Electrical Power and Energy Systems 24 (2002) 573±581

www.elsevier.com/locate/ijepes

Reactive power compensation for radial distribution networks using

genetic algorithm
D. Das*
Department of Electrical Engineering, Indian Institute of Technology Kharagpur, W.B. 721302, India
Received 7 April 1999; revised 9 July 2001; accepted 8 August 2001

Abstract
In the present work, genetic algorithm is applied to select the optimum values of ®xed and switched shunt capacitors required to be placed
on a radial distribution network under varying load conditions so as to minimize the energy loss while keeping the voltage at load buses
within the speci®ed limit by taking the cost of the capacitors into account. Both ®xed and marginal costs of the capacitors are considered to
obtain the best overall performance. The voltage constraint is included as penalty term in the objective function.
Further, instead of only assuming the capacitors as a constant reactive power load, it is also been considered as a constant impedance load
in optimization problem itself and comparison of performances for both considerations are also presented.
Also possible convergence criteria of genetic algorithm based on `the difference between best ®tness and average ®tness' of population is
suggested to reduce the CPU time required to a great extent with a very small sacri®ce in overall savings. q 2002 Elsevier Science Ltd. All
rights reserved.
Keywords: Distribution networks; Shunt capacitors optimization; Genetic algorithm

1. Introduction a near global optimum solution. Chiang et al. [17] have used
the method of simulated annealing to obtain the optimum
As the cost of constructing a new power plant has values of shunt capacitors for radial distribution networks.
skyrocketed, the electric power industry is making every Sundhararajan and Pahwa [18] have used genetic algorithms
effort to reduce the growth of electricity demand. Since a for obtaining the optimum values of shunt capacitors. They
substantial amount of generated power is being wasted as have [18] treated the capacitors as constant reactive power
losses, reduction in losses has been recognized as a viable load and no method is suggested to reduce the CPU time.
option to eliminate to some degree the need for unnecessary Genetic algorithm based solution is capable of determining
additional generating capacity. It is acknowledged that a near global solution with lesser computational burden
much of this power loss occurs in the distribution system. than the simulated annealing method [18]. An optimization
To date, volt-var control has been a commonly implemented method is judged for its ef®ciency by the quality of the
control practice to reduce distribution feeder losses while resulting solution and the number of function evaluation
maintaining acceptable feeder voltage pro®le. required to arrive at the quality solution. In the present
In the past, a lot of work has been carried out in the area paper, the ®nal solutions, cost savings and computer run
of reactive power compensation for radial distribution time are presented and the proposed method has also been
networks [1±16]. These methods are based on different compared with two other existing methods.
non-linear programming techniques and provide only a Genetic algorithm is becoming popular to solve the
local optimum solution with a lesser computer burden. optimization problems in different ®elds of applications
Recently, researchers have attempted to obtain optimum mainly because of its robustness in ®nding optimal solution
values of shunt capacitors for radial distribution networks and ability to provide near optimal solution close to global
using simulated annealing and genetic algorithm (GA). minimum [19,20,22]. Genetic algorithms employ search
Because these two techniques have the capability to search procedures based on the mechanics of natural selection
and survival of the ®ttest. In the GAs, which uses multiple
* Tel.: 191-3222-79507; fax: 191-3222-55303. point search instead of single point search and work with the
E-mail address: [email protected] (D. Das). coded structure of variables instead of the actual variables
0142-0615/02/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0142-061 5(01)00068-0
574 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581

