A Multiagent-Based Particle Swarm Optimization Approach For Optimal Reactive Power Dispatch
A Multiagent-Based Particle Swarm Optimization Approach For Optimal Reactive Power Dispatch
A Multiagent-Based Particle Swarm Optimization Approach For Optimal Reactive Power Dispatch
2, MAY 2005
AbstractReactive power dispatch in power systems is a com- Set of numbers of transformer branches.
plex combinatorial optimization problem involving nonlinear func- Set of numbers of buses on which voltages outside
tions having multiple local minima and nonlinear and discontin- limits.
uous constraints. In this paper, a solution to the reactive power dis-
patch problem with a novel particle swarm optimization approach Demanded active power at bus (p.u.).
based on multiagent systems (MAPSO) is presented. This method Injected active power at bus (p.u.).
integrates the multiagent system (MAS) and the particle swarm op- Active power loss in branch (p.u.).
timization (PSO) algorithm. An agent in MAPSO represents a par- Power flow in branch (p.u.).
ticle to PSO and a candidate solution to the optimization problem. Injected active power at slack bus (p.u.).
All agents live in a lattice-like environment, with each agent fixed
on a lattice point. In order to obtain optimal solution quickly, each Reactive power source installation at bus (p.u.).
agent competes and cooperates with its neighbors, and it can also Demanded reactive power at bus (p.u.).
learn by using its knowledge. Making use of these agentagent in- Injected reactive power at bus (p.u.).
teractions and evolution mechanism of PSO, MAPSO realizes the Tap position of transformer .
purpose of optimizing the value of objective function. MAPSO ap- Voltage magnitude of bus (p.u.).
plied to optimal reactive power dispatch is evaluated on an IEEE
30-bus power system and a practical 118-bus power system. Simu- Voltage vectors of buses (p.u.).
lation results show that the proposed approach converges to better Voltage vectors of buses (p.u.).
solutions much faster than the earlier reported approaches. The
optimization strategy is general and can be used to solve other I. INTRODUCTION
power system optimization problems as well.
Index TermsMultiagent system, particle swarm optimization
(PSO), power system, reactive power dispatch. T HE reactive power dispatch problem has a significant
influence on secure and economic operation of power
systems. Reactive power optimization is a subproblem of the
optimal power-flow (OPF) calculation, which determines all
NOMENCLATURE kinds of controllable variables, such as reactive-power outputs
Voltage angle difference between buses and (rad). of generators and static reactive power compensators, tap ratios
Transfer susceptance between bus and (p.u.). of transformers, outputs of shunt capacitors/reactors, etc., and
Active power loss in network (p.u.). minimizes transmission losses or other appropriate objective
Transfer conductance between bus and (p.u.). functions, while satisfying a given set of physical and oper-
Conductance of branch (p.u.). ating constraints. Since transformer tap ratios and outputs of
Set of numbers of total buses excluding slack bus. shunt capacitors/reactors have a discrete nature, while reactive
Set of numbers of total buses. power outputs of generators and static VAR compensators,
Set of numbers of possible reactive power source bus-voltage magnitudes, and angles are, on the other hand,
installation buses. continuous variables, the reactive power optimization problem
Set of numbers of power demand buses. can be exactly formulated using a mixed-integer/nonlinear pro-
Set of numbers of network branches. gramming model, i.e., cast as a nonlinear optimization problem
Set of numbers of generator buses. with a mixture of discrete and continuous variables.
