www.ietdl.

org
Published in IET Generation, Transmission & Distribution
Received on 14th November 2010
Revised on 5th February 2012
doi: 10.1049/iet-gtd.2011.0559

ISSN 1751-8687

Comparative study of evolutionary computation

methods for active reactive power dispatch
C.-M. Huang1 S.-J. Chen1 Y.-C. Huang2 H.-T. Yang3
1

Department of Electrical Engineering, Kun Shan University, Tainan 710, Taiwan

Department of Electrical Engineering, Cheng Shiu University, Kaohsiung 833, Taiwan
3
Department of Electrical Engineering, Cheng Kung University, Tainan 701, Taiwan
E-mail: h7440@ms21.hinet.net
2

Abstract: This study presents a comparative study for four evolutionary computation (EC) methods to the optimal active
reactive power dispatch (ARPD) problem. Theoretically, there is a coupling relation between ARPDs. However, because of
high X/R ratio existing in the transmission line, the problem of ARPD can be decomposed into two individual sub-problems
by the decoupling concept, that is, ARPD problems. In this study, the evolutionary programming (EP), particle swarm
optimisation (PSO), differential evolution (DE) and the proposed hybrid differential evolution (HDE) algorithms are used to
separately solve the ARPD problem. To evaluate the performance of each method, the IEEE 30-bus and Taiwan Power
Company (TPC) 345 kV simplied systems are employed as the study cases. The results indicate that the proposed HDE can
obtain better results than the other methods in terms of active power transmission losses, voltage deviation, operating cost and
convergence performance.

Introduction

In a deregulated and competitive environment, the active

reactive power dispatch (ARPD) will be more and more
important for power utilities than ever before from the
viewpoint of operating cost, environmental protection and
system security. Traditionally, the reactive power dispatch is
dened as the minimisation of active power transmission
losses by controlling the generator terminal voltages,
transformer tap settings and shunt capacitors/reactors. Based
on the minimisation of the fuel cost function, the active
power dispatch schedules the power output of each
generator without considering the pollutant emission.
However, with the increasing concern for environmental
protection, a revised active power dispatch accounting for
emission functions is required.
Theoretically, there is a coupling relation between reactive
and active power dispatches. However, guided by experiences
that reactive power is most sensitive to voltage magnitudes
and active power to nodal phase angles, the ARPD problem
can be decomposed into two individual sub-problems by
the decoupling concept, that is, ARPD problems [1]. To
more simply, rapidly and reliably solve this sort of problem,
a fast-decoupled load ow method is presented in [2]. On
the other hand, a fully coupled active reactive dispatch
using linear programming approach is presented in [3, 4]. A
united approach using gradient projection method to the
optimal ARPD is proposed in [5]. To minimise load
shedding in severe contingences, the active and reactive
re-dispatch method is proposed in [6]. A new optimal
power ow (OPF) technique using the interior point method
636
& The Institution of Engineering and Technology 2012

with goal programming was proposed to optimise ARPD

while maximising voltage security in [7]. Based on the
derivation of P Q loss sensitivities, a formula with optimal
P Q allocation for minimising total system operating cost
is presented in [8].
The techniques mentioned above have served as effective
tools for solving the ARPD problem. In this paper, four
evolutionary computation (EC) methods including
evolutionary
programming
(EP),
particle
swarm
optimisation (PSO), differential evolution (DE) and the
proposed hybrid differential evolution (HDE) algorithms are
separately used to solve the ARPD problem. EP [9] is a
global search technique that simulates the natural evolution
process and constitutes a stochastic optimisation algorithm.
By simulating the behaviour of a swarm as a simplied
social system, PSO [10] has the fast convergence and easy
implementation characteristics. DE was rst introduced by
Storn and Price in 1995 [11]. It has the parallel search and
rapid convergence naturals. Recently, DE has been
successfully applied to solve the reactive power dispatch
[12] and multi-objective power ow [13] problems.
To enhance the global search ability of DE, an HDE
algorithm combining variable scaling mutation (VSM) [14]
and probabilistic state transition rule used in the ant system
[15] was proposed to solve the optimal ARPD problem.
Based on rapid convergence and global search ability, HDE
can offer higher probability of converging towards global
solution than the other methods.
The rest of the paper is organised as follows. In Section 2, the
problem formulation of ARPD is briey reviewed. Section 3
describes the EC methods. The proposed HDE approach to
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

www.ietdl.org
solve the ARPD problem is also presented in this section. In
Section 4, the simulation results of different methods for the
IEEE 30-bus and TPC 345 kV simplied systems are
illustrated. Finally, conclusions are given in Section 5.

Problem formulation

ARPD is typically a multi-objective optimisation problem.

