(ENSIEG) and the Power Engineering Laboratory of
Grenoble (LEG). His research interests are FACTS
devices, distribution networks operation and
reliability and power system dynamic security.