Electric Power Systems Research 77 (2007) 16541664

Environmental/economic power dispatch using a fuzzied multi-objective particle swarm optimization algorithm
Lingfeng Wang , Chanan Singh
Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX 77843, USA Received 29 August 2006; received in revised form 20 November 2006; accepted 21 November 2006 Available online 28 December 2006

Abstract The environmental issues that arise from the pollutant emissions produced by fossil-fueled electric power plants have become a matter of concern more recently. The conventional economic power dispatch cannot meet the environmental protection requirements, since it only considers minimizing the total fuel cost. The multi-objective generation dispatch in electric power systems treats economic and emission impact as competing objectives, which requires some reasonable tradeoff among objectives to reach an optimal solution. In this paper, a fuzzied multi-objective particle swarm optimization (FMOPSO) algorithm is proposed and implemented to dispatch the electric power considering both economic and environmental issues. The effectiveness of the proposed approach is demonstrated by comparing its performance with other approaches including weighted aggregation (WA) and evolutionary multi-objective optimization algorithms. All the simulations are conducted based on a typical test power system. 2006 Elsevier B.V. All rights reserved.
Keywords: Multi-objective optimization; Particle swarm optimization; Environmental/economic power dispatch; Local/global search; Constraint satisfaction

1. Introduction In recent years, with an increasing awareness of the environmental pollution caused by thermal power plants, limiting emission of pollutants is becoming a crucial issue in economic power dispatch. Power plants using fossil fuels as their major energy source generate particulates and gaseous pollutants including carbon dioxide, oxides of sulphur, and oxides of nitrogen. This may violate the ever-tighter environmental protection regulation since the excessive amount of emissions pollutants inevitably causes detrimental ecological effects. Emissions may be reduced by utilizing fuels with lower emission. Fuel switching is dependant on the price and availability of low-pollution fuel. Furthermore, switching from one fuel to another has many implications such as job losses in coal mining areas. Thus, it is an expensive solution. As an alternative, environmental/economic dispatch (EED) is proposed as a viable solution since it is able to reduce the pollutant emissions with lower generation cost. Most previous economic power dispatch dealt with the case of each generating unit having a single cost function. However,
Corresponding author. E-mail addresses: l.f.wang@ieee.org (L. Wang), singh@ece.tamu.edu (C. Singh). 0378-7796/$ see front matter 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.epsr.2006.11.012

certain practical thermal units use different fuels like coal, natural gas and oil, which have different heat contents. These units are faced with the problem of deciding the most economical fuel to burn. The multiple fuel options lead to piecewise quadratic cost functions (PQCF). Lin and Viviani [20] proposed a hierarchical approach for economic dispatch problems with PQCF. A linear programming based optimization procedure was proposed in [9], where one objective is considered at a time. Rifaat [23] developed the model for economic dispatch of combined cycle cogeneration units considering environmental constraints and used nonlinear programming technique to handle it. Park et al. [22] employed Hopeld neural networks to treat the economic dispatch problems with PQCF. A revised adaptive Hopeld neural network scheme was developed to deal with this problem by Lee et al. [19]. Jayabarathi et al. [14] applied evolutionary programming techniques to solve this type of problems. Later, optimization models were formulated to minimize or maximize a single scalar objective function. For instance, Gaing [11] deployed particle swarm optimization (PSO) to implement the economic power dispatch considering generator constraints. However, it is more natural for the analysts to consider multiple objectives simultaneously. The power system can also operate most economically and incur less environmental problems when optimized with respect to several criteria under a bunch

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

1655

of necessary constraints. Very often, the multi-objective problems are solved by nding non-inferior (i.e., Pareto-optimal, non-dominated) solutions. Prom the qualitative perspective, a non-inferior solution in a multi-objective problem is one where any improvement of one objective function can be attained only at the expense of another. The commonly used methods of producing these non-inferior solutions are the -constraint and weighted minimax methods. A wide range of new techniques is being used these years to tackle the EED problem in a more effective fashion. For instance, Dhillon et al. [8] handled EED using a stochastic approach considering various uncertainties. Song et al. [27] developed genetic algorithms (GAs) for EED, whose parameters are adaptively tuned by fuzzy logic controllers. RoaSepulveda and Herrera [24] dealt with the EED problem using an improved decision tree technique. A comprehensive comparison study regarding genetic algorithm, micro-GA, and evolutionary programming in solving EED was reported by Venkatesh et al. [30]. Abido [13] used multi-objective evolutionary algorithms to dispatch the electric power. Ah King [4] dealt with EED using both deterministic and stochastic evolutionary approaches. Neural-fuzzy techniques were used in [13] for multi-objective generation dispatch. It can be seen that recently the meta-heuristic optimization methods have been signicantly used in EED primarily due to their nice feature of population-based search. Particle swarm optimization is such a technique. We adopt PSO to handle the complexity and nonlinearity of the problem. PSO has several key advantages over other existing optimization techniques in terms of simplicity, convergence speed, and robustness [5,18]. PSO is easy to implement in computer simulations using basic mathematical and logic operations, since its working mechanism only involves two fundamental updating rules. PSO also has fewer operators to adjust in the implementation, and it can be exibly combined with other optimization techniques to build a hybrid algorithm. The mechanism of PSO facilitates a better convergence performance than some other optimization procedures like genetic algorithms, which have computationally expensive evolutionary operations such as crossover and mutation. Unlike the traditional optimization algorithms, PSO is a derivative-free algorithm and thus it is especially effective in dealing with complex and nonlinear problems. PSO is more robust to deal with such problems, since it is less sensitive to the nature of the objective function in terms of convexity and continuity, and the inner working of PSO helps to escape local minima. The robustness of PSO can also be reected by its less sensitivity to the optimizer parameters as well as the initial solutions to start its iteration process. Furthermore, PSO avoids the so-called curse of dimensionality that the dynamic programming (DP) suffers in handling practical economic dispatch problems. Meanwhile, PSO has also certain limitations that should be taken care of in dealing with economic dispatch problems. It still has some dependency on initial points, and the most effective combination of different parameters may need some time to tune. Also the nal outputs have some stochastic characteristic. PSO