The objectives and constraints of the ARPD solved in this
paper are summarised below.
2.1

Reactive power dispatch

Reactive power dispatch is a sub-problem of the OPF

calculation. Typically, it minimises the active power
transmission loss subject to a number of constraints. In this
paper, the voltage magnitude deviation at each load bus is
also considered an objective function.
2.1.1 Active power transmission loss: The objective of
active power transmission loss can be expressed as follows
[16]
fQ =

Pl,k =

L 
L

i=1 j=1

k[(i,j)

[gij (|Vi | + |Vj |2 2|Vi ||Vj | cos(di dj ))]

2

bus I; QBi is the injected reactive power at bus I; QDi is the

reactive power demand at bus I; Qli is the reactive
transmission loss at bus I; |Yij| is the admittance between
buses i and j; and dij is the admittance phase angle between
buses i and j.
Reactive generation constraint: The reactive generation
constraint is usually used on the generation bus (PV bus) as
follows
QBi, min QBi QBi, max

where QBi,min and QBi,max are the lower and upper limits of
the injected reactive power at bus i.
Bus voltage constraint: To preserve the stable operation of
the power system, each bus voltage is controlled within the
constraints. The bus voltage constraint is generally utilised
on the load bus (PQ bus) as follows
Vi, min |Vi | Vi, max

2.1.2 Voltage magnitude deviation: The bus voltage

magnitude is one of the important indices in security
operation. The objective of voltage magnitude deviation is
given as follows [17]
Vdev =

(Vi Viref )

(2)

i=1

where Viref is the ith specied reference value of the voltage

magnitude for load buses, which is generally set at 1.0 pu
and NPQ is the number of load buses.
The reactive power dispatch of models (1) and (2) must
satisfy the following constraints
Power balance constraint
PBi PDi Pli = 0

(3)

QBi QDi Qli = 0

(4)

QC, min QC QC, max

(9)

Tk, min Tk Tk, max

(10)

where QC is the reactive power installed by capacitor, QC,min

and QC,max are the lower and upper limits of the capacitor, Tk
is the position of transformer k and Tk,min and Tk,max are the
lower and upper positions of transformer k. In this paper,
the capacitor and transformer tap settings are regarded as
continuously adjustable variables.
2.2

NPQ


(8)

where Vi,min and Vi,max are the lower and upper limits of the
bus voltage at bus i.
Capacitor and transformer tap setting constraints: The
constraints of the capacitor and transformer tap setting are
dened as follows

(1)

where fQ is the active power transmission loss; Pl,k is the

active power transmission loss of branch k; L is the number
of transmission lines; |Vi| is the voltage magnitude at bus I;
gij is the conductance between bus i and j; and di is the
voltage phase angle of bus i.

(7)

Active power dispatch

Traditionally, the active power dispatch minimises fuel cost

subject to some constraints. However, with the increasing
concern for environment protection, a revised active power
dispatch accounting for emission functions is required. In
this paper, three objectives including fuel cost, power
wheeling cost and pollutant emission are considered.
2.2.1 Fuel cost: The fuel cost function of the system can be
expressed as the quadratic function of the generators active
power output as follows [18]
F(PG ) =

N


2
ai + bi PGi + ci PGi

($/h)

(11)

i=1

and
Pli =

L


|Yij | |Vi | |Vj | cos(fij + dj di )

(5)

|Yij | |Vi | |Vj | sin(fij + dj di )

(6)

j=1

Qli =

L

j=1

where PBi is the injected active power at bus I; PDi is the

power demand at bus I; Pli is the active transmission loss at
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

where F(PG) is the total fuel cost of the system; PGi is the
power output of the ith unit; N indicates the number of
generators; ai , bi and ci are the cost coefcients that are
generally obtained by the curve-tting technique [19].
2.2.2 Power wheeling cost: The evaluation of power
wheeling cost can roughly be divided into four methods:
the postage stamp method, contract path method, megawatt
mile method and marginal cost method [20]. In this paper,
the megawatt mile method is adopted and can be expressed
637

& The Institution of Engineering and Technology 2012

www.ietdl.org
as follows
Ch (PG ) =

4k |Pk (h)| Lk

(12)

all k

where Ch(PG) is the power wheeling cost of the hth exchange;

4k is the weighting value of the kth transmission line; | Pk(h)|
is the absolute value of the active power ow for the kth
transmission line on the hth exchange; and Lk is the length
of the kth transmission line.
The active power ow Pk(h) in (12) can be calculated using
the DC power ow method [21] as follows
Pk (h) =

1
[Duk (h)]
xk

(13)

where xk is the reactance of the kth transmission line, Duk(h)

denotes the variation in voltage phase angle for the kth
transmission line on the hth exchange.
2.2.3 Pollutant emission: The emission function of the
system can be expressed as the polynomial function of the
generator active power output as follows [22]
E(PG ) =

N


2
nk
ai + bi PGi + gi PGi
+ + wi PGi

(ton/h)

i=1

(14)
where a, b, g and w are emission coefcients and nk
represents the order of emission function. The emission
function of (14) may be one of all types of pollutants, such
as NOx , CO2 , SO2 , particulates, or thermal pollutants. The
emissions of NOx and CO2 are separately considered as two
different case studies in this paper.
The optimisation of the active power dispatch must be
subjected to power balance constraints, generation capacity
constraints and line overload prevention constraints, as
dened in (3), (15) and (16), respectively.
PGi, min PGi PGi, max
L


PiC Pi 0

(15)
(16)

Much research has been devoted to dealing with multiobjective power dispatch problem [7, 13, 17, 23, 24, 26].
Among these techniques, the weighted-sum strategy that
converts the conicting objectives into a single-objective
optimisation problem is considered as a simple and efcient
method as follows [23]
F(x) =

Multi-objective power dispatch

The multi-objective power dispatch problem is often

formulated as the minimisation, of several conicting
objectives subject to a number of equality and inequality
constraints as follows

s.t. gj (x) = 0,

EC algorithms

The ECs, including EP, PSO, DE and the proposed HDE are
articial intelligent methods for optimisation based on the
mechanics of natural evolution. These methods share many
similarities. In this section, the EP is introduced rst, and
followed by PSO, DE and the proposed HDE methods.
3.1

