EUROPEAN TRANSACTIONS ON ELECTRICAL POWER

Euro. Trans. Electr. Power (2010)

Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/etep.494

A multi-objective optimal power flow using

particle swarm optimization

J. Hazra1*,y and A. K. Sinha2

1
IBM India Research Lab, Bangalore, India
2
Department of Electrical Engineering, IIT Kharagpur, India

SUMMARY
This paper presents a multi-objective optimal power flow technique using particle swarm optimization. Two
conflicting objectives, generation cost, and environmental pollution are minimized simultaneously. A multi-
objective particle swarm optimization method is used to solve this highly nonlinear and non-convex
optimization problem. A diversity preserving technique is incorporated to generate evenly distributed Pareto
optimal solutions. A fuzzy membership function is proposed to choose a compromise solution from the set of
Pareto optimal solutions. The algorithm is tested on IEEE 30 and 118 bus systems and its effectiveness is
illustrated. Copyright # 2010 John Wiley & Sons, Ltd.

key words: multi-objective optimization; pareto-optimality; particle swarm optimization; optimal

power flow

1. INTRODUCTION

In operation and planning of power systems, operators need to make decisions with respect to different
objectives. Hence, several tools have been developed to assist the operators. Optimal Power Flow
(OPF) is one of them which helps the operators in running the system optimally under specific
constraints. A lot of research starting from early 1960s has been done in this field to minimize the total
generation cost. After the Clean Air Act Amendments (Kyoto Protocol) in 1990, operating at minimum
cost maintaining the security is no longer the sufficient criterion for dispatching electric power.
Minimization of polluted gases is also becoming mandatory for the generation utilities in many
countries. Hence, OPF problem becomes a multiobjective optimization problem.
Main challenges in a multiobjective optimization are generation of good quality solutions,
generation of uniformly distributed Pareto set, maximizing the diversity of the developed Pareto set,
selection of best compromise solution from the Pareto set, computational efficiency, etc. Several
methods have been developed to solve mutli-objective optimization problems. By way of example, the
penalty function method [1], weighted sum method [2], "-constrained method [3], non-dominated
sorting genetic algorithm (NSGA) based approach [4], etc. have been used for solving various multi-
objective optimization problems.
However, these methods have difficulties. For example, in penalty function method choosing the
proper penalty factors is a difficult task [5]. Weighted sum approach combines all the objectives to a
single objective by using weight factors. This formulation may loose the significance of the objective
function and, moreover, there is no rational basis of determining the weight factors for non-
commensurate objectives [4]. " constrained method avoids the use of weight factors for multiple
objectives but the difficulty with this method is that it requires repeated run for each relaxed level of " to
get the Pareto-optimal set [4]. NSGA method is very sensitive to fitness sharing factor [6]. Moreover,

*Correspondence to: J. Hazra, IBM India Research Lab, Bangalore, India.

y
E-mail: [email protected]

Copyright # 2010 John Wiley & Sons, Ltd.

 J. HAZRA AND A. K. SINHA

these methods have not addressed the problem of diversity preserving, maximizing the diversity,
selection of best compromise solution, etc.
This paper emphasizes on the development of a multi-objective optimal power flow technique using
particle swarm optimization method which has overcome some of the above mentioned difficulties.
Instead of combining all the objectives to a single objective, objectives are solved simultaneously and
hence retain the significance of each objective. To maintain diversity among the Pareto-optimal
solutions a diversity preserving technique is proposed. To obtain diversification special care is taken in
the selection process. Special care is also taken care to prevent non-dominated solutions from being
lost. A fuzzy satisfying method is also proposed to help in choosing a compromise solution from
the set of Pareto-optimal solutions. A power flow method which considers realistic situations such as
load characteristics, generator regulation, on-load tap adjustment, etc. is used for solving the OPF
problem. Proposed method has been tested on IEEE 30 and 118 bus systems and its effectiveness is
presented.
The rest of this paper is organized as follows. Section 2 presents the OPF formulation and Section 3
presents the brief overview of the power flow method used. Section 4 describes the PSO technique used
in this paper. Section 5 introduces the Fuzzy membership function for choosing compromise Pareto-
optimal solution. OPF solution strategy is presented in Section 6. Section 7 presents the simulation
results whereas Section 8 concludes the proposed work.

2. PROBLEM FORMULATION

Let fi ð~
xÞ, i ¼ 1; . . .; m be m objective functions defined over n dimensional search space. A multi-
objective optimization problem can then be formulated as:
minimize fi ð~
xÞ ¼ ff1 ð~
xÞ; f2 ð~
xÞ; . . .; fm ð~
xÞg (1)
subjected to the constraints. This will give a set of Pareto-optimal solutions. A decision vector, x (a set
of control parameters) is said to be Pareto optimal, if there is no other decision vector, y dominating x
with respect to the set of objective functions. The decision vector x is said to strictly dominate another
vector y (denoted by ~ x
~ y) if:
fi ð~
xÞ  fi ð~
yÞ 8i ¼ 1; . . .; m (2)
and
fi ð~
xÞ < fi ð~
yÞ for at least one i: (3)

Though primary objective of OPF is to minimize the generation cost, with the increase in
environmental awareness and after the Clean Air Act Amendments (Kyoto Protocol) in 1990, operating
at minimum cost is no longer the sufficient criterion for dispatching electric power. Minimization of
polluted gases such as sulfur dioxide (SO2), carbon dioxide (CO2), nitrogen oxide (NOx) etc. is also
becoming mandatory for the generation utilities. Therefore, in this paper OPF is formulated as two
objectives optimization problem as follows.