themselves, the only information required is the objective power of GAs derives largely from their ability to exploit
function thereby making the searching for global optimum ef®ciently this vast amount of accumulating knowledge by
simpler. means of relatively selection mechanisms.
In the present work, GA is implemented to determine the
optimal sizes of ®xed and switched capacitors under varying 2.1. How genetic algorithm works
load conditions. The objective is to minimize the energy
A GA is an iterative procedure which maintains a
losses in the distribution network by taking the cost of capa-
constant size population p(t) of candidate solutions. The
citors into account. Also the voltage constraint is formulated
algorithm begins with a randomly selected population of
as a penalty function to the objective function.
function inputs represented by strings of bits. During each
Further important contribution of the present work is,
iteration step, called a generation, the structure in the current
instead of only assuming the capacitor as a constant reactive
population is evaluated and on the basis of those evalua-
power (KVAr) load (which has been assumed in all of the
tions, a new population of candidate solution is formed.
previous work), it is also been considered as constant
That is GA uses the current population of string to create
impedance load which can be thought as an approach to
a new population such that the string in the new population
make the problem formulation more realistic from practical
are, an average `better' than those in the current population.
point of view. A comparative study of both considerations
The idea is to use the best elements from the current
(i.e. consideration of capacitors as constant power and
population to help form the new population. If this is done
constant impedance load) in problem formulation is also
correctly, then the new population will, on average, be
highlighted.
`better' than the old population. Three processes Ð selec-
Another important aspect of the present work is to
tion, crossover, and mutation are used to make the transition
propose a secondary convergence criteria. The criteria
from one population generation to the next. The basic
for termination of genetic algorithm used was maximum
genetic algorithm cycle based on these is shown in Fig. 1.
number of generation reached [18,19]. If a secondary
These three steps are repeated to create each new generation
convergence criteria (based on `the difference between
and it continues in this fashion until some stopping condi-
best ®tness and average ®tness' of population) is used, it
tion is reached (such as maximum number of generations or
could reduce a lot of computational time with only a slight
resulting new population not improving fast enough).
reduction in overall savings.
2.2. Initialization

2. Genetic algorithm GAs operate with a set of strings instead of a single string.
This set or population of string goes through the process of
Genetic algorithms (GAs), a way to search for the best evolution to produce new individual strings. To begin with,
answers to tough problems, were ®rst suggested by John the initial population could be seeded with heuristically
Holland in his book `Adaptation in Natural and Arti®cial chosen strings or at random [19]. In either case, the initial
Systems' [20]. Over the last few years, it is becoming popu- population should contain a wide variety of structures.
lar to solve a wide range of search, optimization and
machine learning problems. As their name indicates, genetic 2.3. Evaluation function
algorithm attempts to solve problems in a fashion similar to
The evaluation function is a procedure to determine the
the way in which human genetic processes seem to operate.
®tness of each string in the population and is very much
It's fundamental principle is the ®ttest member of popula-
application oriented. The performance of each structures
tion has the highest probability for survival [22].
of string is evaluated according to its ®tness, which is
A GA (multi path search scheme) is an iterative proce-
de®ned as a non-negative ®gure of merit to be maximized.
dure which maintains a constant size population p(t) of
It is directly associated with the objective function value
candidate solutions. The initial population p(0) can be
[18] in the optimization. GA treats the problem as a black
chosen heuristically or at random [19]. The structures of
box in which the input is the structure of the string and the
the population p(t 1 1) (i.e. for next iteration called genera-
tion) are chosen from p(t) by randomized selection pro-
cedure that ensures that the expected number of times a
structure is chosen is approximately proportional to that
structure's performance relative to the rest of the popula-
tion. In order to search other points in a search space, some
variation is introduced into the new population by means of
genetic operators (crossover and mutation).
While it may seem to be a random search, in fact, the
improvement in each generation indicates that the algorithm
produces an effective directed search technique [22]. The Fig. 1. The basic genetic algorithm cycle.
 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581 575