Set of numbers of buses adjacent to bus , including Up to now, a number of techniques ranging from classical
bus . techniques like gradient-based optimization algorithms to var-
Set of numbers of buses. ious mathematical programming techniques have been applied
Set of numbers of buses. to solve this problem [1][4]. Recently, due to the basic effi-
Set of numbers of buses on which injected reactive ciency of interior-point methods, which offer fast convergence
power outside limits. and convenience in handling inequality constraints in compar-
ison with other methods, interior-point linear programming
[5], quadratic programming [6], and nonlinear programming
Manuscript received August 13, 2004; revised December 27, 2004. This work [7] methods have been widely used to solve the OPF problem
is supported by the Outstanding Young Scholars Fund (no. 60225006) and In-
novative Research Group Fund of Natural Science Foundation of China. Paper of large-scale power systems. However, these techniques
no. TPWRS-00444-2004. have severe limitations in handling nonlinear, discontinuous
The authors are with the College of Electrical Engineering, Zhejiang Uni- functions and constraints, and function having multiple local
versity, Hangzhou, Zhejiang 310027, China (e-mail: zhaobozju@zju.edu.cn;
chuangxinguo@vip.sina.com; yijiacao@zju.edu.cn). minima. Unfortunately, the original reactive power problem
Digital Object Identifier 10.1109/TPWRS.2005.846064 does have these properties. In all these efforts some or the
0885-8950/$20.00 2005 IEEE
ZHAO et al.: A MULTIAGENT-BASED PSO APPROACH 1071
other simplification has been done to get over the inherent optimization problem. All agents live in a lattice-like envi-
limitations of the solution technique. The combinatorial-search ronment, with each agent fixed on a lattice point. In order to
approaches, branch-and-bound and cutting-plane algorithms, obtain optimal solution quickly, they compete and cooperate
are usually used to solve the mixed-integer programming model with their neighbors, and they can also use knowledge. Making
[8]. However, these methods are nonpolynomial and all suffer use of these agentagent interactions and evolution mechanism
from the so-called problem of curse of dimensionality for of PSO in a lattice-like environment, the proposed method can
large-scale applications, making them unsuitable for large-scale find high-quality solutions reliably with the faster convergence
OPF problems. To overcome the drawback of these algorithms, characteristics in a reasonably good computation time. MAPSO
several efficient algorithms have been proposed. Aoki et al. applied for optimal reactive power is evaluated on an IEEE
[9] addressed the issue of discrete variables by an approxima- 30-bus power system and a practical 118-bus power system.
tion-search method for recursive mixed-integer programming Simulation results show that the proposed approach converges
in solving large-scale VAR planning problems. Bakirtzis and to better solutions much faster than the earlier reported ap-
Meliopous [10] proposed a linear-programming methodology proaches.
to handle the discrete shunt capacitors/reactors in an optimiza- The rest of this paper is organized as follows: Section II de-
tion problem by using Driebecks penalty algorithm. Liu et al. scribes mathematical formulation of optimal reactive power dis-
[8] proposed a penalty-based discretization algorithm to handle patch. Section III describes MAPSO in detail. Simulation results
the discreteness of shunt capacitors/reactors during the solution and comparison with other approaches are given in Section IV.
process of a Newton OPF method without combinatorial search, Finally, conclusions are presented in Section V.
which has been implemented in a production-grade Newton
OPF program and tested on actual power networks. II. PROBLEM FORMULATION
In the last decade, many new stochastic search methods have The objective of the reactive power dispatch is to minimize
been developed for the global optimization problems, such the active power loss in the transmission network, which can be
as genetic algorithms, evolutionary programming and particle described as follows:
swarm optimization. Particle swarm optimization (PSO) is
one of the evolutionary computation techniques [11]. It was
developed through simulation of a simplified social system,
and has been found to be robust in solving continuous non- (1)
linear optimization problems. The PSO technique can generate where ; ; . The symbols of the above
high-quality solutions within shorter calculation time and have equation and in the following context are given in the Nomen-
more stable convergence characteristic than other stochastic clature section. The minimization of the above function is sub-
methods. Although the PSO seems to be sensitive to the tuning ject to a number of constraints:
of some weights or parameters, many researches are still
in progress to prove its potential in solving complex power (2)
system problems [12], [13]. Kassabalidis et al. [14] introduced
dynamic security border identification using enhanced PSO.
Naka et al. [15] proposed a hybrid PSO for distribution state
estimation. It has been found that the PSO quickly finds the (3)
high-quality optimal solution for many power system optimiza-
tion problems. Generally, PSO has a more global searching and
ability at the beginning of the run and a local search near the (4)
end of the run. Therefore, while solving problems with more
local optima, there are more possibilities for the PSO to explore (5)
local optima at the end of run. However, the reactive power (6)
optimization problem does have these properties in itself. For (7)
these reasons, a reliable global approach to power system (8)
optimization problems would be of considerable value to power
engineering community. where power flow equations are used as equality constraints, re-
Recently, agent-based computation has been studied in the active power source installation restrictions, reactive generation
field of distributed artificial intelligence [16] and has been restrictions, transformer tap-setting restrictions, bus voltage re-
widely used in other branches of computer science [17]. strictions and power flow of each branch are used as inequality
Problem solving is an area that many multiagent-based ap- constraints.