EP [9]

The EP simulates the natural evolutionary process to reach the

ttest individuals after repeated offspring creation,
competition and selection procedures as follows.
3.1.1 Offspring creation (or mutation): An offspring
Pi+k is created from each parent Pi by adding a random
variation as follows
Pi+k = Pi + N(m, s2i ),

hk (x) 0,

i = 1, 2, . . . , k

(17)

k = 1, 2, . . . , n

where fi(x) is the ith objective; x is a decision vector; l is the

number of objectives; and m and n are the number of equality
and inequality constraints, respectively.
638
& The Institution of Engineering and Technology 2012

(19)

where N(m, s2i ) represents a vector of Gaussian random

variables with mean m and standard deviation si . si is
given according to the following equation
(20)

where Fi is the objective value of the ith individual; b is a

scaling factor; and V represents an offset.
3.1.2 Competition and selection: Competition for
survival is stochastic. In this process, each individual must
compete with the other randomly selected individuals for
win based on their tness value as follows
Wi =

i = 1, 2, . . . , l
j = 1, 2, . . . , m

(18)

where F(x) is the objective value and wi (Swi=1) is the weight

of the ith objective.
Note that the determination of weighting value in (18) is
often difcult but important to the operators. To help the
operators more easily express their preferences, a weight
assessment technique, that is, the analytical hierarchy
process (AHP) approach [25] is employed in this paper.
Details of the AHP method are explicated in the appendix.

si = bFi + V

where PGi,min and PGi,max are the lower and upper bounds of
the ith power generation, Pi is the active power ow of the ith
line and PiC is the maximal power ow capacity of the ith line.

min fi (x),

wi fi (x)

i=1

i=1

2.3

l


Wt =

1,

0,

Pm


Wt

t=1

Fr
Fr + Fi
otherwise
if u1 ,

(21)

where Wi is the score of the ith individuals; Pm is the number

of competitors; Fr is the objective value of the rth competitor;
and u1 is an uniform random number in the interval of (0,1).
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

www.ietdl.org
After all individuals have undergone competition, the 2k
individuals are then ranked in descending order of their
corresponding Wi value. The rst k individuals are then
selected as new parents of the next generation.
3.2

Particle swarm optimisation (PSO) [10]

PSO is an EC technique which simulates the behaviour of a

swarm as a simplied social system. In the PSO algorithm,
each particle tries to adjust its velocity as follows
vdi (t + 1) = wvdi (t) + a1 rand1 [P bestdi Rdi (t)]
+ a2 rand2 [G bestd Rdi (t)]

(22)

where vdi (t) is the current velocity of the ith particle in the d
dimension, i 1, . . . , P, P is the population size; Pbestdi is
the best previous position of the ith particle; G bestd is the
best previous position of the swarm; Rdi (t) is the current
position of the ith particle; a is an acceleration factor; and
rand represents the uniform random number between 0 and
1, and w is the inertial weight.
With model (22), the next position of the ith particle can be
modied by
Rdi (t + 1) = Rdi (t) + vdi (t + 1)

(23)

As shown in (22) and (23), the convergence speed of each

particle is inuenced by the parameters of a1 and a2 . In
general, a smaller value of ai requires a relatively large
number of iterations, whereas a larger one will lead to the
optimisation process becoming numerically unstable. In this
paper, the value of ai is set according to the trial-and-error
experiments.
3.3

Basic DE and proposed HDE approaches

Based on the basic evolutionary strategies, DE achieves the

ttest individual after repeated initialisation, mutation,
recombination and selection operations. The general scheme
of the basic DE is described in this subsection, where the
proposed HDE is also presented.
3.3.1 Initialisation: Let pi [ pi1 , pi2 , . . ., piM] be a trial
vector representing the ith individual (i 1, 2, . . . , P) of
the population to be evolved, where P is the population size
and M is the dimension of each individual. The elements in
vector pi represent the decision variables (genes), which are
randomly generated as follows
pij = pij, min + s (pij, max pij, min ),

j = 1, 2, . . . , M
(24)

where pij represents the jth gene of the ith individual; pij,min
and pij,max mean the lower and upper bounds of pij ,
respectively; and s represents the uniform random number
between 0 and 1. In this paper, the trial vector pi represents
either the desired values of voltage magnitude at generation
bus, transformer tap settings and shunt capacitors for the
reactive power dispatch or the active power output of each
generator for the active power dispatch.
3.3.2 Mutation: The mutation operation of basic DE is
performed by adding a differential vector to the parent
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

individual as follows
pi = pi + fm (pia pib )

(25)

where pia and pib are the randomly selected individuals in the
parent population, (pia pib ) is a differential vector, and
fm [ [0,1] represents the mutation factor.
As shown in (25), DE uses a xed mutation factor to
increase the diversity of the population. In general, a
smaller mutation factor requires more computational time,
whereas the larger one may result in falling into local
minima. Therefore the selection of a mutation operator is a
very important issue in DE. In this paper, the VSM based
on the one-fth success rule [14] is used to overcome the
drawback of the xed mutation factor method. VSM varies
the mutation factor according to the frequency of successful
mutations to avoid falling into local minima and save more
computational time. The rule of updating mutation factor is
as follows:

kd fm (t),
fm (t + 1) = ki fm (t),

fm (t),

if ps (t) , 1/5
if ps (t) . 1/5
if ps (t) = 1/5

(26)