2.1. Minimization of cost of generation

The generator cost function for any turbo generator is obtained from data points taken during ‘‘heat
run’’ tests. Using these data points generator cost curves are approximately fitted by quadratic
functions. However, for large turbo generators this approximation does not hold good. For large turbo
generators, just before opening a valve, a turbine operates at maximum efficiency at that loading.
However, as soon as a valve opens turbine operates at less efficiency because of the throttling of the
steam passing through the throttled control valve. To model the effects of valve-points, a recurring
rectified sinusoid contribution was proposed to add with the quadratic input-output equation as

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

follows [7]:
X
NG
f1 ¼ Fgi
i¼1
(4)
X
NG
¼ ðpi þ qi Pgi þ ri P2gi Þ þ jei  sinðfi  ðPgi Pi;min ÞÞj
i¼1

where f1 is the total fuel cost; Fgi is the generation cost of generator i; NG is the number of participating
generators; Pgi is the generation of ith generator; Pi;min is the minimum generation of ith generator;
pi ; qi ; ri are cost coefficients of generator i; and ei ; fi are coefficients of generator i reflecting valve point
loading effect.

2.2. Minimization of polluted gas emission

Like generator cost curve, emission curve is obtained using curve fitting from practical emission data
of thermal plant. The amount of emission from a fossil-based thermal generator unit depends on
the amount of power generated by the unit. Total emission generated is usually approximated as sum
of a quadratic function and an exponential function of the active power output of the generators as
follows [4,8]:
X
NG
f2 ¼ 102 ðai þ bi Pgi þ g i P2gi Þ þ ji expðli Pgi Þ (5)
i¼1

where, a, b, g, j, and l are emission coefficients.

Subjected to the constraints of load and generation balance and operating limits of each equipment.

3. POWER FLOW MODEL

3.1. Load characteristics

In general majority of the loads are frequency and voltage dependent. To take care of voltage and
frequency dependence, both real and reactive power loads are modeled as follows [9]:

Pdi ¼ Pdio ð1 þ ki Df Þðai þ bi jVi j þ ci jVi j2 Þ (6)

Qdi ¼ Qdoi ð1 þ ki0 Df Þða0i þ b0i jVi j þ c0i jVi j2 Þ (7)

where Pdi ; Qdi are actual active and reactive load at bus i; Pdio ; Qdoi are active and reactive load at bus i
for nominal voltage and nominal frequency; Df is the system frequency deviation from normal
frequency; jVi j is the bus voltage magnitude at bus i; ki ; ai ; bi ; ci are coefficients of active load at bus
i with ai þ bi þ ci ¼ 1; and ki0 ; a0i ; b0i ; c0i are coefficients of reactive load at bus i with a0i þ b0i þ c0i ¼ 1.
All quantities are in per unit.

3.2. Generator characteristics

Real power output of any generator with free governor operation is adjusted by the governor droop
characteristics. This can be expressed as:
Xngi ngi

X 
Pmaxik
Pgi ¼ Pgik ¼ Psetik  Df (8)
k¼1 k¼1
Rik

Pminik Pgik Pmaxik (9)

where ngi is the number of generators at bus i; Pgi is the total active power generation at bus i; Pgik is the
active power generation of kth unit at bus i; Psetik is the set value of active power generation of kth unit

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

at bus i; Pmaxik ; Pminik are maximum and minimum generations of kth generator at bus i; and Rik is the
regulation constant of kth generator at bus i.
All quantities are in per unit.

3.3. Real power flow equations

Static real power flow equations considering load and generation characteristics are derived as
follows [9]:
2 0 3 2 3
DP1 P
ng1
6 jV j 7 6 B11 B12 : : Pmax1k
þ b1 k1 Pdo1 72 Dd 3
6 1 7 6
6 7 k¼1 R1k  jV1 j 7 1
6 DP20 7 6 76 7
6 7 6 P Pmax2k
ng2 76 Dd 7
6 jV2 j 7 6 B21 B22 : : 76
þ b2 k2 Pdo2 76 2 7
6 7 6 k¼1 R2k  jV2 j
7
6 : 7¼6 6
76
76 :
7
7 (10)
6 7 : : : : :
6 : 7 6 76 7
6 7 6 : : : : : 76 : 7
6 7 6 74 5
6 DP 0 7 6 7
4 n5 4 P
ngn 5 DðDðf ÞÞ
Bn1 Bn2 : : Pmaxnk
þ bn kn Pdon
jVn j k¼1 Rnk  jVn j

3.4. Reactive power flow equations

Static reactive power flow equations considering load and generation characteristics are derived as
follows [9]:
2 0 3
DQ1 2 0
32 3
6 jV j 7 B þ 2c Q B : : B DV1
6 1 7 6 11 1 do1 12 1n
76 7
6 0 7
6 DQ2 7 6 0 76 7
6 7 66 B 21 B 22 þ 2c Qdo2 : : B 2n 7
7
6
6 DV2 7
7
6 jV2 j 7 6 2
76 7
6 : 7 6 : : : : : 7 6 7
6 7 6 76 : 7
6 7¼6 : : : : : 7 6 7 (11)
6 : 7 6 76 7
6 7 6 76 7
6 7
6 DQ 0 7 6 6 B B : : B þ 2c
0
Q 76
7 6 :
7
7
6 n7 4
n1 n2 nn n don
5 4 5
6 7
4 jVn j 5
DVn

4. PARTICLE SWARM OPTIMIZATION (PSO)

Particle Swarm Optimization (PSO) is a simple and efficient population based optimization method
proposed by Kennedy and Eberhart [10]. In PSO, potential solutions are called particles and population
of particles is called swarm. Each particle in a swarm flies in the search space toward the optimum or a
quasi-optimum solution based on its own experience and experiences of nearby particles.
Let us define search space, S in n dimension and the swarm consists of N particles. Each particle i,
has its position defined by xit ¼ fxi1;t ; xi2;t ; . . .; xin;t g and a velocity defined by vit ¼ fvi1;t ; vi2;t ; . . .; xin;t g in
variable space S at time t. Position and velocity of each particle changes with time (generation).
Velocity and position of each particle in the next generation (time step) can be calculated as:
vij;tþ1 ¼ w  vij;t þ c1  randðÞ  ðpij;t xij;t Þ þ c2  randðÞ  ðpgj;t xij;t Þ (12)

xij;tþ1 ¼ xij;t þ vij;tþ1 8i ¼ 1; . . .; N (13)

where N is the number of particles in the swarm; n is the dimension of solution vector; w is the inertia
weight; t is the generation number; c1 ; c2 are acceleration constant; randðÞis the uniform random

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

number in the range [0,1]; vij;t is the velocity of particle i at generation t; xij;t is the position of particle i
at generation t; pgj;t is the global best at generation t; pij;t is the best position that particle i could find so
far.
In Equation (12), global best pgj;t conceptually connects all the particles in the population to one
another, so that each particle is influenced by the very best performance of the entire population.
It encourages the particles to move toward itself, and particles move in its direction.

