You are on page 1of 4

Hybrid Differential Evolution Particle Swarm

Optimization Algorithm for Reactive Power


Optimization
Shouzheng Wang, Lixin Ma, Dashuai Sun
Dept. of Electrical Engineering
Univ. of Shanghai for Sci.& Tech.
Shanghai, P.R.China
wlwarren@126.com, ma_eeepsi@usst.edu.cn

Abstract—Reactive power optimization is a mixed integer obtain optimal solutions to some real valued problems
nonlinear programming problem where metaheuristics efficiently [4, 5]. However, because of its no use of global
techniques have proven suitable for providing optimal solutions. information about the search space, DE may be trapped in local
In this paper, swarm and evolutionary algorithm have been optima. Also, basic PSO has drawbacks of prematurity, slow
applied for reactive power optimization. The objective of this search speed and low convergence accuracy [6]. Recently,
nonlinear optimization is minimization of system losses and many improved PSO algorithms have been proposed by
improvement of voltage profiles in a power system. A hybrid incorporating with DE, so as to explore better solutions. Some
differential evolution particle swarm optimization algorithm is hybrid DEPSO algorithms have been applied to reactive power
presented to obtain the global optimum. The proposed algorithm
optimization problem successfully [7, 8], however, so far no
is implemented on the IEEE 14-bus system. To validate the
conclusive conclusion has been reached in which hybrid
effectiveness of the algorithm, the simulation results are
compared with other optimization algorithms’. It is shown that DEPSO algorithm is better than the others.
the approach developed is feasible and efficient. In this paper a hybrid differential evolution particle swarm
optimization algorithm which is inspired by the paper [6], is
Keywords- reactive power optimization; differential evolution; developed to solve the reactive power optimization problem.
particle swarm optimization; hybrid differential evolution particle The proposed algorithm is implemented on the IEEE 14-bus
swarm optimization
system.

I. INTRODUCTION
II. PROBLEM FORMULATION
The reactive power optimization problem has a significant
influence on secure and economic operation of power systems. A. Objective Function
With the regulation of the voltage level of the generators, the Mathematically, the objective function of the reactive
taps of the transformer with OLTC and the switchable shunt power problem can be formulated as follows:
capacitor/reactor groups, and so on, the reactive power
optimization must improve system voltage profiles while NB
minimizing system losses at all times. min Ploss = ∑ G [V ij i
2
+ V j2 − 2ViV j cos(δ i − δ j )] .
Conventional optimization techniques such as nonlinear i =1, j∈i
programming, linear programming, and quadratic (1)
programming et al. have been used to solve the reactive power
optimization problem [1]. However, because of the non- Transformer tap-setting T and generator bus voltages VG are
differential and non-linearity nature of the reactive power control variables so they are self-restricted. Load bus voltages
optimization problem, majority of these techniques converge to VB and reactive power generations QG are state variables, which
a local optimum. In recent years, many new stochastic search are restricted by adding them as the quadratic penalty terms to
methods have been developed for the global optimization the objective function to form a generalized objective function
problem such as genetic algorithm [2], simulated annealing [3] as
and many others. These techniques search for the global or V − Vi lim 2
quasi-global optimum and the results reported were promising min f = Ploss + λV ( i ∑
Vi max − Vi min
)
and encouraging for further research in this direction. i∈ N B

Qi − Qi lim 2
Differential evolution (DE) and particle swarm
optimization (PSO) are population-based optimization
+ λQ ∑ (Q
i∈ N g i max − Qi min
) .
algorithms. Due to their excellent convergence characteristics (2)
and few control parameters, DE and PSO have been applied to

Sponsored by Innovation Foundation of Shanghai (Grant NO.08YZ100)

978-1-4244-4813-5/10/$25.00 ©2010 IEEE


where Ploss is the active power loss in the system; NB and Ng are III. HYBRID DIFFERENTIAL EVOLUTION PARTICLE SWARM
the sets of total buses and generator buses respectively; λV and OPTIMIZATION ALGORITHM
λQ are penalty coefficients of the constraint violations of the
voltages of PQ-buses and the reactive power of PV-buses A. Differential Evolution(DE)
respectively; Vi and Qi are the voltage magnitude of bus i and The main strategy of DE is to generate a new position for
the reactive power of PV-bus i respectively; Vimax ,Vimin ,Qimax an individual by calculating vector differences between other
,Qimin are the upper and lower limits of bus voltages and the randomly selected members of the population [4]. The
reactive power output of PV-buses respectively; Gij is the following are the main operators of DE algorithm.
mutual conductance between buses i and j; δi, δj are the voltage
phase angles of bus i and bus j respectively; and Vilim, Qilim are 1) Mutation: For each vector Xik (target vector), select three
expressed as: distinct vectors Xak, Xbk and Xck randomly from the current
population other than vector Xik. Generate a new population
⎧Vi max if Vi > Vi max vector on the formula
⎪ (3)
Vi lim = ⎨Vi min if Vi < Vi min ,
⎪V X i'k = X ak + F ( X bk − X ck ). (11)
⎩ i others
2) Crossover: Generate trial vectors applying the selected
⎧Qi max if Qi > Qi max crossover scheme. The following expression describes the