where ps(t) is the frequency of successful mutations. The

successful mutations denes the objective function of the
best individual in the next generation as being better than
the best individual in the current generation. The factor kd
is set at 0.85 and ki is the reciprocal of kd in this paper.
As shown in (26), if the frequency of successful mutations
is less than one-fth, the mutation factor of the next
generation is changed to a smaller value to search for better
individuals. On the other hand, if the frequency of
successful mutations is larger than one-fth, a larger
mutation factor is adopted for the next generation to speed
up the optimisation processes.
3.3.3 Recombination: In essence, the mutant individual in
(25) is a noisy replica of pi . When the population diversity is
small, the candidate individuals will rapidly gather together
so the individuals cannot be further improved. To extend
the local diversity of the mutant individuals, a
recombination operation is introduced as follows

pij

pij ,
pij ,

if randij . Rr
if randij Rr

(27)

where pij is the jth gene of the ith individual before mutation,
pij represents the jth gene of the ith offspring individual
following mutation, randij is a random number with normal
distribution, and Rr [ [0, 1] is a recombination factor.
Equation (27) indicates each gene of the ith individual is
reproduced from the current gene pij or the mutant gene pij .
3.3.4 Selection: Each offspring individual must compete
against its parent individual based on the objective values
as follows
pi (t + 1) =

pi (t + 1),
pi (t),

if Fi (t + 1) . Fi (t)
if otherwise

(28)

where Fi(t + 1) and Fi (t) represent the objective values of the

ith individual at t + 1 and t iteration, respectively. As shown
in (28), it is observed that any parent individual will be
639

& The Institution of Engineering and Technology 2012

www.ietdl.org
replaced by its offspring individual if the objective value of
the parent individual is worse than that of its offspring
individual.
As described above, the basic DE used one-to-one
competition to retain its offspring that gives rise to a rapid
convergence rate. This rapid convergence may lead to a
higher probability of obtaining a local optimum point
because the diversity of the population descends more
rapidly during the optimisation processes. To increase the
global search capability, a probabilistic state transition rule
used in the ant system is utilised to replace the selection
operation in the basic DE algorithm.
The ant system was rst applied to the travelling salesman
problem [15]. Informally, ants prefer to move to cities which
are connected by short distance with a high amount of
pheromone. However, the cities with short distance and
high pheromone are not absolutely selected by ants. Each
ant generates a complete tour by choosing the cities
according to a probabilistic state transition rule as follows
[t (t)]g [Fi (t)]z
Pri (t) = P i
g
z
i=1 [ti (t)] [Fi (t)]

(29)

where ti (t) is the pheromone concentration of the ith ant at

tth iteration; Fi(t) is the objective value of the ith ant at tth
iteration; g is the pheromone constant; and z is the
constants of the objective value.
In addition, the pheromone concentration is updated
according to the following formula

ti (t + 1) = rti (t) + Dti

(30)

where

q,
Dti = di

0,

if ith ant is better so far

(31)

otherwise

r [ [0, 1] is a pheromone decay parameter, q is a constant

and di is the Euclidean distance. In this paper, di is
proportional to Fi(t).
Equations (29) (31) show if the objective value of the
offspring individual is better than the other individuals, the
pheromone concentration ti is increased and it has more
probability of surviving as a new individual in the next
generation. Compared with the one-to-one competition
method, the probabilistic state transition rule has the
advantages of retaining the diversity of the population and
escaping from local optimal solutions.
3.4

Scheme of HDE to the optimal ARPD problem

The proposed HDE algorithm for searching the optimal

ARPD is depicted in Fig. 1, which is described in the
following steps:
Step 1: Determine the corresponding weighting values of
each objective using the AHP method.
Step 2: Keep the active power of each generator and phase
angle of each bus unchanged.
Step 3: Perform optimal reactive power dispatch using the
proposed HDE approach in the following steps:
Step 3.1: Randomly generate the initial parent trial vector pi
as shown in (24).
640
& The Institution of Engineering and Technology 2012

Fig. 1 Schematic diagram of the proposed method

Step 3.2: Evaluate the objective value of each parent

individual using (18) subject to a number of constraints
shown from (3) (10).
Step 3.3: Execute the VSM and recombination operations
according to the variable mutation factor fm and
recombination factor Rr , as described in (26) and (27),
respectively.
Step 3.4: Calculate the objective value of each offspring
individual as described in (18).
Step 3.5: Utilise the probabilistic state transition rule
described from (29) to (31) to select P sets of the better
individuals in the population.
Step 3.6: Repeat steps 3.2 3.5 until the optimisation process
converges or the maximum iteration is reached.
Step 4: Keep the reactive power of each generator and voltage
magnitude of each bus unchanged.
Step 5: Perform active power dispatch using the proposed
HDE approach. This step is accomplished like reactive
power dispatch, except the control variables are replaced by
the active power output of each generator.
Step 6: Perform load ow analysis to obtain the apparent
power of the slack bus.
Step 7: Check if the variation in apparent power at the slack
bus is less than the tolerance error. If it is, stop the
optimisation processes and output the dispatch results;
otherwise, go to Step 2 and proceed to perform the optimal
ARPD processes.