plications are concerned with. Liu et al. [18] introduced an In the most of the nonlinear optimization problems, the con-
application of distributed techniques for solving constraint straints are considered by generalizing the objective function
satisfaction problem. Enlightened by multiagent system and using penalty terms. In the reactive power dispatch problem, the
PSO, this paper integrates multiagent system and PSO to form generator bus voltages and , the tap position of trans-
a multiagent-based PSO approach (MAPSO), for solving the former , and the amount of the reactive power source installa-
reactive power optimization problem. In MAPSO, an agent tion are control variables which are self-constrained. Volt-
represents a particle to PSO and a candidate solution to the ages of bus and injected reactive power of bus
1072 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 2, MAY 2005
are constrained by adding them as penalty terms to the ob- Suitable selection of inertia weight provides a balance
jective function (1). The above problem is generalized as fol- between global and local explorations. In general, the inertia
lows: weight is set according to the following equation:
(15)
(14) C. MAPSO
In this study, MAS and PSO are integrated to form the pro-
posed MAPSO method for solving reactive power optimiza-
where and . tion dispatch. In MAPSO, an agent represents not only a candi-
ZHAO et al.: A MULTIAGENT-BASED PSO APPROACH 1073
From (17), each agent has four neighbors. They form a little (20)
local environment in which agent can only sense.
4) Behavioral Strategies for Agents: To quickly and accu- Then, is determined by
rately achieve its purposes, each agent has some behaviors. In
MAPSO, each agent firstly competes and cooperates with its
neighbors to diffuse its useful information to the whole envi-
ronment, and it can also use evolution mechanism of PSO and (21)
1074 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 2, MAY 2005
(22)
(24)
otherwise
ZHAO et al.: A MULTIAGENT-BASED PSO APPROACH 1075
For discrete variables of the th particle , the most straight- Step 9) If one of the stopping criteria is satisfied then go to
forward way is to use the indices of the set of discrete variables Step 10. Otherwise, go to Step 4.
with elements Step 10) Output the agent with the minimum fitness value in
the last generation.
(25)
IV. NUMERICAL RESULTS
Let denote the continuous variables with elements
To verify the effectiveness and efficiency of the proposed
(26) MAPSO based reactive power optimization approach, the IEEE
30-bus power system and a practical 118-bus area power system
then particle is denoted by . For particle , the are used as the test systems. The MAPSO has been implemented
index value of the discrete variable is then optimized in- in Matlab 6.5 programming language and numerical tests are
stead of the discrete value of the variable directly. In the pop- carried on a Pentium IV 2.0-G computer.
ulation, the indices of the discrete variables of the th particle Some parameters must be assigned before MAPSO is used
should be the float point variables before truncation. That is, to solve reactive power optimization dispatch. is
, is the number of discrete variables. Hence, equivalent to the population size in traditional PSO, so
the objective function of the th particle can be expressed as can be chosen from 5 to 10. In self-learning operator, with con-
follows: sideration of the computational cost, it is better to let be
smaller than 5 and range from 5 to 10. In all study sys-
(27)
tems, the following MAPSO parameters are used: ,
where i.e., the members of a particle are 36, ,
, , , inertia weight factor
, acceleration constant , both
penalty factors in (9) are chosen, .
(28)
where and denote the feasible subsets A. IEEE 30-Bus Power System
of continuous and discrete variables of particle , respectively. The IEEE 30-bus system is shown in Fig. 3 and the system
denotes the greatest integer less than the real value . data and operating conditions are given in the [23]. The network
consists of 48 branches, six generator-buses, and 22 load-buses.
E. Implementation of MAPSO for Reactive Power Four branches, (6,9), (6,10), (4,12) and (27,28), are under load
Optimal Dispatch tap setting transformer branches. The possible reactive power
In MAPSO, many different operators are utilized to simulate source installation buses are 3, 10 and 24. Six buses are selected
the behaviors of agents, and realize their purposes. In order to as -buses and -buses as follows: -buses: bus 2, 5, 8,
reduce the computational cost, the self-learning operator is only 11,13, -bus: bus 1. The others are -buses. The variable
performed on the agent with the minimum fitness value in each limits are listed in Table I. The transformer taps and the reactive
generation, but it has an important effect on the performance power source installation are discrete variables with the changes
of MAPSO. The proposed method can quickly and accurately step of 0.01 p.u.
converge to the optimal solutions for reactive power optimal To demonstrate the superiority of the proposed MAPSO
dispatch. The details of the overall algorithm are as follows. approach, simulation results have been compared with various
Step 1) Input parameters of system and algorithm, and techniques available in literature, namely, standard genetic
specify the lower and upper boundaries of each algorithm (SGA), adaptive GA (AGA) in [23], EP method
variable. presented in [24], Broydens nonlinear programming method
Step 2) Generate a lattice-like environment , and initialize [25] and PSO method. The initial conditions for all the methods
randomly each agent. are same and are given as follows:
Step 3) Evaluate the fitness value of each agent based on the
Newton-Raphson power flow analysis results and p.u. p.u.
the proposed mixed-variable handing methods.