is being used in diverse optimization problems. However, it was seldom used in multi-objective optimization problems (MOPs). In this paper, a novel fuzzied multi-objective particle swarm optimization (FMOPSO) algorithm is proposed to deal with the environmental/economic dispatch problem. The novelty of the developed algorithm can be demonstrated by the development of several distribution preservation mechanisms for dealing with multi-objective optimization case. The remainder of the paper is organized as follows. Section 2 formulates the target EED problem including its multiple objectives as well as design constraints. Section 3 presents the inner working of particle swarm optimization algorithms. The major concepts in multi-objective optimization problem are introduced in Section 4. The proposed FMOPSO algorithm is discussed in detail in Section 5. The simulation results as well as evaluation work are presented in Section 6. Finally, conclusion is given and future research directions are suggested. 2. Problem statement The typical optimal environmental/economic power dispatch problem can be formulated as a bi-criteria optimization model. The two conicting objectives, i.e., fuel cost and pollutants emission, should be minimized simultaneously while fullling certain system constraints. This problem is formulated in this section. 2.1. Problem objectives Objective 1: Minimization of fuel cost The generators cost curves are represented by quadratic functions with sine components. The superimposed sine components represent the rippling effects produced by the steam admission valve openings [3]. The total $/h fuel cost F(PG ) can be represented as follows:
M

F (PG ) =
i=1

2 min ai + bi PGi + ci PGi + |di sin[ei (PGi PGi )]|

(2.1)

where M is the number of generators committed to the operating system, ai , bi , ci , di , ei are the cost coefcients of the i-th generator, and PGi is the real power output of the ith generator. PG is the vector of real power outputs of generators and dened as PG = [PG1 , PG2 , . . . , PGM ] (2.2)

Objective 2: Minimization of pollutants emission The most important emissions considered in the power generation industry due to their effects on the environment are sulfur dioxide (SO2 ) and nitrogen oxides (NOx ). These emissions can be modeled through functions that associate emissions with power production for each unit. Sulfur dioxide emissions are dependent on fuel consumption and they take the same form as the fuel cost functions used for economic dispatch. NOx emissions are more difcult to model since they come from different sources and their production is associated with several factors such as boiler temperature and

1656

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

air content. One approach to represent SO2 and NOx emissions is to use a combination of polynomial and exponential terms [12]:
M

E(PG ) =
i=1

2 i + i PGi + i PGi + i exp(i PGi )

(2.3)

where i , i , i , i , and i are coefcients of the i-th generator emission characteristics. 2.2. Problem constraints There are two constraints considered in the problem, i.e., the generation capacity of each generator and the power balance of the entire power system. Constraint 1: Generation capacity constraint For normal system operations, real power output of each generator is restricted by lower and upper bounds as follows:
min max PGi PGi PGi

(2.4)

max min where PGi and PGi are the minimum and maximum power generated by generator i, respectively. Constraint 2: Power balance constraint The total power generation must cover the total demand PD and the real power loss in transmission lines PL . This relation can be expressed as M

by itself and its neighbors. At each generation, they observe the tness of themselves and their neighbors and move toward those with a better position. Let x and v denote a particle coordinates (position) and its corresponding ight speed (velocity) in the search space. Therefore, the i-th particle is represented as xi = [xi1 , xi2 , . . ., xid , . . ., xiM ] in the M-dimensional space. Each particle keeps track of its coordinates in the solution space which are associated with the best solution it has achieved so far. This tness value is called pbest. The best previous position of the i-th particle is recorded and represented as pbesti = [pbesti1 , pbesti2 , . . ., pbestid , . . ., pbestiM ]. Another best value that is tracked by the particle swarm optimizer is the best value obtained so far by any particle in the neighbors of the particle. When a particle takes all the population as its topological neighbors, the best value is a global best and is called gbest. The index of the best particle among all the particles in the group is represented by the gbestd . The rate of the velocity for particle i is represented as vi = (vi1 , vi2 , . . . , vid , . . . , viM ). The modied velocity and position of each particle can be calculated using the current velocity and the distance from pbestid to gbestd as shown in the following formulas: id
(t+1)

= w id + c1 rand() (pbestid xid ) + c2 Rand() (gbestd xid ),

(t)

(t)

(t)

(3.7)

xid

(t+1)

= xid + id

(t)

(t+1)

i = 1, 2, . . . , N, (3.8)

PGi = PD + PL
i=1

(2.5)

d = 1, 2, . . . , M.