Numerical results

The proposed approach was veried on the IEEE 30-bus and

TPC 345 kV simplied systems. For comparison, the EP [9],
PSO [10] and DE [11] methods implemented using the
commercial MATLAB package were also tested using the
same database.
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

www.ietdl.org
4.1

Case I (IEEE 30-bus 6-generator system [26])

In case I, the bus 1 is slack bus, buses 2, 5, 8, 11 and 13 are

generation bus, and the others are load bus. For reactive
power dispatch, the objectives of active power loss and
voltage deviation are converted into a single objective
optimisation problem as given in (18). The control variables
are the desired values of voltage magnitude at the PV bus,
transformer tap settings and shunt capacitors. Table 1 shows
the reactive generation constraints of each PV bus. The
lower and upper limits of voltage magnitude at PV bus are
0.90 and 1.10 pu, respectively, whereas the transformer tap
settings are varied between 0.95 and 1.05 pu The weighting
values obtained by the AHP method for both active power
loss and voltage deviation are 0.6667 and 0.3333,
respectively.
For the active power dispatch, the objectives considered are
fuel cost, power wheeling cost and NOx emission. The active
power output of each generator is regarded as the control
variable. To evaluate the power wheeling cost, the length of
the transmission line dened in (12) is assumed to be
proportional to the reactance of the line. Table 2 shows the
coefcients of fuel cost and NOx emission functions. The
obtained weighting values are 0.4294, 0.1440 and 0.4266
for fuel cost, power wheeling cost and NOx emission,
respectively. Table 3 shows the parameter settings of
different methods. The proposed HDE uses the VSM
method to vary the mutation factor between 0.05 and 0.35.
For each method, the population size is set at 30. The
maximum iterations of optimisation and sub-optimisation
schemes are set at 100 and 500, respectively. Owing to the
optimisation processes of diverse methods are almost
converged within 25 iterations through many trials with
different random seeds, the numbers of iteration for these
two schemes are then, respectively, presented within 30 and
50 iterations in order to illustrate the convergence
performance of diverse methods.
Figs. 2 4 show the optimisation processes of different
methods for ARPD, where Fig. 2 represents the suboptimisation processes of the reactive power dispatch at the
last iteration, Fig. 3 shows the sub-optimisation processes of
Table 1

2
5
8
11
13

Methods
EP
PSO
DE

HDE

Parameters

Value

no. of competitors Pm
scaling factor b
inertia weights w
acceleration factors [a1 a2]
no. of differential vector
recombination factor Rr
mutation factor fm
VSM fm
pheromone decay parameter r

15
0.9
0.95
[0.35 0.35]
4
0.6
0.1
0.05 0.35
0.8

active power dispatch at the last iteration and Fig. 4

represents the optimisation processes of apparent power
variation at the slack bus. As shown in Fig. 4, after about
17 iterations, the proposed HDE converges towards the best
one, whereas the DE, EP and PSO methods require about
23, 23 and 22 iterations, respectively. Table 4 shows the
results of reactive power dispatch for different methods,
where the basic load ow solution is regarded as the
benchmark. The results reveal the proposed HDE can save
more active power transmission losses with lower voltage
deviation than the other methods.
Table 5 shows the active power output of each generator.
The associated objective values of fuel cost, power
wheeling cost and NOx emission are also provided in this
table. The table indicates the proposed HDE can achieve

QBi,min (pu)

QBi,max (pu)

2 0.40
2 0.40
2 0.10
2 0.06
2 0.06

0.50
0.40
0.40
0.24
0.24

Fig. 2 Sub-optimisation of reactive power dispatch at the last

iteration

Coefficients of fuel cost and NOx emission functions

Unit

PG1
PG2
PG3
PG4
PG5
PG6

Parameter settings for different methods

Reactive generation constraints for PV bus

Bus

Table 2

Table 3

NOx emission function (tonne/h)

2
Ei (PGi ) = ai + bi PGi + gi PGi

Fuel cost function ($/h)

2
Fi (PGi ) = ai + bi PGi + ci PGi
a

10
10
20
10
20
10

200
150
180
100
180
150

100
120
40
60
40
100

0.04091
0.02543
0.04257
0.05326
0.04257
0.06131

2 0.05554
2 0.06047
2 0.05094
2 0.03550
2 0.05094
2 0.05555

0.06490
0.05638
0.04586
0.03380
0.04586
0.05151

Lower and upper limits of each generating unit are 0.05 and 1.5 pu, respectively
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

641

& The Institution of Engineering and Technology 2012

www.ietdl.org

Fig. 3 Sub-optimisation of active power dispatch at the last

iteration

Fig. 5 Pareto optimal fronts of different methods

better objective value as well as lower operating cost than the

other methods. In addition, to illustrate the trade-off nature
between fuel cost and NOx emission, the Pareto optimal
fronts for different methods obtained by using (18) with
different weights are depicted in Fig. 5, where the Pareto
optimal solution of each method (marked with circle) is
also provided. Note that the dispatch results of each Pareto
optimal solution have been presented in Tables 4 and 5.
To verify the convergence performance of different
methods, simulations of average, minimum and maximum
variation of apparent power at slack bus through 30 trials
using different random numbers are implemented and the
results are shown in Table 6. The results indicate the HDE
can offer shorter execution time with less iteration than the
other methods.
4.2
Fig. 4 Optimisation of apparent power variation at the slack bus

Table 4

Case II (TPC 345 kV simplified system) [24]

The TPC 345 kV simplied system has 33 buses and 16

generators. Most generators are composed of several units.