4.1. Initialization
PSO is initialized with a group of random particles (solutions) and then searches for optima by
updating particles. In a PSO, each particle is a k-dimensional real-valued vector, where k is the number
of the control variables. Each control variable of any particle i is initialized randomly within its range
½xmin
j ; xj
max
 as follows:

xij;t¼0 ¼ xmin
j þ randðÞ  ½xmax
j xmin
j  for j ¼ 1 to k (14)

At initial step (t ¼ 0) as there is no idea about particles velocity, at t ¼ 0, velocities of the particles are
initialized with their initial positions as follows:
vij;t¼0 ¼ xij;t¼0 for j ¼ 1 to k (15)

4.2. Limiting velocity and position

The particle swarm algorithm proceeds by modifying the distance that each particle moves on each
dimension per iteration. Changes in the velocity are stochastic and can expand into wider and wider
cycles through the problem space, eventually approaching infinity. Commonly used method is to limit
the velocity within its maximum value (Vmax ) as follows:
if vij;t >Vmax ; vij;t ¼ Vmax (16)

if vij;t <Vmax ; vij;t ¼ Vmax (17)

Similarly, positions may be limited as follows:

if xij;t >xmax
j ; xij;t ¼ xmax
j (18)

if vij;t <xmax
j ; vij;t ¼ xmax
j (19)

Clerc and Kennedy [11] have suggested constriction factor approach to limit the velocity oscillation.
When constriction factor is used in PSO, velocity equation is modified as follows:
h i
vij;tþ1 ¼ j vij;t þ c1  randðÞ  ðPij;t xij;t Þ þ c2  randðÞ  ðPgj;t xij;t Þ (20)

where
2k
j¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi k"½0; 1; ’ ¼ c1 þ c2 ; ’>4 (21)
j2’ ’2 4’j

Where parameter k allows the user to control the degree of convergence of the PSO algorithm [11].

4.3. Parameter selection

4.3.1. Inertia weight (w). Suitable selection of weight factor w helps in quick convergence. A large
weight factor facilitates global exploration (i.e., searching of new areas) while small weight factor
facilitates local exploration. Therefore, it is convenient to choose large weight factor for initial

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

iterations and gradually smaller weight factor for successive iterations. Following function is used for
dynamic inertia weight:

wi wf
w ¼ wi   iter (22)
itermax

where wi is the initial weight factor; wf is the final weight factor; itermax is the maximum number of
iterations allowed.

4.3.2. Constants ðc1 ; c2 Þ. Acceleration constant c1 called cognitive parameter pulls each particle
toward local best position whereas constant c2 called social parameter pulls the particle toward global
best position. Usually values of c1 and c2 lies between 0 and 4.

4.3.3. Maximum velocity (Vmax ). If velocity limit (Vmax ) is very low particle may not explore
sufficiently and on the other hand if velocity limit (Vmax ) is very high it may oscillate about optimal
solution. In this paper, a velocity controlling scheme is used which helps in better convergence. Instead
of fixing the maximum velocity, it is decreased linearly with the increase in generation as follows:

vi vf
vmax ¼ vi   iter (23)
itermax

where vmax is the maximum velocity; vi is the initial velocity; vf is the final velocity; and itermax is the
maximum number of iterations allowed.
This helps to maintain balance between exploration and exploitation and ensure convergence of the
algorithm.

4.4. Finding best local guide

In case of multi objective optimization, there may not exist a single best solution and instead of a single
solution it provides a set of Pareto-optimal solutions. Pareto-optimal solutions obtained in each
iteration are stored in an archive jAj and the archive is updated in each iteration to make it domination
free. All the Pareto-optimal solutions in the archive are equally good. Therefore, in the proposed
technique, best local guide (Pgj;t ) for each particle is identified from the archive using sigma method [12]
as presented in Figure 1. In sigma method, each particle is assigned a s value with coordinate say (f1 ; f2 )
for two objectives so that all the points which are on the line f2 ¼ af1 have same s value. Therefore, for
two objectives s can be calculated as:

s ¼ ðf12 f22 Þ=ðf12 þ f22 Þ (24)

f2
non-dominated solution
particle
=0
ma
sig
sigma=-1

f1
sigma=1

Figure 1. Finding best local guide using sigma method.

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

where f1 and f2 are the fitness values for objective 1 and objective 2, respectively. For m objective
functions s is calculated as below:
2 3
f12 f22
6 7
6 f 2 f 2 7
6 2 37
6 7
6 7
s ¼ 6 : 7=ðf12 þ f22 þ . . . þ fm2 Þ (25)
6 : 7
6 7
6 7
4 2 25
fm f1

When two s values are close to each other, two particles are on two lines which are close to each other.
To find the best local guide for any particle (k), sigma value of that particle (k) is compared with the
sigma values of all the archive members and the best matched archive member ( j) is selected as best
local guide for that particle (k).