Here a reduction is applied to model transmission losses as a function of the generators output through Krons loss coefcients. The Krons loss formula can be expressed as follows:
M M M

PL =
i=1 j=1

PGi Bij PGj +

i=1

B0i PGi + B00

(2.6)

where Bij , B0i , B00 are the transmission network power loss Bcoefcients, which are assumed to be constant, and reasonable accuracy can be achieved when the actual operating conditions are close to the base case where the B-coefcients were derived [25]. In summary, the objective of environmental/economic power dispatch optimization is to minimize F(PG ) and E(PG ) simultaneously subject to the constraints (2.4) and (2.5). 3. Mechanism of particle swarm optimization Inspired by the movement pattern in a bird ock or sh school, Kennedy and Eberhart proposed particle swarm optimization in 1995, which is a population-based stochastic optimization technique [15,16]. In PSO, individuals (particles) move around in a multidimensional search space to approach the optima, where each point represents a solution to the target problem. Initially a bunch of particles are randomly created and set into motion through this space. In their movement, each particle adjusts its position based on its own experience as well as the experience of a neighboring particle by utilizing the best position encountered

where N is the number of particles in a group, M is the number of members in a particle, t is the pointer of generations, [0,1] is the constriction factor which controls the velocity magnitude, w is the inertia weight factor, c1 and c2 are acceleration constants, rand() and Rand() are uniform random values in a range [0,1], (t) (t) t is the velocity of particle i at generation t, and xi is the current position of particle i at generation t. The particle swarm optimization concept consists of, at each time step, changing the velocity of (accelerating) each particle toward its pbest and gbest locations. Acceleration is weighted by a random term, with separate random numbers being generated for acceleration toward pbest and gbest locations. 4. Principles of multi-objective optimization Real-world applications usually involve simultaneous optimization of multiple objectives, which are generally noncommensurable and conicting with each other. The general multi-objective optimization problem is to nd the vector x = [x1 , x2 , . . . , xn ]T which satises the m inequality constraints: gi (x) 0, i = 1, 2, . . . , m (4.9)

the p equality constraints hi (x) = 0, i = 1, 2, . . . , p (4.10)

and optimizes the vector function f (x) = [f1 (x), f2 (x), . . . , fk (x)]T (4.11)

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

1657

In other words, the aim is to determine, from among the set of all numbers which satises (4.9) and (4.10), the particular set x1 , x2 , . . . , xn which yields the optimum values of all the objective functions. The constraints given by (4.9) and (4.10) dene the feasible region and any point x in denes a feasible solution. The vector function f (x) is a function which maps the set into the set which represents all possible values of the objective functions. The k components of the vector f (x) represent the noncommensurable criteria which must be considered. The constraints gi (x) and hi (x) represent the restrictions imposed on the decision variables. The vector x denotes the optimal solutions. It is often preferable to give a set of optimal solutions instead of a single one for such optimization problems, since no solution can be better than other ones with respect to all the objectives under consideration. These solutions are called Pareto-optimal solutions. More formally, a solution x is Pareto-optimal if for each x and I = {1, 2, . . ., k}, either i I (fi (x)) = fi (x ) or there exists at least i I such that fi (x) > fi (x ) (4.13) (4.12)

from its historical memory and the updated set of non-dominated solutions stored in the archive, respectively. In some evolutionary approaches such as genetic algorithms, elitism is often used to avoid losing the best-found solutions in order to expedite the evolutionary process [10]. In principle, the major distinction between multi-objective evolutionary optimization algorithms (MOEAs) and SOEAs lies in the mechanism of solution evaluation and elitism/archiving [17]. Archiving can be deemed as a subfunction of elitism since it applies the similar concept as in elitism to nd out and memorize the best individuals in each iteration. The archive-based MOPSO algorithm can be represented using the pseudo code as shown in Algorithm 1. Algorithm 1. An archive-based MOPSO algorithm.