Results of reactive power dispatch for different methods

SPG (pu)

Methods

load flow solution

DE
EP
PSO
HDE

SQG (pu)

2.88683
2.86310
2.86660
2.85760
2.85120

Qloss (pu)

1.02766
0.99456
0.99898
0.96496
0.91818

(0.2343
(0.2674
(0.2630
(0.2970
(0.3440

Objective value

DPloss (%)

Ploss(pu)

Vdev(pu)

0.5283
0.2910
0.3260
0.2360
0.1720

0.0243
0.0160
0.0156
0.0101
0.0032

0.00
44.92
38.29
55.33
67.44

1. Total active and reactive power demands are 2.834 and 1.262 pu, respectively
2. Term DPloss(%) means the percentage that active power loss saves when compared with the load flow solution

Table 5

Results of active power dispatch for different methods

Methods PG1 (pu) PG2 (pu) PG3 (pu) PG4 (pu) PG5 (pu) PG6 (pu) SPG (pu)

Fuel
Power wheeling NOx emission Objective
cost ($/h)
cost ($/h)
(tonne/h)
value

DE
EP
PSO
HDE

630.8845
622.0571
619.1943
616.9819

0.1651
0.1066
0.1989
0.2020

0.5688
0.4988
0.3366
0.3709

0.7646
0.7846
0.8832
0.7402

0.5835
0.6835
0.6670
0.6422

0.3481
0.4081
0.4237
0.4352

642
& The Institution of Engineering and Technology 2012

0.4330
0.3850
0.3482
0.4607

2.8631
2.8666
2.8556
2.8512

73.7208
73.7762
72.4395
71.0113

0.1946
0.1976
0.1993
0.1932

583.6449
583.7928
583.5567
573.6030

IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

www.ietdl.org
Table 6

Sensitive analysis of different methods (30 trials with different random numbers)

Methods

Variation of apparent power at slack bus

DE
EP
PSO
HDE

Ave.

Min.

Max.

Max. Min.

0.032
0.032
0.027
0.016

0.023
0.021
0.015
0.012

0.039
0.039
0.033
0.016

0.016
0.019
0.018
0.004

The system contains three nuclear units, three hydro units,

one pumped-storage hydro unit and nine thermal units.
Except for the thermal units, the other units are regarded as
must-run units. For the 33 buses, bus 1 is the slack bus,
buses 9, 11, 15, 16, 18, 20, 21, 22 and 27 33 are the PV
bus, and the others are the PQ bus. Therefore 15 control
variables are considered for the reactive power dispatch.
For the active power dispatch, the objectives considered are
fuel cost, power wheeling cost and CO2 emission. The fuel
cost function adopted in this case was a cube curve, and the
CO2 emission was calculated as follows
Ei,CO2 (PG ) = kHe(PGi ) Cai Oxi

(32)

where k is a constant; and He(PGi) is the heat rate of the ith

unit, which is obtained from the fuel cost function. Cai
represents the carbon emission coefcient of the ith unit
and Oxi is the oxidising rate of carbon for the ith unit.

Average time (s)

No. of average iteration

17.02
17.03
16.86
14.45

23
23
22
17

The weighting value of each objective is the same as case

I. Fig. 6 shows the optimisation processes of apparent power
variation at slack bus for different methods. After about 14
iterations, the proposed HDE method converges towards
the value of 0.08 pu, whereas the basic DE, EP and PSO
methods converge towards the values of 0.10, 0.11 and
0.08 pu after about 20, 19 and 17 iterations, respectively.
Table 7 shows the results of reactive power dispatch for
different methods. The results show that the proposed HDE
can obtain lower active power transmission loss with less
voltage deviation than the other methods.
Table 8 shows the active power output of each generator.
The associated fuel cost, power wheeling cost, CO2
emission and the lower and upper limits of each generating
unit are also provided in the table. Results indicate that the
proposed HDE can achieve better objective value as well as
lower operating cost than the other methods. The average
execution times for DE, EP, PSO and HDE methods
through 30 random runs are 19.12, 18.75, 18.05 and
16.24 s, respectively.
4.3

Discussions

Considering the results obtained from the two cases, the

following observations are in order:

Fig. 6 Optimisation of apparent power variation at slack bus for

different methods in case II
Table 7

1. Obtained from Tables 4 and 7 reveal that the proposed

HDE can acquire lower active power loss than the other
methods, which provides the benet of using lower power
generation to schedule power output of each generator for
active power dispatch.
2. Since two sub-optimisation processes, including ARPD,
are executed in a single simulation run, the average
convergence time is perhaps a bit longer than the single
optimisation scheme.
3. Because of the uses of VSM and probabilistic state
transition rule, the proposed HDE converges fast and can
reach better objective value as well as lower operating cost
than the other methods. For comparison, the mutation and
selection operations of different methods are summarised in
Table 9.

Results of reactive power dispatch for different methods in case II

Methods

load flow solution

DE
EP
PSO
HDE

SPG (MW)

23 174
23 104
23 068
23 072
23 052

SQG (MVAR)

15 058
14 798
14 662
14 365
14 654

Qloss (MVAR)

Objective value

24892
25152
25288
25585
25296

Ploss(MW)

Vdev(pu)

496.5
426.4
390.1
394.9
375.0

0.031
0.019
0.015
0.016
0.011

DPloss (%)

0.00
14.12
21.43
20.46
24.47

1. Total active and reactive power demands are 22 678 MW and 1199 MVAR, respectively
2. Power outputs include nuclear, hydro and thermal units
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

643

& The Institution of Engineering and Technology 2012

www.ietdl.org
Table 8

Results of active power dispatch for different methods in case II

Methods PG1
PG2
PG3
PG4
PG5
PG6
PG7
PG8
PG9
(MW) (MW) (MW) (MW) (MW) (MW) (MW) (MW) (MW)