output is its ®tness [22]. Because GA proceeds only accord-

ing to the ®tness of the strings and not to other information,
the properties of the ®tness will in¯uence the GAs perfor-
mance. The evaluation function Fe for binary vector C is
equivalent to the function F, i.e. Fe C† ˆ F X†; where the Fig. 2. The crossover operation.
chromosome C corresponds to real value X. The evaluation
function plays the role of the environment rating potential 2.6. Mutation
solutions in terms of their ®tness.
In many problems, the objective is more naturally stated Selection and crossover effectively search and recombine
as the minimization of some cost function S (as the present the existing chromosomes. However, they do not create any
capacitor placement problem) rather than the maximization new genetic material in the population. Mutation is capable
of some utility or pro®t function U(x). Hence, it is often of overcoming this shortcoming. It is an occasional random
necessary to map the objective function to a ®tness function alteration of string position. It is not a primary operator but
form. it ensures that the probability of searching any region in the
problem space is never zero and prevents complete loss of
2.4. Selection genetic material through selection and crossover.
For this, a ®xed small mutation probability (Pm) is set at
This is the ®rst step of the three genetic operations. This the start of the algorithm and bits in all the new strings
determines which strings in the current generation will be (offspring) are then subject to change (from 1 to 0 or vice
used to create the next generation. This is done by using a versa) based on this mutation probability. In Fig. 3, bits 6
biased random Ð selection methodology [22]. That is, and 10 (for 11 bit string) are mutated. Bit 6 goes from dark
parents are randomly selected from the current population to light and bit 10 goes from light to dark. The result is a new
in such a way that the `best' strings in the population have generation string.
the greatest chance of being selected. There are many ways
to do this, one commonly used technique is Roulette Wheel 2.7. Genetic control parameter selection
parent selection (used in the present work). The following
steps are carried out in Roulette Wheel parent selection Genetic parameters (namely population size, crossover
algorithm: rate and mutation rate) are the entities that help to tune
the performance of the genetic algorithm [19,22]. The selec-
(i) Sum the ®tness of all the population members. Call it tion of values for these parameters plays an important role in
as total ®tness. obtaining an optimal solution. There are some general
(ii) Generate a random number (ri) between zero and the guidelines, which could be followed to arrive at an optimal
total ®tness. values for these parameters [19].
(iii) Select a population member whose cumulative Population size. Genetic algorithms (GAs) do poorly with
®tness obtained from adding its ®tness to the ®tness of very small populations, because the population provides an
the proceeding population members, is greater than or insuf®cient sample size for most hyper planes. A large popu-
equal to ri. lation is more likely to contain representative from a large
number of hyper planes. Hence GAs can perform a more
This algorithm is referred to as Roulette Wheel selection informed search. As a result, a large population discourages
because it can be viewed as allocating pi-shaped slice on premature convergence to suboptimal solution. On the other
Roulette Wheel to population members, with each slice hand, a large population requires more evaluations per
proportional to member's ®tness. This parent selection generation, possibly resulting in an unacceptably slow rate
technique has the advantage that it directly promotes of convergence.
reproduction of the ®ttest population member by biasing Crossover rate (Pc). The higher the crossover rate, the
each member's chance of selection in accordance with its more quickly new structures are introduced into the popula-
evaluation. tion. If the crossover rate is too high, high performance

2.5. Crossover

This (mating process) is the second step, which deter-

mines the actual form of the strings in the next generation.
At this point, two of the selected parents are paired. If the
length of the each string is r, then a random number between
1 and r is selected, say d. The mating process is one of
swapping bits d 1 1 through, r of the ®rst parent with bits
d 1 1 through r of the second parents. In this way two new
strings (called offspring) are created. This is shown in Fig. 2. Fig. 3. The mutation operator.
576 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581

structures are discarded faster than the selection can produce also to place the minimum number of capacitors so that
improvements. If the crossover rate is low, the search may installation cost could be minimum. Hence it will be worth-
stagnate due to the lower exploration rate. while, if the installation cost of the capacitor could be
Mutation rate. A low level of mutation serves to prevent incorporated in the capacitor cost calculation itself.
any given bit position from getting stuck to a single value Hence, in this work, both the installation cost and
where as a high mutation rate results in essentially a random marginal cost of capacitor has been considered separately.
search. The capacitor cost is considered as the linear function of
its size, i.e. [15]

3. Handling of constraints Kc Qci †Qci ˆ Kcf 1 Kc Qci 2†

Since GAs are domain independent, one of the conse-

where Kcf is the installation cost.
quence of the neatness of GAs is their inability to deal
with functional constraints. Few approaches to handle the 4.1. Handling of voltage constraints
constraints in genetic algorithms have previously been
proposed. One of these uses penalty functions as an adjust- As pointed out in Section 3, genetic algorithms are
ment to objective function. Other approaches use decoder or essentially unconstrained procedures within the given repre-
repair algorithms which avoid building an illegal individual sentation space. Hence, in this work voltage constraint is
or repair one, respectively [22]. formulated as a penalty function to the basic objective func-
If one incorporates a high penalty into the evaluation tion of Eq. (1). That is, the constrained objective function
routine and the domain is one in which production of an could be represented as:
individual violating the constraints is likely, one runs Minimize
the risk of creating a genetic algorithm that spends
most of its time evaluating illegal individuals. Further, X
L X
ncap
l s
it can happen that when a legal individual is found, it S ˆ Ke T j Pj 1 Kc Qci †Qci 1 l Vmin 2 Vmin †2 3†
drives the others out and the population converges on it jˆ1 iˆ1

without ®nding the better individuals, since the likely l s

paths to other legal individual requires the production of where Vmin is the minimum voltage limit, Vmin the mini-
illegal individuals as intermediate structures, and the mum voltage of the system and l is the constant
penalties for violating the constraints make it unlikely multiplier
that such intermediate structure will reproduce. If one The value of l plays an important role, while handling
imposes moderate penalties, the system may evolve indivi- with constrained optimization. A higher value of l can
duals that violate the constraints but are rated better than cause premature convergence while with lower value of
those that do not because the rest of the evaluation function l , the system may evolve individuals that violates the
can be satis®ed better by accepting the moderate constraint constraint but are rated better than those that do not.
penalty than by avoiding it. Hence suitable value of l , satisfying the constraint limit
and also giving best possible result could only be found
out by trial and error.
4. Basic objective function