Furthermore, Pareto-dominance is also an important concept in multi-objective optimization. An objective vector fa = (fa1 , . . . , fak ) dominates fb = (fb1 , . . . , fbk ) (often denoted as fa fb ) if only if fa is partially less than fb , i.e., i {1, . . ., k}, fai fbi i {1, . . ., k}, fai < fbi holds. Pareto-dominance is used in this study to compare the performance of each individual in the population, and thus no specic evaluation function needs to be explicitly dened anymore. 5. The proposed approach The standard PSO algorithm is not suited to resolve multiobjective optimization problems in that no absolute global optimum exists there, but rather a set of non-dominated solutions. Thus, to render the PSO algorithm capable of dealing with MO problems, some modications become necessary. When using stochastic search based algorithms to optimize multiobjective problems, two major issues normally arise during the algorithm design. First, the tness assignment and selection should be addressed so that the search can move towards the Pareto-optimal set. Second, the diversity of the population should be maintained to prevent premature convergence and obtain an evenly distributed Pareto-optimal front. In this paper, the standard PSO algorithm is modied and improved so as to facilitate a multi-objective optimization approach, i.e., multi-objective particle swarm optimization (MOPSO). The Pareto-dominance concept is used to extend the approach to tackle multi-objective problems. 5.1. Archiving mechanism According to (3.7) and (3.8), each particle needs to change its position based on its local and global guides, which are chosen An archive (a.k.a. repository), A, is employed to tackle the MO problems, which stores the non-dominated solutions found by the optimizer so far. Initially, A is empty and the positions and velocities of all the particles are initialized in a random manner. Both personal guide and global guide for each particle are initialized to be the starting position. At every iteration t, the velocities vi and positions xi of each particle are updated according to (3.7) and (3.8). The objectives are then evaluated provided that xi falls into the feasible region. Any solutions which are not weakly dominated by any element of the archive are added to A, and meanwhile any members of A which are dominated by xi are culled. In doing so, the archive A can be guaranteed to be a non-dominating set as the optimization process proceeds. The key point for achieving a competitive MOPSO algorithm lies in the appropriate selection of the personal and global guides. For the selection of pbesti , it is somehow straightforward. If the current position of the i-th particle xi weakly dominates pbesti or xi , and pbesti are mutually non-dominating, then pbesti is chosen as the current position. For the selection of global guide gbesti , it is much more complicated. Since elements in A are mutually non-dominating and no element in the archive is dominated by any xi , so the archive can be deemed to be globally superior

1658

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

with respect to each swarm member. Thus, each member in A can be a candidate for the potential global guide. There are several schemes which can be used to choose a global guide for each particle from the archive [6,7,28]. In this paper, a distinct selection scheme termed fuzzied MOPSO is proposed to deal with the MO problem in a more effective fashion by preserving the diversity of the solutions during the entire optimization run. The major function of the archive is to store a historical record of the non-dominated solutions found along the heuristic search process. The archive interacts with the generational population in each iteration so as to absorb superior current non-dominated solutions and eliminate inferior solutions currently stored in the archive. The non-dominated solutions obtained at every iteration in the generational population (swarm) are compared with the contents of archive in a one-per-one basis. A candidate solution can be added to the archive if it meets any of the following conditions: the archive is empty; the archive is not full and the candidate solution is not dominated by or equal to any solution currently stored in the archive; the candidate solution dominates any existing solution in the archive; the archive is full but the candidate solution is non-dominated and is in a less crowded region than at least one solution. In addition, the archive should be maintained such that all the solutions are non-dominated. If converged to the actual Pareto front, the non-dominated solutions stored in the archive are just a portion of the entire set of Pareto-optimal solutions. Therefore, it is highly necessary to provide the decision-maker a set of representative non-dominated solutions with good diversity. In this way, the decision-maker may be able to access the whole Pareto front from these selected solutions. Thus, one important criterion to measure the performance of an MO optimization algorithm is to check if the solutions derived from this algorithm spread along the entire Pareto front in an elegant fashion. 5.2. Fuzzied PSO The major difculty in extending PSO to handle multiobjective problems is to gure out a scheme for choosing both local and global guides for each particle in the swarm. Unlike single-objective problems, there are no explicit concepts on how personal and global bests can be clearly identied in MO problems. In the single-objective PSO, the global best particle can be readily found by choosing the particle with the best position. In the MO optimization problems, the optimum solutions are Pareto-optimal. Thus, each particle should select the globally best particle based on the Pareto-optimal concept. Very often, the key task in MOPSO is to determine the best global particle for each particle in the population. The proposed multi-objective particle swarm optimization algorithm is termed as Fuzzied MOPSO (i.e., FMOPSO) since a fuzzication mechanism is introduced for the selection of global best position gbest. Here we interpret gbest not just as a point but as an area, and each