4.5. Diversity preserving

Multi-objective optimization techniques finally provide a set of Pareto-optimal solutions to the
Decision Maker (DM) while optimizing a set of conflicting objectives. Sometimes these Pareto-
optimal solutions may be concentrated in a part of the search space which may overlook more efficient
solutions in the other parts of the search space. Hence, DM always prefers a well-distributed set of
solutions throughout the search space. Therefore, in multi-objective optimization, preserving diversity
among Pareto-optimal solutions is an important issue.
In the proposed diversity preserving algorithm, in each PSO iteration non-dominated solutions are
copied to archive. During iterations, fitness values of each particle of population is compared with the
fitness values of the archive members. Three situations may arise i.e., (i) particle may be dominated
by some archive members (ii) particle may be non-dominated with respect to all archive members
and (iii) particle may be non-dominated with respect to some archive members and dominate other
members. In first case, algorithm does not consider the particle to enter into archive. In case 2, particle
is included into the archive and in case 3, particle is included into the archive and at the same time
all the archive members dominated by the particle are removed from archive. As non-dominated
solutions are copied to archive in iteration, the size of archive increases rapidly which increases
the computational burden. To maintain a well distributed fixed size archive, a diversity preserving
technique is proposed based on crowding distances of each solution.
In the proposed approach, solutions are sorted according to each objective value and for any solution
i, crowding distance with respect to objective j, is defined as crwd½j½i ¼ minðdi1;i ; di;iþ1 Þ where,
d represents Euclidian distance between two solutions. Overall crowding for any solution i is defined as the
cumulative sum of individual crowding of i corresponding to all the objectives. To preserve the end
solutions, large crowding values are assigned to them. To preserve diversity, in each iteration solutions
with least crowding value is removed. The complete diversity preserving algorithm is shown in Table I.

5. SELECTION OF COMPROMISE SOLUTION

Multi-objective optimization provides a set of Pareto-optimal solutions. However, decisions with

Pareto optimality or efficiency are not uniquely determined, the final decision must be selected from
among the set of Pareto-optimal solutions. Consequently, the aim in solving multi-objective
optimization problems is to derive a compromise or satisficing solution of a decision maker which is
also Pareto-optimal based on subjective value judgements. Therefore, fuzzy set theory may be useful in
this decision-making problem.
By considering the vague nature of human judgements, it is quite natural to assume that the decision
maker (DM) may have a fuzzy goal for each of the objectives. These fuzzy goals can be quantified by

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

Table I. Proposed diversity preserving technique.

l ¼ jAj Number of solutions in archive
for p ¼ 1 to N for each particle
for y ¼ 1 to l for each archive member
if fi ð~
x p Þ<fi ð~x y Þ for at least one i if the particle is non-dominated
and fi ð~ x Þ> fi ð~
p
x y Þ for at least one i copy the w.r.t. any archive member
particle to archive, l ¼ l þ 1
if fi ð~
x p Þ fi ð~
x yÞ 8 i if the particle dominates
and fi ð~ x Þ< fi ð~
p
x y Þ for at least one i delete archive any archive member y
member y, l ¼ l1
if (l>archive size)
for i ¼ 1 to l, set crwd½i ¼ 0 initialize distances
for j ¼ 1 to m for each objective
A ¼ sort(A, m) sort using each objective value
crwd½j½1 ¼ crwd½j½l ¼ 1 end solutions are preserved
for i ¼ 2 to (l  1) crwd½j½i ¼ minðdi1;i ; di;iþ1 Þ for all other solutions d represents
Euclidian distance
for i ¼ 1 to l
P
crwd½i ¼ m j¼1 ½crwd½j½i
delete archive members with least
crowding values and maintain the archive size

eliciting the corresponding linear membership functions through the interaction with the DM. The best
compromised solution should have the closest distance from the positive ideal solution and the farthest
distance from the negative ideal solution. However, it is not possible in real life situation that one
particular alternative satisfy both these conditions simultaneously. Therefore, a relative importance has
to be given to both of these distances.
To formulate fuzzy membership function, DM is asked to assess a unacceptable value of a objective f
denoted by fimax, and a satisfactory value of f denoted by fimin. With these fuzzy goals, membership
function could be represented by a linear function fimin and fimax as shown in Figure 2. Here membership
value 0 means least satisfaction whereas 1 indicates maximum satisfaction. Mathematically fuzzy
membership function for each objective can be defined as:

1 if fi fimin
fimax fi
mi ¼ if fimin <fi <fimax (26)
fi fimax
max

0 if fi fimax

where mi is the membership value of objective i; fimin is the value of objective i which is completely
satisfactory; fimax is the value of objective i which is completely unsatisfactory.

ui

0 fi
min max
fi fi

Figure 2. Fuzzy membership function.

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

Each Pareto-optimal solution has a fuzzy membership for each objective. Overall membership value
of a Pareto-optimal solution is defined by summing the membership values for all the objectives.
Normalized membership function for each solution is defined as:

P
N obj
mki
i¼1
m ¼
k
(27)
P
M NP obj
mki
k¼1 i¼1

where Nobj is the number of objective functions; M is number of Pareto-optimal solutions; and mk is
membership value of non-dominated solution k.
The Pareto-optimal solution that attains the maximum membership mk is chosen as the best
compromise solution.

6. SOLUTION STRATEGY

In the proposed method, objectives are optimized by controlling generators active power
generations, generators voltage magnitudes and transformers tap positions. In order to take care of
the effect of automatic controllers, in each iteration of the PSO algorithm load-generation balance is
maintained after updating particles velocity and position. For each particle, total generation is
calculated by summing all the updated control variables corresponding to each generator. Effective
load at each bus is calculated using Equations (6) and (7). Total system load is calculated by
summing all the individual bus loads. Now PSO control variables corresponding to each generator
are adjusted to balance the total system load and loss (typically 3% of the system total load) as
follows:
(1) initialize PSO population and archive,
(2) update velocity and position of each particle,
(3) calculate total system load, totload ¼ Snbus
i¼1 ,
(4) calculate total generation capacity, totpmax ¼ Sngen
i¼1 Pmaxi ,
(5) for particle calculate total generation, totgen ¼ Sngen
i¼1 Pgi ,
(6) diff ¼ totgenðtotload þ lossÞ,
(7) if diff<0:001 go to step 10
(8) for i ¼ 1 to ngen
Pgi ¼ Pgi þ diff  Pmaxi =totpmax
if Pgi >Pmaxi Pgi ¼ Pmaxi if Pgi <Pmini Pgi ¼ Pmini
(9) go to step 5,
(10) solve Equations (10) and (11),
(11) obtained generations, voltage magnitudes, and transformer taps are assigned as particle’s latest
position,
(12) repeat the same procedure for all the particles,
(13) update archive using proposed diversity preserving technique,
(14) go to step 2 for PSO next iteration,
(15) stop simulation when converges.