DE
EP
PSO
HDE

3117.9
3019.8
3084.1
3119.4

3120.0
2692.7
3120.0
3120.0

1525.2
2193.5
1279.7
1020.0

1744.7
1702.5
1122.8
1189.1

1473.6
2065.5
2666.6
2880.0

1020.0
1020.0
1094.8
1020.0

1414.5
1020.0
1243.4
1145.0

1025.2
1020.0
1020.0
1020.0

1292.9
1020.0
1064.2
1020.0

SPG
(MW)

Fuel cost
($/h)

15 734.3
15 754.0
15 695.8
15 533.6

15 607 913
14 368 442
14 731 799
14 016 092

Power
CO2 emission Objective
(tonne/h)
value
wheeling
cost ($/h)
164 413
166 669
164 938
164 823

3 931 289
3 880 663
3 886 864
3 841 249

17 455 023
16 847 297
16 991 627
16 585 340

Lower and upper limits of each generating unit are as follows:

PG1 and PG2: 1020 MW 3120 MW; PG3 and PG4: 728 MW2767 MW; PG5: 1020 MW 2880 MW; PG6: 520 MW 2910 MW;
PG7: 1020 MW1717 MW; PG8: 700 MW 3150 MW; PG9: 867 MW 1300 MW

Table 9

Summary of mutation and selection operations for different methods

Methods
DE
EP
PSO
HDE

Mutation operation

add a differential vector to the current individual

use one-to-one (offspring to-parent) competition to retain offspring
with fixed mutation factor
add a Gaussian random variable to the current individual
select the first k winners following random competition for all
individuals
add a new velocity to the current position
one-to-one (offspring to-parent) competition to retain offspring
add a differential vector to the current individual
select the first k winners following probabilistic state transition
with variable mutation factor
rule for all individuals

Conclusions

A new algorithm to alternatively solve the optimal ARPD

problem has been proposed and implemented in this paper.
The paper rst introduces the formulation of ARPD
problem. Then the EC algorithms including EP, PSO, DE
and the proposed HDE for determining the optimal ARPD
problem in a single simulation run are briey reviewed.
Owing to the efcient global or near global search scheme,
the proposed HDE algorithm can offer higher probability of
converging towards global solution than the other methods.
Testing on the IEEE 30-bus and TPC 345 kV simplied
systems showed that the proposed HDE can save more
active power transmission losses with lower voltage
deviation and offer lower operating cost with better
convergence performance than the basic DE, EP and PSO
methods.

Acknowledgments

Financial supports from the National Science Council,

Taiwan, R.O.C. under the grant no. NSC 101-3113-P-006020 are acknowledged.

Selection operation

References

1 Arvanitidis, C.H.N.V., Luenberger, D.G.: Decomposition of real and

reactive power ows: a method suited for on-line applications, IEEE
Trans. Power Appar. Syst., 1972, PAS-91, (2), pp. 661670
2 Stott, B., Alsac, O.: Fast decoupled load ow, IEEE Trans. Power
Appar. Syst., 1974, PAS-93, (3), pp. 859 869
3 Chebbo, A.M., Irving, M.R.: Combined active reactive dispatch, part 1:
problem formulation and solution algorithm, IEE Proc. Gener. Transm.
Distrib., 1995, 142, (4), pp. 393399
4 Chebbo, A.M., Irving, M.R., Dandachi, N.H.: Combined active reactive
dispatch, part 2: test results, IEE Proc. Gener. Transm. Distrib., 1995,
142, (4), pp. 401405
5 Lee, K.Y., Park, Y.M., Ortiz, J.L.: A united approach to optimal real
and reactive power dispatch, IEEE Trans. Power Appar. Syst., 1985,
PAS-104, (5), pp. 661670