point in the area has different possibilities of being chosen as the gbest. The fuzzication formula used in the study is N(gbest, std2 ), which represents a set of normally distributed particles with gbest as their mean value and std as standard deviation. Fig. 1 depicts the global guide selection process in the proposed fuzzied PSO algorithm. First, in each iteration, the original gbest is selected from the archive, which is however not used directly to update the particle speed and position. Instead, an area around it is randomly generated based on the normal distribution. Then, tournament selection is applied to choose the gbest from this area, which will be used to update the particle speed and position. It is obvious that large std values will result in large generated selection regions. Furthermore, in tournament selection, local competition is used to determine survivors. In this study, binary tournament selection is used where the individual with the higher tness in the group of two individuals is selected, and the other is removed. This selection scheme can be deemed as an effective measure to increase the population diversity during the optimization process. During the heuristic multi-objective optimization process, since the MO optimization algorithm is attempting to build up a discrete picture of a possibly continuous Pareto front, it is often desired to distribute the solutions as diversely as possible on the discovered tradeoff curve. Furthermore, the uniformity among the distributed solutions is also crucial so as to achieve consistent and smooth transition among the solution points when searching for the best compromise solution based on the particular requirements of the target problem. Therefore, to accomplish these challenges, it is highly necessary to preserve the diversity of solutions distribution during the optimization process. This global guide selection scheme in MOPSO can be regarded as an effective measure for preserving distribution diversity and uniformity. 5.3. Niching and tness sharing The selection mechanism in PSO may drive the population towards a uniform distribution of mutants of the highly t individual. This phenomenon is called genetic drift and it is damaging to the formation and maintenance of good spread of solutions. It may also trap the PSO in local optima. To avoid genetic shift caused by the selection pressure, the niching (or speciation) mechanism is used in the proposed algorithm. Fitness sharing is one of the widely used niching methods, where the tness represents the resource for which the individuals belonging to the same niche compete. Sharing was originally proposed to promote the population distribution to prevent genetic drift as well as to search for possible multiple peaks in single-objective optimization. Recently, it has received great attention as one of the crucial measures to preserve distribution in coping with MO optimization problems [10]. From the inner working of tness sharing we adopted, it is evident that the impact of genetic shift can be reduced or even eliminated since the shared tness of the replicas is reduced when their number increases. In doing so, other individuals with lower raw tness or belonging to less populated niches will have higher chances to compete for selection and reproduction. To sum up, the adop-

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

1659

Fig. 1. Global guide selection in the fuzzied PSO.

tion of tness sharing can maintain or increase the diversity of the individuals during the whole optimization process. 5.4. Turbulence factor

its control. It helps to prevent the particle from becoming stuck in local minima as well as to enable wide exploration in the decision space. 5.5. Constraints handling

In this study, a turbulence operator is also used in order to improve the performance of PSO by enhancing the population diversity. The turbulence operator is essentially a mutation operator, which acts on the current velocity value of each particle obtained from (3.8) in the standard PSO procedure. xid
(t+1)

= xid

(t+1)

+ RT xid

(t+1)

i = 1, 2, . . . , N, (5.14)

d = 1, 2, . . . , M.

The nal term added in (5.14) is a small stochastic perturbation, known as the turbulence factor, where Rt [1,1]. The turbulence factor is a random value added to the updated position of each particle with a certain probability. It is equivalent to the perturbation operation in Evolutionary Strategy (ES), which reects the variation during a particles ight which is beyond

Since PSO is essentially an unconstrained optimization algorithm, the constraints handling scheme needs to be incorporated into it in order to deal with the constrained power dispatch problem. Here a straightforward constraint checking procedure called rejecting strategy is added. When any two individuals are compared, their constraints are examined rst. If both satisfy the constraints, the concept of Pareto-dominance is then applied to determine which one is the survivor. If one is feasible and the other is not, the feasible candidate dominates. Though this constraint satisfaction checking scheme is simple, it turns out to be very effective in guaranteeing the feasibility of the non-dominated solutions. A numerical example can be used to illustrate this process. In iteration 275 of the test problem, among the 100 individuals generated based on the MOPSO mechanism,

1660

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

Table 1 Three representative candidate solutions in iteration 275 Generators/objectives PG1 PG2 PG3 PG4 PG5 PG6 PG7 PG8 PG9 PG10 PG11 PG12 PG13 PG14 Fuel cost ($/h) Emission (ton/h) Indi. 1 (MW) 108.008 103.795 45.0323 52.1524 68.2602 84.8787 50 53.5470 81.3461 50.2011 79.2568 123.078 50.1936 50.2905 12962.8 42.0344 Indi. 2 (MW) 91.0041 117.961 51.3432 52.5802 72.2774 90.3492 50.2911 59.7973 82.7082 52.2715 75.9077 103.868 51.8011 50.3800 13124.1 38.8393 Indi. 3 (MW) 85.0549 102.861 52.1282 53.2562 74.2796 86.3631 52.7150 59.1250 82.3278 51.9967 85.1762 108.583 56.1753 50 13120.8 37.9634