Evolutionary optimizations are unconstrained optimization techniques. However, OPF is associated

with some equality and inequality constraints. In this paper a dynamic penalty function method is used
for constraint handling. A penalty proportional to the extent of constraint violation is added to the
objective function value if any violation occurs. The penalty function used is given by

FðxÞ ¼ f ðxÞ þ hx HðxÞ (28)

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

Read input data

Select the objectives to be optimized

Select control variables

Select PSO control variables

Iteration t = 0

Initialize Population, Pt

Evaluate population, Pt

Send non-dominated particles ( w.r.t.

to archive members) to the archive
Modify velocity and
position of each particle
Number of members in
archive is more?
No
Yes
Truncate Archive iteratively
using crowding distance

No
t > tmax t = t +1

Yes
Compute fuzzy membership of
each member in the archive

Select the solution with highest

fuzzy membership value as
best compromised solution

Figure 3. Computational flow of the proposed OPF.

where f ðxÞ is the objective function of the constrained problem; hx is the penalty value which is a
function of iteration number; HðxÞ is the penalty factor and is given by

X
n2 X
n1
HðxÞ ¼ ðmax½0; gi ðxÞÞ2 þ ðhj ðxÞÞ2 (29)
i¼1 j¼1

where gi ðxÞ<0 are the n2 inequality constraints and hj ðxÞ ¼ 0 are the n1 equality constraints.
Complete flowchart of the proposed method is shown in Figure 3.

7. SIMULATION RESULTS AND DISCUSSIONS

7.1. Test system details

To verify the effectiveness of the proposed method, simulations were carried out on IEEE 30 and 118
bus systems. For IEEE 30 bus system minimum and maximum generation limits and transmission line
capacities were taken from Reference [13]. Minimum and maximum generation limits for 118 bus
were taken from Reference [14]. Due to unavailability of line capacity data for 118 bus system,
following data were used: 150 MVA for voltage level V1, 250 MVA for voltage level V2 and 350 MVA

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

for the transformers. For IEEE 118 bus system negative generations were treated as load. Generator
cost coefficients and emission coefficients used in simulation are given in Appendix-A.

7.2. Convergence characteristics

Evolutionary computation techniques are sensitive to proper selection of the control parameters. These
parameters are selected through experiment. For the simulations, PSO parameters were selected from
Reference [15]. With the selected parameters an experiment was done to optimize the swarm size for
each system. Convergence characteristics with different size of swarm are given in Figure 4. From
Figure 4 it is clear that swarm of 10 and 50 particles for IEEE 30 and 118 bus systems, respectively,
provide satisfactory convergence characteristic.

7.3. Dynamic velocity control of PSO

Like any heuristic method, within limited number of iterations (100) proposed method may not
converge to the global optimal solution at all the time. However, if we run the simulation for large
number of iterations (>1000), it will converge to global solution in each run but computational time
will be very large. Therefore, within limited iterations, performances could not be judged by the results
of a single run. Many trials should be done to reach a useful conclusion about the performance of the
algorithm. Therefore, 100 independent runs were carried to test the effectiveness of the velocity control

Figure 4. Convergence characteristics with swarm size.

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

Table II. Test results for 100 independent trials.

System Proposed CFA PI
best worst avg. best worst avg.
30 bus 531 244 541 362 532 158 531 244 549 579 533 277 64
118 bus 7 120 814 7 236 824 7 130 309 7 124 263 7 350 657 7 141 897 73

effect incorporated into PSO. Best, worst, and average minima obtained by the proposed method are
given in Table II. Table II shows that For IEEE 30 bus system best solution obtained is 531 244 Rs/h and
the average solution obtained is 532 158 Rs/h. Therefore, cost variation band is only 0.17% of best cost.
Similarly, cost variation band for IEEE 118 bus system is 0.13%. These results show cost variation
band is very small.
Proposed method was compared with the commonly used PSO with Constriction Factor Approach
(CFA) [11]. Table II shows that for IEEE 30 bus system best costs obtained by both the methods are
same, however, worst and average costs are better in proposed approach. For IEEE 118 bus system best,
worst and average costs by proposed method are better than CFA approach. Out of 100 independent
runs, proposed method outperforms CFA based approach in 64 and 73 runs for IEEE 30 and 118 bus
systems, respectively. This shows that proposed method can explore better solution than CFA
approach.

7.4. Comparison with other optimization techniques

To show the effectiveness in solving multiple objectives, proposed method was compared with the
multi-objective economic dispatch methods in References [16,17] and [4]. Tables III and IV present

Table III. Best fuel cost and corresponding emission.

Generator LP [16] MOSST [17] NSGA [4] Proposed
Case1 Case2 Case1 Case2
P1 0.15 0.1125 0.1567 0.1168 0.1198 0.117
P2 0.30 0.302 0.287 0.3165 0.31 0.3151
P3 0.55 0.5311 0.4671 0.5441 0.5328 0.6041
P4 1.05 1.0208 1.0467 0.9447 0.9904 0.9725
P5 0.46 0.5311 0.5037 0.5498 0.5197 0.5042
P6 0.35 0.3625 0.3729 0.3964 0.3611 0.354
Best Cost ($/h) 606.31 605.89 600.57 608.24 600.13 607.84
Emission (ton/h) 0.2233 0.2222 0.2228 0.2166 0.2199 0.2192

Table IV. Best emission and corresponding fuel cost.