644

& The Institution of Engineering and Technology 2012

6 Affonso, C.M., da Silva, L.C.P., Lima, F.G.M., Soares, S.: Optimal

MW/MVAR dispatch and minimal load shedding strategy for
improving voltage stability margin. IEEE Power Eng. Society
General Meeting, 13 17 July 2003, vol. 2, pp. 890 895
7 Rosehart,
W.D.,
Canizares,
C.A.,
Quintana,
V.H.:
Multiobjective optimal power ows to evaluate voltage security costs
in power networks, IEEE Trans. Power Syst., 2003, 18, (2),
pp. 578586
8 Lee, S.J., Yang, S.D.: Derivation of PQ loss sensitivities by angle
reference transposition and an application, IEEE Trans. Power Syst.
Lett., 2006, 21, (1), pp. 428 430
9 Fogel, D.B.: System identication through simulated evolutionary:
a machine learning approach to modeling (Ginn Press, Needham,
MA, 1991)
10 Kennedy, J., Eberhart, R.: Particle swarm optimization. Proc. IEEE Int.
Conf. Neural Networks, 1995, vol. 4, pp. 19421948
11 Storn, R., Price, K.: Differential evolution: a simple and efcient
adaptive scheme for global optimization over continues space.
Technique report TR-95012, International Computer Science
Institute, Berkeley, 1995
12 Alireza, A., Tabatabaii, I., Hosseini, S.H.: Optimal reactive power
dispatch in electricity markets using a multiagent-based differential
evolution algorithm. Int. Conf. Power Eng., Energy and Elect.
Drives, Setubal, Portugal, 1214 April 2007, pp. 249254
13 Varadarajan, M., Swarup, K.S.: Solving multi-objective optimal power
ow using differential evolution, IET Gener. Transm. Distrib., 2008, 2,
(5), pp. 720 730
14 Back, T., Schwefel, H.P.: An overview of evolutionary algorithms for
parameter optimization, Evol. Comput., 1993, 1, (1), pp. 1 23
15 Dorigo, M., Maniezzo, V., Colorni, A.: Ant system: optimization by a
colony of cooperative agents, IEEE Trans. Sys. Man Cyber., B, 1996,
26, (1), pp. 29 41
16 Zhao, B., Guo, C.X., Cao, Y.J.: A multiagent-based particle swarm
optimization approach for optimal reactive power dispatch, IEEE
Trans. Power Syst., 2005, 20, (2), pp. 10701078
17 Zhang, C., Chen, M., Luo, C.: A multi-objective optimization method
for power system reactive power dispatch. Proc. Eighth IEEE Congress
on Intelligent Controll and Automation, Jinan, China, July 2010,
pp. 6 10
18 Chakraborty, S., Senjyu, T., Yona, A., Saber, A.Y., Funabashi, T.:
Solving economic load dispatch problem with valve-point effects
using a hybrid quantum mechanics inspired particle swarm
optimization, IET Gener. Transm. Distrib., 2011, 5, (10),
pp. 1042 1052
19 Ramanathan, R.: Emission constrained economic dispatch, IEEE
Trans. Power Syst., 1994, 9, (4), pp. 1994 2000

IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

www.ietdl.org
20 Sood, Y.R., Padhy, N.P., Gupta, H.O.: Wheeling of power under
deregulated environment of power system a bibliographical survey,
IEEE Trans. Power Syst., 2002, 17, (3), pp. 870 878
21 Wood, A.J., Wollenberg, B.F.: Power generation operation and control
(John Wiley & Sons, Inc., New York, 1996, 2nd edn.)
22 Liu, X.: Emission minimisation dispatch constrained by cost and wind
power, IET Gener. Transm. Distrib., 2011, 5, (7), pp. 735742
23 Wong, K.P., Fan, B., Chang, C.S., Liew, A.C.: Multi-objective
generation dispatch using bi-criterion global optimization, IEEE
Trans. Power Syst., 1995, 10, pp. 1813 1819
24 Huang, C.M., Yang, H.T., Hong, Y.Y., Hong, S.P., Liou, K.P.: Power
dispatching considering fuel cost and CO2 emission, Monthly
J. Taipowers Eng., 1999, 610, pp. 32 48
25 Satty, T.L.: The analytical hierarchy process: planning, priority setting,
resource allocation (McGraw-Hill, New York, 1980)
26 Yokoyama, R., Bae, S.H., Morita, T.: Multiobjective optimal
generation dispatch based on probability security criteria, IEEE
Trans. Power Syst., 1988, 3, (1), pp. 317324

Table 10 Intensity scale of importance [25]

Intensity of importance

Definition

1
3
5
7
9
2, 4, 6, 8

equal importance
weak importance
strong importance
very strong importance
absolute importance
intermediate values between
adjacent scale values

corresponding questionnaire matrix, an index of consistency

ratio (CR) is calculated as follows


Appendix

The AHP approach [25] is a weight assessment technique

which is employed to help the operators determine the
weighting value of each objective in this paper. As shown
in Table 10, the AHP creates an intensity scale of
importance to transform these linguistic terms into
numerical intensity values. Let C1 , C2 , . . . , Cp be the set of
objectives considered, the quantied judgments on pairs of
objectives Ci , Cj , are then represented by a p-by-p matrix.

1
1/a12

A = [aij ] = .
..
1/a1p

a12
1
..
.
1/a2p

..
.

a1p
a2p

..
.
1

(33)

where the entries aij (i, j 1, 2, . . . , p, where p is the number

of the objective) are determined by the operators. If aij a,
then aji 1/a, where a is an intensity value determined by
the operators, as shown in Table 10. On the other hand, if
Ci is judged to be of equal importance as Cj , then aij 1,
and aji 1; in particular, aii 1 for all i.
As shown in (33), an eigenvector of matrix A
corresponding to the largest eigenvalue can then be
obtained. Furthermore, to check the effectiveness of the

IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 7, pp. 636 645
doi: 10.1049/iet-gtd.2011.0559

CR =


lmax p
/RI
p1

(34)

where lmax is the largest eigenvalue of matrix A and RI

denotes the random index [25]. In general, a CR of 0.10 or
less is considered acceptable.
For instance, as shown in case I of this paper, if fuel cost
(F) is weak important than power wheeling cost (Ch) and
equal important with NOx emission (E), the questionnaire
matrix in (33) can be obtained by

F
1 3
A = [aij ] = Ch 1/3 1
E 1/1 3

1
1/3
1

(35)

The largest eigenvalue of this matrix is 3.00. After

normalising the associated eigenvector, we can obtain the
weighting vector [0.4294, 0.1440, 0.4266]. It means the
weighting values of fuel cost, power wheeling cost and NOx
emission are 0.4294, 0.1440 and 0.4266, respectively. In
the present example, lmax is 3.00, p is 3, and RI is 0.58.
Thus, CR is 0.00. With the acceptable CR pre-specied at
0.1 or less value, this assessment is regarded as an effective
process.

645

& The Institution of Engineering and Technology 2012