With the delivery of customer load on the distribution 5. Implementation of GA for capacitor optimization
system there are demand (kW) losses which over a period problem
of time will re¯ect as energy losses. Hence the objective
function for optimization can be stated mathematically as to Here, the purpose of the GA is to determine the
[16,17] capacitor sizes at the candidate locations for each load
Minimize level. If we have `ncap' candidate locations and L load
levels, the GA returns the value for `ncap £ L' design
X
L X
ncap variables.
S ˆ Ke Tj Pj 1 Kc Qci †Qci 1† As mentioned in Section 3, in GA, the `®tness' is de®ned
jˆ1 iˆ1 as nonnegative ®gure of merit to be maximized which is
directly associated with objective function. Here the objec-
where Pj is the power loss at any load level j with a time tive is to minimize Eq. (1) or (3), which of course guarantees
duration Tj ; Kc Qci † the cost of the ith capacitor of rating Qci ; that S is always positive, therefore if we consider the inverse
ncap the number of candidate locations for capacitor place- function of S as the ®tness than it will satisfy both conditions
ment, L the number of load level, T the design period and Ke (i.e. positive and ®gure of merit to be maximized). There-
is the energy cost of losses. fore in the present work ®tness has been de®ned as K/S,
One of the main objectives of compensation problem is where K is a constant multiplier. A suitable value of K
 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581 577

adjusts the ®tness function in a close range of decimal STEP-4:

places.
Since GA is a multipath search technique, the number set generation number ˆ 1
of functions evaluation are greater. One of the major tasks STEP-5:
to be carried out for each objective function calculation is
`load ¯ow'. In this work an ef®cient load ¯ow algorithm for for i ˆ 1 to i ˆ ªpopulation size £ crossover rateº; DO;
solving radial distribution networks developed in Ref. [21]
is used. The computer program was written in FORTRAN-
(i) Select two parents from population(based on roulette
77 and implemented on METEOR-400 VT with 66 MHz
wheeling method).
clock.
(ii) Perform crossover and hence generate two offspring.
The complete evaluation procedure is given below:
(iii) Mutate these two offspring based on mutation
STEP-1:
probability.
Form initial binary strings (chromosomes) equal to
population size as:
STEP-6:
‰Cap11 Cap12 ¼Cap1ncap ; Cap21 Cap22 ¼Cap2ncap ; ¼; Calculate ®tness of each offspring (as in STEP-2 and
STEP-3).
CapL1 Capl2 ¼CapLncapŠ STEP-7:
Combine the old population and new population to
where Capij represents the binary substrings (0 and 1), repre-
create a single population best of both. For example, at
senting the value of shunt at jth candidate location for ith
(t)-th generation, total population size is NPOP (old popula-
load level without any voltage constraint, at any load level
tion). Now in STEP-5, a set of new population is obtained
Capij is taken as 11 bits, i.e. allowing the capacitor to vary
size is also NPOP. When combining these old population
from to 2047 kV Ar and with voltage constraint Capij is
and new population, total population size is 2 £ NPOP. Out
taken as 12 bits, i.e. allowing the capacitor to vary from 0
of this population size of 2 £ NPOP, best NPOP size popu-
to 4095 kV Ar, ncap the number of candidate location and L
lation (based on their ®tness value) is selected (t 1 1)-th
is the number of load levels.
generation.
STEP-2:
STEP-8:
Evaluation of the ®tness for each string is as follows:
Repeat STEP-5 to STEP-7 till the solution converges or
(i) each binary substring (genes) in a string into decimal maximum number of generation is reached.
number representing the size of each capacitor; In the present work maximum number of generation was
(ii) set load level ˆ 1; set to 200 and the new convergence criteria based on the
(iii) assign capacitor sizes for the given locations (larger difference between `best ®tness and average ®tness' of the
than the minimum required); population is explained in Section 9.3.
(iv) carry out load ¯ow to ®nd out: power loss, energy
loss and voltage constraints at this load level;
6. Selection of ®xed and switched capacitors
(v) increment load level, i.e. load level ˆ load level 1 1;
(vi) check if load level . L; if not repeat (iii)±(v);
After performing STEP-1 to STEP-8 mentioned in
(vii) sum up the energy losses to getPtotal energy loss;
Section 5, it will give the capacitors sizes for each
(viii) calculate the capacitor Cost ˆ ncap
iˆ1 Kcf 1 Kc Qci †; load levels. At any location, the minimum capacitor
(ix) evaluate the objective function:
sizes that is required for any load level can be consid-
S ˆ total energy loss 1 capacitor cost ered as the size of ®xed capacitor that could be placed
in that location. Any extra capacitor over ®xed capacitor
1 voltage constraint penalty; that will be required at that location could be taken as
switchable capacitor.
(x) calculate the ®tness ˆ K/S.
7. Capacitor as a constant impedance load
STEP-3:
As pointed out by Roger C. Dugan in the discussion
(i) Find out the cumulative ®tness of each string by part of Ref. [15], since capacitor placement on a feeder
adding its ®tness to the ®tness of the proceeding popula- signi®cantly affects the voltage and a capacitor is a
tion members. constant impedance element, the assumption that capa-
(ii) Sum the ®tness of all the population members. citor is a source of constant reactive power, weakens
(iii) Find the best ®tted string and send it to the solution any claim that the result presented by this consideration
vector. is an optimal solution. In this work, a comparative study
578 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581