5.7. Computational ow In principle, an archive-based MOPSO algorithm can be illustrated in Fig. 3. As seen from the gure, initially the population is randomly created, and then the selection pressure from the PSO algorithm drives the population move towards the better positions. At each iteration, the generational population is updated and certain elite individuals from it are chosen to store in the elite population (archive) based on the Pareto-dominance concept. It should be noted that each individual also maintains a memory which records the best positions the individual has encountered so far. The personal best for each particle is selected from this memory. Meanwhile, among the individuals stored in the archive, the global best is singled out according to our fuzzied global guide selection strategy. Both guides are then used by the PSO to steer the search to promising regions. The procedure is repeated until the maximum number of iterations is reached or the solutions cease improving for a certain number of generations. Under this framework, other multi-objective optimization algorithms can also be developed based on different meta-heuristics such as evolutionary computation, simulated annealing, and tabu search. The proposed FMOPSO is applied to the constrained power dispatch problem for searching for the optimal or near-optimal solutions. Its computational ow is described in the following procedure: Step 1: Specify the lower and upper bound generation power of each unit. Step 2: Randomly initialize the individuals of the population. Note that the speed and position of each particle should be initialized such that each candidate solution (particle) locates within the feasible search space. Step 3: For each individual PGi of the population, the transmission loss PLi is calculated based on B-coemcient loss formula. Step 4: Evaluate each individual PGi in the population based on the concept of Pareto-dominance. Step 5: Store the non-dominated members found thus far in the archive. Step 6: Initialize the memory of each particle in which a single local best pbest is stored. The memory is contained in another archive. Step 7: Increment iteration counter. Step 8: Choose the personal best position pbest for each particle based on the memory record; choose the global best gbest from the fuzzied region using binary tournament selection. The niching and tness sharing mechanism is also applied throughout this process for choosing both local and global guides.

we select three representative individuals as shown in Table 1 to demonstrate how the rejecting strategy is applied during the optimization process. All of these three individuals satisfy the real power balance constraint. However, individual 1 violates the capacity constraint since the output of generator 3 is smaller than the lower limit of 50 MW. Thus, this individual will be discarded and no further optimization operations will be applied on it in the next iteration. Both the remaining candidate solutions satisfy the capacity constraints. Then comparisons are made between these two potential solutions and the solutions currently stored in the archive in terms of Pareto-optimality. In this numerical example, individual 2 is dominated and individual 3 is not dominated by the solutions currently stored in the archive. Thus, only individual 3 is chosen to be stored in the archive. 5.6. Encoding scheme It is crucial to appropriately encode the individuals of the population in PSO for handling the EED application. The power output of each generating unit is selected as the gene and several genes constitute the individual, which is a candidate solution for the target problem. The genes here are all real-coded. The i-th individual PGi can be represented as follows: PGi = [PGi1 , PGi2 , . . . , PGid , . . . , PGiM ], i = 1, 2, . . . , N (5.15) where M is the number of generators and N is the population size. PGid is the power generated by the d-th unit in i-th individual. Thus, the dimension of a population is N M. The structure of individual (particle) i can be illustrated in Fig. 2.

Fig. 2. Encoding scheme for each particle.

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

1661

Fig. 3. Data ow diagram of the proposed algorithm.

Step 9: Update the member velocity v of each individual PGi based on (3.7) as follows: id
(t+1)

6.1. Performance of FMOPSO Since PSO algorithms are sometimes quite sensitive to certain parameters, the simulation parameters should be carefully chosen. In principle, the larger the generational population size is, the more diverse the obtained solutions will be. However, a large number of individuals inevitably incurs high computational cost. Archive size is an important factor, which may determine the quality of the Pareto front produced. It species the maximum number of non-dominated solutions that can be stored in the archive. In the simulations, after some trials, both the population size and archive size are set to 100, and the number of generations is set to 500. The constants c1 and c2 represent the weighting of the stochastic acceleration terms that pull each particle toward the pbest and gbest positions, respectively. Low values may allow particles to y away from the target areas before being pulled back, while high values incur abrupt movement toward, or even past the target areas. After careful tuning, both acceleration constants in the algorithm are chosen as 1. The selection of inertia weight factor w is crucial to the convergence behavior of the PSO algorithm. A large w facilitates search of unexplored areas since it facilitates global exploration by giving large weight to the current velocity. On the other hand, a small one motivates the exploitation of the current area under search. A properly chosen inertia factor can balance the global and local exploration of the swarm, which may lead to better solutions. Thus, in this study, the inertia weight factor w will decrease when the number of generations increases. It decreases linearly during the optimization run according to w = wmax wmax wmin iter itermax (6.19)

= w i + c1 rand() (pbestid PGid ) + c2 Rand() (gbestd PGid ),

(t)

(t)

(t)

i = 1, . . . , N;

d = 1, . . . , M

(5.16)

where N is the population size, M is the number of generating units, and w is the inertia weight factor. Step 10: Modify the member position of each individual PGi based on (3.8) as follows: PGid = PGid + id
(t+1) (t+1) (t) (t+1)

(5.17)

If PGid violates the constraints, then it should be modied accordingly in order to stay within the search space. Following this, add the turbulence factor into the current position as follows: PGid = PGid + RT PGid
(t+1) (t+1) (t+1)

(5.18)

Step 11: Update the archive which stores non-dominated solutions according to the four selection criteria discussed previously. Step 12: If the current individual is dominated by the pbest in the memory, then keep the pbest in the memory; otherwise, replace the pbest in the memory with the current individual. Step 13: If the maximum iterations are reached, then go to Step 14. Otherwise, go to Step 7. Step 14: Output a set of Pareto-optimal solutions from the archive as the nal solutions. 6. Simulation and evaluation of the proposed approach In this study, a typical IEEE 118-bus test system with 14generators [12] is used to investigate the effectiveness of the proposed FMOPSO approach. This is the largest practical test system we could nd in the existing literature with the complete data required for our environmental/economic dispatch problem. However, there is no inherent limitation of the proposed method for larger systems. The load demand used in the simulations is 950 MW.