Qi lim = ⎨Qi min if Qi < Qi min . (4) crossover process
⎪Q others ⎧⎪ X i' k
⎩ i if rand ≤ C R
X i'' k = ⎨ . (12)
B. Constraints ⎪⎩ X ik otherwise
The minimization of the above objective function is subject 3) Selection: Evaluate the trial vector and decide whether
to a number of constraints: or not it will be part of the next generation by greedy strategy.
The following expression represents the selection method:
Pgi = Pdi − Vi ∑ V (G
j∈ N i
j ij cos δ ij + Bij sin δ ij ) , i ∈ N 0 (5) ⎧⎪ X i'' k
X ik = ⎨
if fitness ( X i'' k ) ≥ fitness ( X ik ) (13)
⎪⎩ X ik otherwise
Qgi = Qdi − Vi ∑V (G
j∈N i
j ij sin δ ij − Bij cos δ ij ) , i ∈ N PQ (6) F and CR are the control parameters. F guides the
amplitude of the influence of the difference vector and CR
represents the amount of the candidate solution that is used.
Vi min ≤ Vi ≤ Vi max i ∈ NB (7)
B. Particle Swarm Optimization(PSO)
In PSO, a swarm of particles flies through a D
Ti min ≤ Ti ≤ Ti max k ∈ NT (8) dimensional search space with a velocity which is dynamically
adjustable according to its own flying experience and its
companion’s. The best previous position of a particle is
(9) recorded and represented as pbest. The position of the best
Qgi min ≤ Qgi ≤ Qgi max i ∈ NG
particle among all the particles is represented as gbest. The
velocity and position of each particle in the k-th iteration of
Qci min ≤ Qci ≤ Qci max i ∈ NC (10) the swarm can be computed using the following formulas
which are proposed by [5]:

where N0, NPQ, NT, NG and NC are the sets of total buses vidk +1 = wvidk + c1r1 ( pbestidk − xidk ) + c2 r2 ( gbestdk − xidk ), (14)
excluding slack bus, PQ buses, transformer branches, generator
buses and capacitor buses, respectively; Pgi and Qgi are the
specified active and reactive power supply at bus i; Pdi and Qdi xidk +1 = xidk + vidk +1. (15)
are the specified active and reactive power demand at bus i; δij
is the voltage phase angle difference between buses i and j; Bij where r1 and r2 are uniformly random numbers between 0 and
is the susceptance between buses i and j; Tk is the tap position 1; vidk and xidk are the current velocity and position of
of transformer k; Qgi and Qci are the reactive power outputs of individual i in dimension d at iteration k respectively; pbestidk
generator at bus i and capacitor at bus j respectively; and Tkmin, is the previous position of individual i in dimension d; gbestdk
Tkmax, Qgimin, Qgimax, Qcimin, Qcimax are the lower and upper limits is the position of the best particle in the swarm in dimension d;
of transformer tap positions, reactive power output of c1 and c2 are the cognitive and social acceleration constants
generators, and reactive power output of capacitors respectively.
respectively.
The use of linearly decreasing inertia weight factor w has
provided improved performance in all the applications.
Suitable selection of the inertia weight provides a balance Step 1) Input parameters of the system and the algorithm.
between global and local exploration and exploitation by Initialize all the particles (velocity vi1and position xi1)
dynamically adjusting the velocity and it is calculated using randomly in the target population P1, and the positions of
(16). individuals zi1 in the extra population P2 in boundaries, and
wmax − wmin evaluate the fitness fitness(xi1) of each target particle, k=1.
w = wmax − ⋅ iter Step 2) Perform a Newton-Raphson power flow for every
itermax (16)
particle and calculate the fitness of pbest and gbest. Stop the
where itermax is the maximum number of iterations; iter is the algorithm if the terminal criterion is satisfied, otherwise
current iteration number; wmax, wmin are the maximum and k=k+1.
minimum inertia weight. Step 3) Calculate the average position pbestave,dk-1 and
The particle velocity is limited by the maximum value vmax. average velocity vave,dk-1 as (18) and (19). Update the velocity
If vmax is too high, then the particles might fly past good vik and position xik of each particle according to (17) and (15),
solutions. If vmax is too low, then the particles will take a long and evaluate the fitness fitness(xik).
time to reach the desired solution or not explore sufficiently Step 4) Implement prior crossover between xik and the
beyond local solutions. In this paper, 20% of the variable particle zik in P2 as (20). Compare the offspring yi’k with xik,
dynamic range is adopted as the limit of vmax. and replace xik with yi’k if yi’k is better.
Step 5) Generate the mutant vector Xi’k using (11).
C. Hybrid Differential Evolution Particle Swarm Step 6) Generate the trial vector Xi’’k from Xi’k and xik
Optimization Algorithm(PSOPDE) according to (12).
Inspired by advantages and disadvantages of DE and PSO Step 7) Compete Xi’’k with xik using (13). If
respectively, a hybrid of DE and PSO gives a new method of fitness(Xi’’k)≥fitness(xik), replace xik with Xi’’k. Otherwise
optimization called PSO with Prior-crossover DE (PSOPDE) replace yi’k with Xi’’k if fitness(Xi’’k) ≥fitness(yi’k) in (20). Go to
[6]. In PSOPDE, the average position and average velocity of
step 2.
particles also leads the particles to optimal solution besides the
pbest position and the gbest position. Thus the novel velocity IV. SIMULATION RESULTS
updated formula is described as (17). In order to improve local
search performance of PSO, DE operator is integrated with The proposed PSOPDE is applied for reactive power
PSO computation. Additionally, a new crossover operation as optimization on the IEEE 14-bus system in MATLAB 7.0
expressed in (20) between target population and an extra language and run on a Celeron(R) 2.66-G computer.
population is implemented before DE component, which The IEEE 14-bus system consists of 20 branches, four
increases the diversity of the particle swarm and helps to reach generators, and three transformers. The possible reactive power
the global optimum. source installation bus is 9. Bus 1 is selected as the slack bus
and designated to correct transmission loss changes. Bus 2, 3,
vidk +1 = wvidk + c1r1 ( pbestidk − xidk ) + c2 r2 ( gbestdk − xidk ) 6, 8 are PV buses, and the remaining buses of the test system
are PQ buses. Branch 5-6, 4-7, 4-9 are the transformer
k k k k
+ c3 r3 ( pbestave , id − xid ) + c4 r4 ( gbest ave, d − xid ) branches. Variable limits are similar to those in [9]. Reactive
power output of capacitor banks and transformer taps are
(17)
discrete variables with the change step of 0.05p.u. and
0.0125p.u. respectively. The basic apparent power of the test
N
system is set to 100MVA.
k
pbest ave ,d = ∑ pbest
i =1
k
id N (18)
To verify the performance of PSOPDE, the simulation
N
results are compared to those of DE and basic PSO. The three
k
vave ,d = ∑v
i =1
k
id N (19) methods have the same population size, 30, and maximum
iteration itermax=100. The parameters of PSOPDE are as
follows: wmax=0.9, wmin=0.4, c1=c2=2.0, c3=c4=0.01, CR’=0.5
⎧⎪ zidk if rand ≤ C R' (20) and F=0.15; and the crossover factor CR, rise by CR=CRmin+(
yid' k = ⎨ CRmax- CRmin)*(iter-1)/itermax, while CRmin=0.5 and CRmax=1.0 [6]
⎪⎩ xidk otherwise . The mutation factor F and the crossover factor CR of the DE
approach are set to 1 and 0.4 respectively. To the basic PSO, its
where r3, r4 are uniformly random numbers in the range (0, 1); acceleration constants are set to c1=c2=2.0, and its inertia
c3, c4 are the scale factors like c1 and c2; pbestave,dk and vave,dk weight factor is set to w=1.
are the average position and velocity respectively; zik is an
individual from the extra population that has the same size as Each method is executed 20 times when applied to the test
the initial target population; CR’ is the prior crossover factor system because of the random nature of all stochastic search
like CR. algorithms. The best, worst and average solutions of these
algorithms are presented in Table I. Obviously the PSOPDE
The brief procedure of implementation of PSOPDE for can obtain optimal results among the three methods, which
reactive power optimization can be written as follows: confirm that the PSOPDE is well capable of finding the global
or quasi-global optimum solution. It can be seen that the 0.165