of optimal results obtained by both assumptions is carried Table 2

out. Power losses and minimum voltage for bare system
Consideration of capacitor as a constant impedance Load Power Minimum voltage
element means that the kV Ar provided by a capacitor level loss (kW) magnitude (pu)
varies by square of the voltage magnitude and suitable
0.50 51.597 V54 ˆ 0:95668
modi®cations in the load ¯ow subroutine has been done.
1.0 224.96 V54 ˆ 0:90919
For this, in ®rst iteration it has been assumed that the 1.80 867.16 V54 ˆ 0:82030
bus voltages of each node is 1.0 pu and hence capacitor
provides its full rated kV Ar but from next iteration
node voltage magnitudes will be known quantity and 9. Test results
kV Ar provided by the capacitor is the multiple of
the square of the voltage magnitude of that location GA, presented in this work, is applied for study of the
to its rated kV Ar. following test cases (each test case includes further
subcases):

8. Analysis CASE-A. Locations of the capacitors are known and

capacitors are treated as constant reactive power load.
In order to test the proposed genetic algorithm a 69 CASE-B. With consideration of capacitors as constant
node radial distribution system has been taken as an impedance load for test Case-A.
example network [15,16]. Cost ®gure adopted [15] is CASE-C. With convergence criteria based on the
as follows: Ke ˆ 0:06$=kwh; capacitor ®xed cost, Kcf ˆ difference between `best ®tness and average ®tness' of
$1000; capacitor marginal cost, Kc ˆ 3$=KV Ar: The population for test Case-A and Case-B.
load duration data [15] for the system is given in Table 1.
The power losses and minimum voltage for bare system
(without shunt capacitors) at different load level are given 9.1. Test Case-A
in Table 2.
In this case, three candidate location node numbers 20, 50
and 53 of the example network are chosen and capacitors
8.1. Selection of control parameters are treated as constant reactive power load. The algorithm is
tested for optimal sizing of the capacitors of the following
First of all, effect of population sizes for different test
subcases:
cases were observed. Different population sizes (40, 50,
60, 70 and 80) were taken and it has been observed that
Case-A1. Obtain the optimum value of shunt capacitors
when the number of variables were lesser, population size
without voltage constraint.
of 60 or even 50 were satisfactory, but with large number of
Case-A2. Obtain the optimum value of shunt capacitors
variables (i.e. increasing the number of candidate locations),
with voltage constraint.
population size of 70 was performing well and this has been
taken accordingly. For the purpose of comparison results for test Case-
It is also worth mentioning here that with different
A1 and Case-A2 are shown in Table 3. Table 3 shows
initial population, ®nal results were more or less same
the values of ®xed and switched capacitors, real power
for capacitors optimization of radial distribution networks.
loss and minimum voltage at different load level.
Also note that ®nal results of GA depend on the crossover
Table 3 also shows the annual saving and total CPU
and mutation probabilities. For best performances, some
time.
experiments are necessary for obtaining suitable combina-
From Tables 2 and 3 (Case-A1), we can observe that
tion of crossover and mutation probabilities.
the minimum voltage at peak load without considering
Experiments were carried out for different combination of voltage constraint has been increased from 0.82030 pu
four crossover probabilities (0.6, 0.8, 0.9 and 1.0) and four
(minimum voltage for bare system at peak load, Table 2) to
mutation probabilities (0.0001, 0.001, 0.006 and 0.01). It
0.85998 pu and the annual saving is $42209 (Case-A1,
has been found that a crossover probability, Pc ˆ 1:0 and
Table 3).
mutation probability Pm ˆ 0:006 give the better perfor-
For test Case-A2, minimum voltage limit of 0.90 pu
mances for different test cases.
was set and the purpose was to get the solution which
Table 1 also satisfy this voltage constraints at all nodes (for
Load duration data each load level). Different values of l were tried out and
the value of l for the best performance (giving best saving
Load (pu) 1.80 1.0 (nominal) 0.50
while nearly satisfying the voltage constraint) was found to
Duration (h) 1000 6760 1000
be 1.0 £ 10 6.
 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581 579