Generator LP [16] MOSST [17] NSGA [4] Proposed
Case1 Case2 Case1 Case2
P1 0.40 0.4095 0.4394 0.4113 0.3915 0.4087
P2 0.45 0.4626 0.4511 0.4591 0.47 0.4574
P3 0.55 0.5426 0.5105 0.5117 0.5178 0.557
P4 0.40 0.3884 0.3871 0.3724 0.4003 0.4095
P5 0.55 0.5427 0.5553 0.581 0.5434 0.5346
P6 0.50 0.5142 0.4905 0.5304 0.5111 0.4976
Emission(ton/h) 0.1942 0.1942 0.1944 0.1943 0.1943 0.1942
Cost ($/h) 639.6 644.11 639.23 647.25 636.62 642.9

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

results for best fuel cost with corresponding emission and best emission with corresponding fuel cost,
respectively, from a set of Pareto-optimal solutions obtained. In methods [16] and [17], test system was
considered as lossless. To compare, case 1 simulations were done considering the system as lossless
while case 2 simulations were done considering transmission losses, which is more realistic.
In Table III, for case 1, best cost (per hour) obtained by References [16,17] and [4] are $606.31,
$605.89, and $600.57, respectively, and corresponding emissions (ton/h) are 0.2233, 0.2222, and
0.2228, respectively. In this case best cost obtained by proposed method is $600.13 and corresponding
emission is 0.2199, both are better than other methods. In case 2, best cost obtained by proposed
method is little less than the method in Reference [4] but corresponding emission is little higher.
In Table IV, for case 1, best emission (ton/h) obtained by References [16,17] and [4] are 0.1942,
0.1942, 0.1944, respectively, and corresponding costs (per hour) are $639.6, $644.11, and $639.23,
respectively. For this case best emission (0.1943 ton/h) obtained by proposed method is very similar to
other methods but corresponding cost ($636.62) is less compared to other methods. Same is true for
case 2 also. This shows that proposed method can efficiently explore the end Pareto-optimal solutions
in case of multi-objective optimization.

7.5. Diversity preserving

In multi-objective optimization, solution quality can not be judged only seeing the end solutions.
Diversity among Pareto-optimal solutions is a very important issue. To show the effectiveness of the
proposed diversity preserving technique, the method was compared with NSGA-II [18], SPEA2 [19],
and OMOPSO [20]. Figures 5 and 6 show the Pareto fronts obtained by different methods. For IEEE 30
645 645

640 640

635 635

630 630
Cost ($/h)

Cost ($/h)

625 625

620 620

615 615

610 610

605 605
0.19 0.195 0.2 0.205 0.21 0.215 0.22 0.19 0.195 0.2 0.205 0.21 0.215 0.22
Emission (ton/h) Emission (ton/h)

(a) NSGAII (b) SPEA2

645 645

640 640

635 635

630 630
Cost ($/h)

Cost ($/h)

625 625

620 620

615 615

610 610

605 605
0.19 0.195 0.2 0.205 0.21 0.215 0.19 0.195 0.2 0.205 0.21 0.215 0.22
Emission (ton/h) Emission (ton/h)

(c) OMOPSO (d) Proposed

Figure 5. Pareto optimal solutions for IEEE 30 bus system.

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

6 6
x 10 x 10
7.25 7.34

7.24
7.32

7.23

7.3
7.22
Cost (Rs/h)

Cost (Rs/h)
7.21 7.28

7.2
7.26

7.19

7.24
7.18

7.17 7.22
4 4.5 5 5.5 6 6.5 3.5 4 4.5 5
Emission (ton/h) Emission (ton/h)

(a) NSGAII (b) SPEA2

6
x 10
7.7 7.48

7.65
7.43
7.6

7.55
7.38
7.5
Cost (Rs/h)

Cost (Rs/h)

7.45 7.33

7.4
7.28
7.35

7.3
7.23
7.25

7.2 7.18
2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 2.5 3 3.5 4 4.5 5 5.5 6
Emission (ton/h) Emission (ton/h)

(c) OMOPSO (d) Proposed

Figure 6. Pareto optimal solutions for IEEE 118 bus system.

bus system front size was taken as 20 and for IEEE 118 bus system it was taken as 50. For 118 bus
system front is less uniform compared to 30 bus system because of higher nonlinearity. However, for
both the systems Pareto-optimal solutions obtained by the proposed method are more uniformly
distributed throughout the Pareto front compared to other methods.

7.6. OPF solutions

Table V presents end Pareto-optimal solutions and best compromise solution obtained from a single
run for the test systems with two competitive objectives cost and emission. From the simulation results
it is clear that if the operator wants to minimize emission, cost of generation increases and vice versa.

Table V. End solutions in a single run.

System Cost (C) (Rs/h) Emission (E) (ton/h)
30 bus min C 545 556 0.2469
min E 594 248 0.2328
compromise 551 050 0.2450
118 bus min C 7 926 112 6.8020
min E 8 404 232 3.4157
compromise 8 364 959 3.4300

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

Table VI. Optimal setting of control variables for IEEE 30 bus system.
Bus Generation (MW) Voltage mag. (p.u.) Tr. Tap
1 115.19 1.0600 6–9 1.0492
2 69.58 1.0267 6–10 1.0056
5 25.04 0.9889 4–12 1.0272
8 26.74 1.0107 28–27 0.9695
11 27.18 1.0318
13 26.33 1.0279

However operator can also choose a compromise (somewhat higher cost and higher emission from
their lowest optimum values) solution as well.
For IEEE 30 bus system, 16 control variables (6 generator outputs + 6 generator voltages + 4
transformer taps) were optimized. Similarly, for IEEE 118 bus system 117 (54 + 54 + 9) variables were
optimized. Control variables for the best compromise solutions are given in Tables VI and VII for IEEE
30 and 118 bus systems, respectively. Complete OPF solution for IEEE 30 bus system is given in
Table VIII.