average power losses reduction of 10.937% (from 0.13770p.u. 0.16


DE
PSO

to 0.12264p.u.) was accomplished by PSOPDE while DE 0.155


PSOPDE

achieved 8.214% (from 0.13770p.u. to 0.12639p.u.) and basic 0.15

Power Losses(p.u.)
PSO achieved 10.653% (from 0.13770p.u. to 0.12302p.u.). 0.145

Meanwhile, PSOPDE shows considerable performance by 0.14

keeping the difference between the best and worst solutions


0.135

0.13

within 0.1%, while this difference for DE and PSO is 4.3% and 0.125

1.0% respectively. It is shown that the convergence 0.12


0 10 20 30 40 50 60 70 80 90 100

performance of PSOPDE is stable and robust. Number of Iterations

However, on the one hand, according to the average Fig.1 Constringency curve with DE, PSO and PSOPDE Methods for
execution time presented in Table I, PSOPDE takes the most IEEE 14-bus system
CPU execution time among the three algorithms under the
same iterations. Because there are DE operators and extra
population updates operators in the PSOPDE method, from the
High and Low Voltage Profile Limits
1.1
simulation most time is spent on the fitness evaluation for those
operators, which needs to run an AC load flow in the test 1.05
system. On the other hand, from the convergence

Voltage (p.u.)
characteristics of the three methods depicted in Fig.1, it is clear 1

that the active power loss is drastically reduced and the


solution by PSOPDE is converged to high quality solutions at 0.95

the early iterations. The results proved that the proposed Initial Voltage Profile
PSOPDE algorithm provides higher-quality solution with 0.9 Final Voltage Profile(DE)
Final Voltage Profile(PSO)
smaller iterations than the other two methods. Final Voltage Profile(PSOPDE)
0.85
The results of the voltage profile corrections for DE, basic
2 4 6 8 10 12 14
Bus Number

PSO and PSOPDE methods are comparatively presented in Fig. 2 Voltage Profile Correction for DE, PSO and PSOPDE
Fig.2. For this case, the three algorithms brought the bus
voltages within the voltage limits. Also, PSOPDE methods can
REFERENCES
keep the bus voltages at higher level than the other two
methods while satisfying voltage constraints and minimizing [1] J.A.Momoh, M.E.El-Hawary and R.Adapa, “A Review of Selected
Optimal Power Flow Literature to 1993 Part I&II”, IEEE Transactions
the system losses. on Power Systems, Vol.14, No.1, pp.96-111, February 1999.
ƒ V. CONCLUSION [2] K.Iba, “Reactive power optimization by genetic algorithm,” IEEE
Transactions on Power Systems, vol.9, pp.685-692, 1994.
This paper has presented and compared three algorithms [3] C. A. Roa-Sepulveda, B. J. Pavez-Lazo, “A solution to the optimal
based on swarm intelligence and evolutionary techniques for power flow using simulated annealing,” International Journal of
solving the reactive power optimization problem. Case studies Electrical Power & Energy Systems, vol.25, no.1, pp.47-57, January
on the IEEE 14-bus system illustrate the effectiveness of these 2003.
algorithms in terms of the quality of the solutions found and [4] R.Stone, K.Price, “Differential evolution-a simple and efficient heuristic
for global optimization over continuous spaces,” Global Optimization,
their convergence characteristics. All three algorithms are able vol.11, pp.341-359, 1997.
to successfully restore the bus voltages to prescribed limits [5] R.C. Eberhart, Y.Shi, “Particle swarm optimization: developments,
while lowering the system active power losses. It is shown by applications and resourses,” Proceedings of the congress on
averaging the results over a multitude of trial runs that Evolutionary Computation, vol.1, pp.81-86, 2001.
PSOPDE indeed outperforms DE or basic PSO method on this [6] Wei Xu, Xingsheng Gu, “A hybrid particle swarm optimization
problem when comparing power loss reduction and number of approach with prior crossover differential evolution,” 2009 World
iterations required to achieve. It should be noted that the Summit on Genetic and Evolutionary Computation (2009 GEC Summit),
pp.671-677, June 2009.
optimization strategy is general and can be applied to other
power system optimization problems as well. [7] Bo Zhao, Yi-Jia Cao, “A multi-agent particle swarm optimization
algorithm for reactive power optimization,” Proceedings of the CSEE,
vol.25, pp.1-7, March 2005.
TABLE I. COMPARISON OF BEST, WORST AND AVERAGE VALUES FOR [8] Jia Liu, Dan Li, Li-qun Gao and Li-xin Song, “Vector evaluated
DIFFERENT METHODS adaptive particle swarm optimization algorithm for multi-objective
reactive power optimization,” Proceedings of the CSEE, vol.28, no.31,
Power losses (p.u.) Average pp.22-28, November 2008.
Different
Execution
Methods Best Worst Average [9] Jin-chang Duan, “Applications of the modified genetic algorithm in
Time(seconds)
reactive power optimization in power system,” Advances of Power
DE 0.12407 0.12956 0.12639 96.295 System&Hydroelectric Engineering, vol.24, no.6, pp.15-20, June 2008.
PSO 0.12271 0.12382 0.12302 45.297

PSOPDE 0.12263 0.12273 0.12264 164.208

You might also like