Step 4) Update the time counter . The initial generator bus voltages and transformer taps are
Step 5) Perform the neighborhood competition and cooper- set to 1.0. The total generations and power losses are obtained
ation operator on each agent. as follows:
Step 6) Execute the PSO operator and further adjust its po-
sition in the search space on each agent according p.u. p.u.
to (12) and (13). p.u. p.u.
Step 7) Evaluate the fitness value of each agent based on the
Newton-Raphson power flow analysis results and The voltages outside the limits on three -buses are given
the proposed mixed-variable handing methods. as follows:
Step 8) Find the best agent with the minimum fitness value,
and then perform the self-learning operator.
1076 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 2, MAY 2005
TABLE I TABLE II
VARIABLE LIMITS (P.U.) COMPARISON OF OPTIMAL TRANSMISSION LOSS FOR DIFFERENT
METHODS (P.U.)
TABLE IV
VALUES OF CONTROL VARIABLES AFTER OPTIMIZATION BY SGA,
PSO, AND MAPSO (P.U)
Fig. 5. Optimization procedure by SGA, PSO, and MAPSO for the practical
118-bus system.
faster convergence properties The performance of the proposed [19] R. C. Eberhart and Y. Shi, Comparing inertia weights and constriction
method demonstrated through its evaluation on the IEEE 30-bus factors in particle swarm optimization, in Proc. Congr. Evol. Comput.,
2000, pp. 8488.
power system and a practical 118-bus power system shows that [20] M. Wooldridge, An Introduction to MultiAgent System. New York:
MAPSO is able to undertake global search with a fast conver- Wiley, 2002.
gence rate and a feature of robust computation. From the sim- [21] W. Z. Zhong et al., A multiagent genetic algorithm for global numerical
optimization, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 34, pp.
ulation study, it has been found that MAPSO converges to the 11281141, Apr. 2004.
global optimum. The optimization strategy is general and can [22] S. A. Kazarlis, S. E. Papadakis, J. B. Theocharis, and V. Petridis, Mi-
be used to other power system optimization problems as well. crogenetic algorithms as generalized hill-climbing operators for GA op-
timization, IEEE Trans. Evol. Comput., vol. 5, no. 3, pp. 204217, Jun.
2001.
REFERENCES [23] Q. H. Wu, Y. J. Cao, and J. Y. Wen, Optimal reactive power dispatch
using an adaptive genetic algorithm, Int. J. Elect. Power Energy Syst.,
[1] K. Y. Lee, Y. M. Park, and J. L. Ortiz, A united approach to optimal vol. 20, pp. 563569, Aug. 1998.
real and reactive power dispatch, IEEE Trans. Power Syst., vol. 1, no. [24] L. L. Lai and J. T. Ma, Application of evolutionary programming to
2, pp. 11471153, May 1985. reactive power planning Comparison with nonlinear programming ap-
[2] Y. Y. Hong, D. I. Sun, S. Y. Lin, and C. J. Lin, Multi-year multi-case proach, IEEE Trans. Power Syst., vol. 12, no. 1, pp. 198204, Feb. 1997.
optimal AVR planning, IEEE Trans. Power Syst., vol. 5, no. 4, pp. [25] D. B. Das and C. Patvardhan, Reactive power dispatch with a hybrid
12941301, Nov. 1990. stochastic search technique, Int. J. Elect. Power Energy Syst., vol. 24,
[3] N. Deeb and S. M. Shaidepour, Linear reactive power optimization in pp. 731736, Sep. 2002.
a large power network using the decomposition approach, IEEE Trans. [26] R. Zimmerman and D. Gan, MATPOWER: A Matlab Power System Sim-
Power Syst., vol. 5, no. 2, pp. 428435, May 1990. ulation Package. Ithaca, NY: Cornell Univ. Press, 1997.
[4] J. R. S. Manlovani and A. V. Garcia, A heuristic method for reactive [27] Y. W. Zhong and J. G. Yang, A genetic algorithm for tasks scheduling
power planning, IEEE Trans. Power Syst., vol. 11, no. 1, pp. 6874, in parallel multiprocessor systems, in Proc. Mach. Learning Cybern.,
Feb. 1995. vol. 3, Nov. 2003, pp. 17851790.
[5] S. Granville, Optimal reactive dispatch through interior point methods,
IEEE Trans. Power Syst., vol. 9, no. 1, pp. 136146, Feb. 1994.