7.7. Computational efficiency

Computational efficiency of the proposed method has been compared with NSGA-II [18], SPEA2 [19],
and OMOPSO [20]. Computational complexity of NSGA-II is OðMN 2 Þ as in each iteration, it
calculates set of solutions dominated by solution k as well as the set of solutions dominated the solution
k. Where N is the number of solutions and M is number of objectives. Computational complexity of
OMOPSO and proposed method is also OðMN 2 Þ as non-dominated sorting is required in both

Table VII. Optimal setting of control variables for IEEE 118 bus system.
Bus Generation (MW) Voltage mag. (p.u.) Bus Generation (MW) Voltage mag. (p.u.) Tr. Tap
1 22.19 1.0007 65 160.38 1.0561 8–5 1.0506
4 80.17 1.0431 66 172.77 1.0668 26–25 1.0013
6 60.24 1.0327 69 27.69 1.0177 30–17 1.0105
8 46.31 1.0600 70 85.68 1.0166 38–37 1.0412
10 110.05 1.0600 72 75.43 1.0134 63–59 1.0778
12 151.62 1.0376 73 52.00 1.0403 64–61 1.1000
15 0.00 1.0019 74 59.12 0.9923 65–66 0.9403
18 92.87 1.0157 76 59.86 0.9670 68–69 0.9985
19 70.47 1.0022 77 70.52 0.9768 81–80 0.9757
24 49.28 1.0600 80 150.12 0.9678
25 200.61 1.0332 85 0.00 0.9923
26 64.71 1.0382 87 63.23 1.0600
27 54.75 1.0220 89 163.99 1.0229
31 79.18 1.0272 90 97.08 0.9629
32 39.15 1.0164 91 43.56 1.0125
34 82.88 0.9860 92 97.08 1.0117
36 87.69 0.9843 99 10.04 0.9655
40 68.69 0.9500 100 156.74 1.0155
42 37.97 1.0405 103 41.41 1.0284
46 69.33 1.0462 104 61.98 1.0177
49 119.68 1.0405 105 65.66 1.0181
54 143.68 1.0366 107 65.63 1.0181
55 72.05 1.0224 110 18.68 1.0415
56 75.74 1.0269 111 84.98 1.0600
59 71.99 0.9759 112 97.08 1.0453
61 124.52 1.0166 113 47.87 1.0085
62 97.08 1.0184 116 97.08 1.0180

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

Table VIII. Power flow solutions for IEEE 30-Bus test system.
Bus code Pg (MW) Qg (MVAR) Volt. (p.u.) Delta (deg.)
1 115.19 43.39 1.06 0
2 69.58 27.64 1.0267 1.76
3 0 0 1.0274 3.59
4 0 0 1.0191 4.38
5 25.04 13.95 0.9889 7.84
6 0 0 1.0118 5.25
7 0 0 0.9946 6.86
8 26.74 27.87 1.0107 5.39
9 0 0 0.9936 6.08
10 0 0 0.9904 8.25
11 27.18 19.71 1.0318 2.92
12 0 0 0.9991 7.18
13 26.33 21.58 1.0279 5.12
14 0 0 0.9846 8.22
15 0 0 0.9806 8.39
16 0 0 0.9872 7.93
17 0 0 0.9842 8.39
18 0 0 0.9712 9.12
19 0 0 0.9689 9.34
20 0 0 0.9735 9.13
21 0 0 0.9785 8.79
22 0 0 0.9795 8.78
23 0 0 0.973 8.99
24 0 0 0.9719 9.41
25 0 0 0.9834 9.65
26 0 0 0.9651 10.1
27 0 0 0.9997 9.5
28 0 0 1.0057 5.68
29 0 0 0.9793 10.79
30 0 0 0.9675 11.72

Table IX. Comparison of OPF execution time.

Test system Execution time (s)
NSGA-II [18] SPEA2 [19] OMOPSO [20] Proposed
30 bus 1.136 6.028 1.021 0.849
118 bus 53.14 66.87 51.81 40.08

approaches. Computational complexity of SPEA2 is OðMN 2 logNÞ as in each iteration it calculates near
neighbors for each solution.
Table IX illustrates a comparison of computational time to solve the OPF problem using different
methods. The simulations were carried out using a P4 3.0 GHz, 512 MB RAM machine. Though
computational complexities of NSGA-II, OMOPSO and proposed method are same, NSGA-II takes
more time because PSO based approaches (OMOPSO and proposed method) converges quickly
compared to GA (NSGA-II) based approach. OMOPSO approach takes little more time compared to
proposed method because of incorporation of mutation operation in the PSO algorithm. SPEA2 takes
largest computational time because in each iteration, it calculates inter-distance between all
individuals.

8. CONCLUSIONS

This paper proposes a optimal power flow technique with two competitive objectives, cost of
generation, and emission of thermal plants. A multi-objective particle swarm optimization technique

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 A MULTI-OBJECTIVE OPTIMAL POWER FLOW

with dynamic velocity controlling is used to solve this optimization problem. To maintain diversity
among Pareto-optimal solutions a diversity preserving technique is proposed. Simulation results
show that PSO with dynamic velocity control ensure good convergence and explores better
solutions than PSO with Constriction Factor Approach (CFA). Simulation results also show that
proposed method is computationally efficient and generates a set of uniformly distributed set of
Pareto-optimal solutions in single run. Fuzzy approach is useful in choosing the best compromise
solution from the set of Pareto-optimal solutions. This OPF method can optimize any number of
objectives and may be helpful to the system operators in choosing judicious decision in running the
system efficiently.