where itermax is the number of generations and iter is the current number of iterations. Both turbulence factor and niche radius are set to 0.02. In the rst 200 generations, the gbest is fuzzied using a large standard deviation to generate a region around the gbest according to Gaussian distribution. In the remaining 300 generations, its value is decreased. This makes sense since, similar to the choice of w values, initial large standard deviation enables global search while the following small standard deviation facilitates local exploration using small movement in

1662

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

objectives as follows: Minimize WF (PG ) + (1 W)E(PG ) (6.20)

where W is a weight factor indicating the relative importance of its associated objective during the optimization and is the scaling factor. The approach is often called weighted aggregation approach (a.k.a. naive approach). However, in most practical applications, there is no sufcient information on the current operating conditions. Thus, it is hard for decision-makers to articulate their preferences for the multiple objectives represented by their associated weighting factors. Also it was found that this approach does not work in problems which have a nonconvex Pareto-optimal front. By varying the value of weights, the Pareto-optimal front can be approximated by the weighted aggregation approach. 6.2.2. Multi-objective optimization evolutionary algorithms As mentioned earlier, a variety of multi-objective evolutionary algorithms have been applied to handle the EED problem. MOEAs are more established as compared to MOPSO, and they are most often used to handle the MO problems which cannot be well resolved by the exact methods such as linear programming, nonlinear programming, and other gradient-based search algorithms [26,29]. The evolutionary optimization methods normally adopt effective parallel search and the Pareto-optimal set can be obtained in a single optimization run. Unlike the constructive cooperation principle used in PSO, most evolutionary algorithms are based on the philosophy of survival of the ttest. Here, a representative multi-objective evolutionary algorithm based on genetic algorithm is implemented to solve the EED problem [28]. The Pareto fronts obtained using WA and MOEA are shown in Fig. 5. By comparison, we can appreciate that the solutions obtained by these two approaches are dominated by those derived through FMOPSO. Also the diversity of the solutions obtained using FMOPSO is the highest among all the three approaches due to the diversity preservation mechanisms used throughout the optimization process, including fuzzication, niching and tness sharing, and turbulence factor. Furthermore,

Fig. 4. Pareto front obtained using FMOPSO.

each iteration. The simulation program is written in C++ and run in 2.20 GHz Pentium-4 processor with 512 MB of RAM. The Pareto-optimal front obtained using FMOPSO is shown in Fig. 4. It is evident that the Pareto-optimal solutions are well-distributed over the tradeoff curve primarily due to several diversity-preserving mechanisms used throughout the entire optimization process in the proposed algorithm. 6.2. Comparative studies To demonstrate the effectiveness of the proposed algorithm, two other approaches are used to study the EED problem for comparison purpose, i.e., the weighted aggregation and the multi-objective evolutionary optimization algorithm. 6.2.1. Weighted aggregation Here, the PSO algorithm is used to nd the Pareto-optimal solutions where the problem is treated as a single-objective optimization one by linear combination of cost and emission

Fig. 5. Comparison of different Pareto fronts obtained.

Fig. 6. Linear membership function.

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664

1663

our proposed PSO-based approach takes 18.8 s to derive the set of non-dominated solutions while the GA-based approach spends 76.9 s. 6.3. Best tradeoff solution After obtaining the Pareto-optimal solution, the decisionmaker may need to choose one best compromise solution according to the specic preference for different applications [21]. However, due to the inaccurate nature of human judgment, it is very often not possible to explicitly dene what is really needed. Thus, fuzzy set is introduced here to handle the dilemma. Here a linear membership function is dened for each of the objective functions Fi : max Fi Fi , F min < F < F max max i i i F Fimin i ui = (6.21) 1, Fi Fimin 0, Fi Fimax In the above denition, Fimin is the value of an objective function i which is completely satisfactory to the decision-maker, while Fimax is the value that is clearly unsatisfactory. It is evident that this membership function indicates the degree of achievement of the objective functions. The function shape is depicted in Fig. 6. For every non-dominated solution k, the membership function can be normalized as follows: uk =
O k i=1 ui S O k k=1 i=1 ui

max min min F1 = 12900 $/h, F1 = 13400 $/h, F2 = 35 ton/h, and max = 51 ton/h. F2 From all the previous comparative studies, it is evident that the average cost and emission are both reduced by using FMOPSO. These values are not neglectable due to the continuous operations of power dispatch throughout the year as well as the numerous power plants worldwide. Thus, the proposed power dispatch algorithm has a very positive signicance in reducing both fuel costs and adverse environment impacts.