Table 3
Results for test CASE-A1 and CASE-A2

Different Load Capacitors (kV Ar) Loss Vmin (pu) Annual CPU
cases level (kW) saving ($) time (s)
Fixed Switched

CASE A1 0.50 Qc20 ˆ 192 Qc20 ˆ 0:0 34.43 V54 ˆ 0:96624 42,209 634
Qc50 ˆ 511 Qc50 ˆ 0:0
Qc53 ˆ 70 Qc53 ˆ 0:0
1.0 Qc20 ˆ 192 Qc20 ˆ 118 146.33 V54 ˆ 0:93186
Qc50 ˆ 511 Qc50 ˆ 512
Qc53 ˆ 70 Qc53 ˆ 186
1.80 Qc20 ˆ 192 Qc20 ˆ 118 556.08 V54 ˆ 0:85998
Qc50 ˆ 511 Qc53 ˆ 736
Qc53 ˆ 70 Qc53 ˆ 500
CASE A2 0.50 Qc20 ˆ 164 Qc20 ˆ 0:0 34.32 V54 ˆ 0:96696 27,062 639
Qc50 ˆ 517 Qc50 ˆ 0:0
Qc53 ˆ 111 Qc53 ˆ 500
1.0 Qc20 ˆ 164 Qc20 ˆ 168 146.30 V54 ˆ 0:93179
Qc50 ˆ 517 Qc50 ˆ 506
Qc53 ˆ 111 Qc53 ˆ 137
1.80 Qc20 ˆ 164 Qc20 ˆ 348 681.21 V54 ˆ 0:89720
Qc50 ˆ 517 Qc50 ˆ 2301
Qc53 ˆ 111 Qc53 ˆ 1239

Again from Tables 2 and 3 (Case-A2), we can observe of the test Case-A1 (Section 9.1) considering capacitor
that the minimum voltage at peak load considering as constant impedance load.
voltage constraint has been increased from 0.82030 pu Comparison of the results are summarized in Tables 4
(minimum voltage for bare system at peak load, Table and 5. From Table 4, it is seen that when capacitors are
2) to 0.89720 pu (very much closer to minimum voltage treated as constant impedance load, total amount of capaci-
limit) but with sacri®ce in peak power loss, annual tors (®xed 1 switched) during peak load period is much
saving and higher capacitor cost (Case-A2, Table 3). higher as compared to that of capacitors as constant reactive
For the test Case-A2, annual saving is $27,062. However, power load. From Table 5, it is seen that there is not much
for test Case-A1 and Case-A2, CPU time is more or less difference in peak power loss. From Table 5, it is also seen that
same (Table 3). the annual saving is less and this is due to higher investment of
capacitors when they are treated as constant impedance load.
9.2. CASE-B: test results with consideration of capacitors Similar ®ndings were also observed for test Case-A2.
as constant impedance load
9.3. Results with convergence criteria based on difference of
For test cases A1 and A2, capacitors were treated as best ®tness and average ®tness
constant reactive power load. But if we treat the capa-
citor as constant impedance load, than the capacitor In the results presented in previous sections, the criterion
optimization problem will be more realistic. Due to for termination of algorithm used was maximum number of
limitation of space, here we only present the results generation reached. But it was observed that after some
Table 4
Comparison of the results of the test Case-A1 considering capacitor as constant reactive power and constant impedance load