9. LIST OF ABBREVIATIONS AND SYMBOLS

NG Number of generator.
Pgi Active power generation of ith generator.
Pi;min Minimum generation limit of ith generator.
Pdi Active power load at bus i.
Qdi Reactive power load at bus i.
Pdio Active power load at bus i for nominal voltage and frequency.
Qdio Reactive power load at bus i for nominal voltage and frequency.
Vi Voltage magnitude of ith bus.
di Voltage phase angle of ith bus.
Df Change in voltage magnitude.
ngi Number of generator at bus i.
Psetik Set active power of kth generator unit at bus i.
Pmaxik Maximum generation of kth generator unit at bus i.
Pminik Minimum generation of kth generator unit at bus i.
Rik Regulation constant of kth generator unit at bus i.
crwd Crowding distance.
xij;t Position of jth element of particle i at generation t.
vij;t Velocity of jth element of particle i at generation t.
mi Fuzzy membership value of objective i.
mk Fuzzy membership value of non-dominated solution k.

APPENDIX A

Table X Generators cost coefficients.

Unit (MW) p (Rs/h) q (Rs/MWh) r (Rs/MW2h) e  104 (Rs/h) f (1/MW)

25 0 2025 1.50 0.0 0.0
26–50 0 1875 1.425 1.2 0.126
51–100 0 1800 1.35 2.4 0.063
101–200 0 1650 1.25 3.24 0.047
201–250 0 1575 1.50 3.0 0.05
251–300 0 1575 1.25 3.6 0.042
301–350 0 1500 1.35 3.35 0.045
351–400 0 1500 1.25 3.85 0.039
401–500 0 1200 1.50 4.0 0.038
>500 0 1200 1.00 4.4 0.034

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep
 J. HAZRA AND A. K. SINHA

Table XI Generators emission coefficients.

Unit (MW) a (lb/h) b (lb/MWh) g (lb/MW2h) j  105 (lb/h) l (1/MW)

30 6.131 5.555 5.151 1 6.667
31–50 4.258 5.094 4.586 1 8.000
51–100 5.326 3.550 3.380 2 2.000
101–200 4.258 5.094 4.586 1 8.000
201–300 2.543 6.047 5.638 5 3.333
>300 4.091 5.554 6.490 2 2.857

REFERENCES
1. Arya LD, Choube SC, Kothari DP. Emission constrained secure economic dispatch. International Journal of Electric
Power Energy Systems 1997; 19(4):279–285.
2. Rosehart WD, Canizares CA, Quintana VH. Multiobjective optimal power flows to evaluate voltage security costs in
power networks. IEEE Transactions on Power Systems 2003; 18(2):578–587.
3. Yokoyama R, Bae SH, Morita T, Sasaki H. Multiobjective optimal generation dispatch based on probability security
criteria. IEEE Transactions on Power Systems 1988; 3(1):317–324.
4. Abido MA. A novel multiobjective evolutionary algorithm for environmental/economic power dispatch. Electric
Power Systems Research 2003; 65(1):71–81.
5. Ramesh VC, Li X. A fuzzy multiobjective approach to contingency constrained opf. IEEE Transactions on Power
Systems 1997; 12(3):1348–1354.
6. Coello CA. An updated survey of ga-based multiobjective optimization techniques. ACM Computing Surveys
(CSUR) 2000; 32(2):109–143.
7. Walters DC, Sheble GB. Genetic algorithm solution of economic dispatch with valve point loading. IEEE
Transactions on Power Systems 1993; 8(3):1325–1332.
8. Abido MA. Multiobjective evolutionary algorithms for electric power dispatch problem. IEEE Transactions on
Evolutionary Computation 2006; 10(3):315–329.
9. Hazra J, Sinha AK. A new power flow model incorporating effects of automatic controllers. WSEAS Transactions on
Power Systems 2007; 8(2):202–207.
10. Kennedy J, Eberhart R. Particle swarm optimization. Proceedings of IEEE Intlernational Conference on Neural
Networks, Perth, Australia, 1995; 1942–1948.
11. Clerc M, Kennedy J. The particle swarm - explosion, stability, and convergence in a multidimensional complex
space. IEEE Transactions on Evolutionary Computation 2002; 6(1):58–73.
12. Mostaghim S, Teich J. Strategies for finding good local guides in multi-objective particle swarm optimization
(mopso). Proceedings of IEEE Swarm Intelligence Symposium, 2003; 26–33.
13. Alsac O, Stott B. Optimal load flow with steady-state security. IEEE Transactions on Power Apparatus Systems
1974; PAS-93(3):745–751.
14. http://www.pserc.cornell.edu/matpower/
15. Hazra J, Sinha AK. Congestion management using multiobjective particle swarm optimization. IEEE Transactions
on Power Systems 2007; 22(4):1726–1734.
16. Farag A, Al-Baiyat S, Cheng TC. Economic load dispatch multiobjective optimization procedures using linear
programming techniques. IEEE Transactions on Power Systems 1995; 10(2):731–738.
17. Das DB, Patvardhan C. New multi-objective stochastic search technique for economic load dispatch. IEE
Proceedings of General Transmission Distribution 1998; 145(6):747–752.
18. Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE
Transactions on Evolutionary Computation 2002; 6(2):182–197.
19. Zitzler E, Laumanns M, Thiele T. SPE A2: Improving the Strength Pareto Evolutionary Algorithm. Technical Report
103, Swiss Federal Institute of Tehnology (ETH) Zurich, Gloriastrasse 35, CH-8092 Zurich, Switzerland Sept 2001.
20. Sierra MR, Coello Coello CA. Evolutionary Multi-Criterion Optimization, Lecture Notes in Computer Science, vol.
3410/2005, Ch. Improving PSO-based multi-objective optimization using crowding, mutation and e-dominance.
Springer-Verlag: Berlin Heidelberg, 2005; 505–519.

Copyright # 2010 John Wiley & Sons, Ltd. Euro. Trans. Electr. Power (2010)
DOI: 10.1002/etep