[6] J. A. Momoh, S. X. Guo, E. C. Ogbuobiri, and R. Adapa, The quadratic
interior point method solving power system optimization problems,
IEEE Trans. Power Syst., vol. 9, no. 3, pp. 13271336, Aug. 1994.
[7] Y. C. Wu, A. S. Debs, and R. E. Marsten, A direct nonlinear predictor- B. Zhao was born in 1977 in Guizhou, China. He received the B.Sc. and M.Sc.
corrector primal-dual interior point algorithm for optimal power flows, degrees in the field of electrical engineering from Guizhou University, Guiyang,
IEEE Trans. Power Syst., vol. 9, no. 2, pp. 876883, May 1994. China, in 1999 and 2002, respectively. He is currently working toward the Ph.D
[8] W. H. E. Liu, A. D. Papalexopoulos, and W. F. Tinney, Discrete shunt degree at Zhejiang University, Zhejiang, China.
controls in a Newton optimal power flow, IEEE Trans. Power Syst., vol. His research interest is application of intelligent systems and evolutionary
7, no. 4, pp. 15091518, Nov. 1992. computation to power system planning and operations.
[9] K. Aoki, M. Fan, and A. Nishikori, Optimal var planning by approxi-
mation method for recursive mixed-integer linear programming, IEEE
Trans. Power Syst., vol. 3, no. 4, pp. 17411747, Nov. 1988.
[10] A. G. Bakirtzis and A. P. Meliopous, Incorporation of switching op-
erations in power system corrective control computation, IEEE Trans. C. X. Guo was born in 1969 in Hubei, China. He received the B.Sc. and Ph.D.
Power Syst., vol. 2, no. 3, pp. 669676, Aug. 1987. degrees in the field of electrical engineering from Huazhong University of Sci-
[11] J. Kennedy and R. Eberhart, Particle swarm optimization, in Proc. ence and Technology (HUST), Wuhan, China, in 1992 and 1997, respectively.
IEEE Int. Conf. Neural Netw., vol. 4, Nov. 1995, pp. 19421948. From 1997 to 2003, he worked as a Senior Engineer and General Manager
[12] Z. L. Gaing, Particle swarm optimization to solving the economic dis- with the Department of Electric Power System Automation, Dongfang Elec-
patch considering the generator constraints, IEEE Trans. Power Syst., tronics Company, Yantai, China. Currently, he is a post-doctoral research fellow
vol. 18, no. 3, pp. 11871195, Aug.. 2003. at Zhejiang University, Zhejiang, China. His research interest is in power system
[13] H. Yoshida et al., A particle swarm optimization for reactive power and operation and automation.
voltage control considering voltage security assessment, IEEE Trans.
Power Syst., vol. 15, no. 4, pp. 12321239, Nov. 2000.
[14] I. N. Kassabalidis et al., Dynamic security border identification using
enhanced particle swarm optimization, IEEE Trans. Power Syst., vol.
17, no. 3, pp. 723729, Aug. 2002.
[15] S. Naka et al., A hybrid particle swarm optimization for distribution Y. J. Cao (M97) received the B.Sc. degree from Xian Jiaotong University,
state estimation, IEEE Trans. Power Syst., vol. 18, no. 1, pp. 6068, Xian, China, in 1988 and the M.Sc. and Ph.D. degrees from Huazhong Uni-
Feb. 2003. versity of Science and Technology (HUST), Wuhan, China, in 1991 and 1994,
[16] J. Liu, Autonomous Agents and Multiagent Systems: Explorations in respectively.
Learning Self-Organization, and Adaptive Computation, Singapore: From September 1994 to April 2000, he worked as a visiting research fellow
World Scientific, 2001. at Loughborough University, Loughborough, U.K., Liverpool University, Liver-
[17] J. Liu, Y. Y. Tang, and Y. C. Cao, An evolutionary autonomous agents pool, U.K., and University of the West England, Bristol, U.K. He was a Full Pro-
approach to image feature extraction, IEEE Trans. Evol. Comput., vol. fessor with HUST from 2000 to 2001. Currently, he is a Full Professor with Zhe-
1, no. 1, pp. 141158, Feb. 1997. jiang University, Zhejiang, China. His research interests are in power system sta-
[18] J. Liu, H. Jing, and Y. Y. Tang, Multiagent oriented constraint satisfac- bility control and application of evolutionary computation and intelligent agents
tion, Artificial Intell., vol. 136, pp. 101144, Jan. 2002. in power systems.