Test Case-A1 Load Capacitors as constant reactive power Capacitors as constant impedance load
level load (kV Ar) (kV Ar)

Fixed Switched Fixed Switched

Optimum value of 0.50 Qc20 ˆ 192 Qc20 ˆ 0:0 Qc20 ˆ 192 Qc20 ˆ 0:0
shunt capacitors without Qc50 ˆ 511 Qc50 ˆ 0:0 Qc50 ˆ 511 Qc50 ˆ 0:0
voltage constraint Qc53 ˆ 70 Qc53 ˆ 0:0 Qc53 ˆ 127 Qc53 ˆ 0:0
1.0 Qc20 ˆ 192 Qc20 ˆ 118 Qc20 ˆ 192 Qc20 ˆ 162
Qc50 ˆ 511 Qc50 ˆ 512 Qc50 ˆ 511 Qc50 ˆ 696
Qc53 ˆ 70 Qc53 ˆ 186 Qc53 ˆ 127 Qc53 ˆ 130
1.80 Qc20 ˆ 192 Qc20 ˆ 118 Qc20 ˆ 192 Qc20 ˆ 192
Qc50 ˆ 511 Qc50 ˆ 736 Qc50 ˆ 511 Qc50 ˆ 1024
Qc53 ˆ 70 Qc53 ˆ 500 Qc53 ˆ 127 Qc53 ˆ 650
580 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581

Table 5 Table 7
Performance comparison of the test Case-A1 considering capacitors as Comparison of the results for test Case-A1
constant reactive power and constant impedance load
Different methods Annual CPU
Test Case-A1 Peak power Capacitors Annual saving ($) time (s)
loss (kW) cost ($) saving ($)
Proposed GA considering 42173 199
Capacitor as constant 556.08 9738.0 42,209 criteria based on ` difference
reactive power load of best ®tness average ®tness'
Capacitor as constant 561.4 11088.0 40,198 of population
impedance load GA proposed by Sundhararajan 42,169 644
and Pahwa [18]
Simulated annealing (SA) 42,138 711
generations, the members of population get saturated and proposed by Chiang et al. [17]
beyond that only slight improvement in overall performance
is achieved at expense of lot of computational time. Hence
with consideration of this fact, it was observed that, instead
for obtaining the optimum values of shunt capacitors (®xed
of specifying the maximum of generations, if a secondary
and switched) in radial distribution system under varying
convergence criteria based on the difference between best
load conditions with consideration of voltage constraint.
®tness and average ®tness of population is used, it could
From the practical point of view, it has also been observed
reduce a lot of CPU time with only a slight reduction in
that consideration of capacitor as constant impedance load
overall saving. Therefore, a convergence criteria, `if the
in optimization problem itself makes the problem more
difference between best ®tness and average ®tness of popu-
realistic.
lation is less than 0.000001 than terminate the algorithm',
Of course, genetic control parameters (i.e. Pm, Pc and
has been used and the performance of this criteria was tested
population size) play an important role in the performance
for all four cases (Section 9.1).
of GA, and some permutations and combinations of these
Table 6 depicts the performance of the test Case-A1
parameters are required to be tested to get the best
(Section 9.1) for two different termination criteria. From
performance. But the results indicate that GA can provide
Table 6, we can observe that the convergence criteria is
approximate global optimum solution. The only problem in
acceptable considering the great reduction in CPU time at
GA is the requirement of higher CPU time, but it has been
the expense of only slight sacri®ce in overall savings. Simi-
observed that almost all of the CPU time is for objective
lar ®ndings were also observed for test Case-A2.
function evaluation. Genetic operations (i.e. selection,
Table 7 shows comparison of the results (test Case-A1) of
crossover and mutation) require very less CPU time. As
the proposed GA considering the convergence criteria based
the objective function evaluation for each string is indepen-
on the `difference of best ®tness and average ®tness' of
dent, suitability of use of parallel processors is feasible.
population with that of GA proposed by Sundhararajan
A new convergence criteria based on `the difference
and Pahwa [18] and simulated annealing proposed by Chiang
between best ®tness and average ®tness' has also been
et al. [17]. From Table 7, it is seen that annual savings using all
proposed. The new convergence criteria can provide not
the three methods are more or less same but the proposed GA
only suf®cient reduction in CPU time but also acceptable
takes very less CPU time as compared to other two methods.
accuracy in overall results.
Similar ®ndings were also observed for test Case-A2.

References
10. Conclusions
[1] Neagle NM, Samson DR. Loss reduction from capacitors installed on
Genetic algorithm (GA) has been successfully implemented primary feeders. AIEE Trans 1956;75:950±9.
[2] Maxwel M. The economic application of capacitors to distribution
Table 6 feeders. AIEE Trans 1960;79:353±9.
Performance comparison of test Case-A1 for two different convergence [3] Cook RF. Optimizing the application of shunt capacitors for reactive
criteria volt Ð ampere and loss reduction. AIEE Trans 1961;80:430±44.
[4] Schmill JV. Optimum size and location of shunt capacitors on distri-
Termination criteria Saving $ CPU bution feeders. IEEE Trans Power Apparatus Syst 1965;84:825±32.
of the genetic algorithm time (s) [5] Duran H. Optimum number, location and size of shunt capacitors in
(Test Case-A1) radial distribution feeders: a dynamic programming approach. IEEE
Trans Power Apparatus Syst 1968;87:1769±74.
Maximum number of 42,208 634 [6] Chang RE. Locating shunt capacitors on primary feeder for voltage
generation reached control and loss reduction. IEEE Trans Power Apparatus Syst 1969;
Difference between `best 42,173 199 88:1574±7.
®tness and average ®tness' [7] Ponnavaiko M, Prakasa Rao KS. Optimal choice of ®xed and switched
of the population capacitors on radial distributors by the method of local variations.
IEEE Trans Power Apparatus Syst 1983;102:1607±15.
 D. Das / Electrical Power and Energy Systems 24 (2002) 573±581 581

[8] Lee SH, Grainger JJ. Optimum placement of ®xed and switched Part-I, Part-II and Part-III. Trans Power Apparatus Syst
capacitors on primary distribution feeders. IEEE Trans Power Appa- 1985;104:3278±97.
ratus Syst 1981;100:345±52. [15] Baran ME, Wu FF. Optimal capacitor placement on radial distribution
[9] Grainger JJ, Lee SH. Optimum size and location of shunt capacitors systems. IEEE Trans Power Delivery 1989;4:725±34.
for reduction of losses on distribution feeders. IEEE Trans Power [16] Baran ME, Wu FF. Optimal sizing of capacitors placed on a radial
Apparatus Syst 1982;100:1105±16. distribution system. IEEE Trans Power Delivery 1989;4:735±43.
[10] Grainger JJ, Lee SH. Capacity release by shunt capacitor placement [17] Chiang HD, Wang JC, Cockings O, Shin HD. Optimal capacitor
on feeders: a new voltage dependent model. IEEE Trans Power Appa- placements in distribution systems, Part-I and Part-II. IEEE Trans
ratus Syst 1982;100:1236±44. Power Delivery 1990;5:634±64.
[11] Grainger JJ, Civanlar S, Lee SH. Optimal design and control scheme [18] Sundhararajan S, Pahwa A. Optimal selection of capacitors for radial
for compensation of distribution feeders. IEEE Trans Power Appara- distribution systems using a genetic algorithm. IEEE Trans Power
tus Syst 1983;102:3271±8. Syst 1994:1499±507.
[12] Grainger JJ, Civanlar S, Clinard KN. Optimal voltage dependent [19] Grefenst JJ. Optimization of control parameters for genetic algorithm.
continuous-time control of reactive power on primary feeders. IEEE IEEE Trans Syst, Man Cybernetics 1986;SMC-16:122±8.
Trans Power Syst 1984;103:2714±22. [20] Spillman R. Genetic algorithms, nature's way to search for the best.
[13] El-kib AA, Grainger JJ, Clinard KN, Gale LJ. Placement of ®xed and/ Dr Dobb's J 1993:26±30.
or non-simultaneously switched capacitors on unbalanced three-phase [21] Das D, Kothari DP, Kalam A. Simple and ef®cient method for load
feeders involving laterals. IEEE Trans Power Apparatus Syst 104, ¯ow solution of radial distribution networks. Int J Electr Power
3298±305. Energy Syst 1995;17(5):335±46.
[14] Grainger JJ, Civanlar S. Volt/Var control on distribution systems [22] Goldberg DE. Genetic algorithms in search, optimization and learn-
with lateral branches using shunt capacitors and voltage regulators: ing, a book. Reading, MA: Addisson-Wesley, 1989.