7. Concluding remarks In this paper, we investigated the environmental/economic power dispatch problem by employing a fuzzied multiobjective particle swarm optimization algorithm. The power dispatch is formulated into a bi-objective optimization problem, which is to minimize the fuel cost as well as pollutant emission simultaneously. In this work, the basic particle swarm algorithm is revised and extended to tackle the multi-objective case and some modications are made in order to render it more effective in dealing with multi-objective optimization problems. The proposed algorithm turns out to be effective since the obtained solutions have some superior characteristics as compared with other approaches including WA and MOEA. Moreover, the proposed approach tackles the multi-objective optimization problems in a more efcient fashion since it is able to derive a manageable set of non-dominated solutions in a single run. There are many issues which may be further investigated. For instance, nonlinear characteristics of the generator such as ramp rate limits and prohibited operating zone for actual power system operations can be considered for more practical power dispatch. Furthermore, power system reliability can be introduced, which makes power dispatch a stochastic problem. References
[1] M.A. Abido, A novel multiobjective evolutionary algorithm for environmental/economic power dispatch, Electric Power Syst. Res. 65 (2003) 7181. [2] M.A. Abido, A niched Pareto genetic algorithm for multiobjective environmental/economic dispatch, Electrical Power Energy Syst. 25 (2) (2003) 97105. [3] M.A. Abido, Environmental/economic power dispatch using multiobjective evolutionary algorithms, IEEE Trans. Power Syst. 18 (4) (2003) 15291537. [4] R.T.F. Ah King, H.C.S. Rughooputh, K. Deb, Evolutionary MultiObjective Environmental/Economic Dispatch: Stochastic vs. Deterministic Approaches, KanGAL Report Number 2004019, 2004. [5] M.R. Alrashidi, M.E. El-Hawary, A survey of particle swarm optimization applications in power system operations, Electric Power Components Syst. 34 (12) (2006) 13491357. [6] C.A. Coello Coello, G.T. Pulido, M.S. Lechuga, Handling multiple objectives with particle swarm optimization, IEEE Trans. Evol. Comput. 8 (3) (2004) 256279. [7] A. Delia Cioppa, C. De Stefano, A. Marcelli, On the role of population size and niche radius in tness sharing, IEEE Trans. Evol. Comput. 8 (6) (2004) 580592. [8] J.S. Dhillon, S.C. Parti, D.P. Kothari, Stochastic economic emission load dispatch, Electric Power Syst. Res. 26 (1993) 179186.

(6.22)

where O and S are the number of objective functions and the number of non-dominated solutions, respectively. The solution with the maximum membership uk can be seen as the best compromise solution. For instance, from the non-dominated solutions for our EED problem, the best compromise solutions for different algorithms are shown in Table 2 if we choose
Table 2 Comparison of the best compromise solutions obtained from different algorithms Generators/objectives PG1 PG2 PG3 PG4 PG5 PG6 PG7 PG8 PG9 PG10 PG11 PG12 PG13 PG14 Fuel cost ($/h) Emission (ton/h) FMOPSO (MW) 94.5703 105.728 50.929 50 75.7894 84.6362 53.3723 54.8911 83.6218 52.5273 79.515 106.104 58.1926 50.1546 13094.9 38.0501 WA (MW) 91.1562 109.584 51.4286 50.1945 68.3609 90.6869 53.5931 56.4637 77.0796 51.234 87.3122 110.159 55.1502 50.722 13113.7 39.2491 MOEA (MW) 81.6684 108.597 50.3574 50.0378 88.2061 89.5116 50 51.6133 82.3149 54.5174 84.3849 112.184 51.427 50.408 13128.8 39.7978

1664

L. Wang, C. Singh / Electric Power Systems Research 77 (2007) 16541664 [21] T. Niimura, T. Nakashima, Multiobjective tradeoff analysis of deregulated electricity transactions, Electrical Power Energy Syst. 25 (2003) 179185. [22] J.H. Park, Y.S. Kim, I.K. Eom, K.Y. Lee, Economic load dispatch for piecewise quadratic cost function using Hopeld neural network, IEEE Trans. Power Syst. 8 (3) (1993) 10301038. [23] R.M. Rifaat, Economic dispatch of combined cycle cogeneration plants with environmental constraints, Proceedings of International Conference on Energy Management and Power Delivery, vol. 1, Singapore, 1998, pp. 149153. [24] C.A. Roa-Sepulveda, M. Herrera, A solution to the economic dispatch problem using decision trees, Electric Power Syst. Res. 73 (2000) 255259. [25] H. Saadat, Power System Analysis, The McGraw-Hill Companies, New York, 1999. [26] J.R. Schott, Fault Tolerant Design Using Single and Multicriteria Genetic Algorithm Optimization, M.S. thesis, Dept. Aeronautics and Astronautics, Massachusetts Inst. Technol., Cambridge, MA, May 1995. [27] Y.H. Song, G.S. Wang, P.V. Wang, A.T. Johns, Environmental/economic dispatch using fuzzy logic controlled genetic algorithms, IEE Proceedings of Generation, Transmission, Distribution, vol. 144, no. 4, 1997, pp. 377382. [28] K.C. Tan, T.H. Lee, E.F. Khor, Evolutionary algorithm with dynamic population size and local exploration for multiobjective optimization, IEEE Trans. Evol. Comput. 5 (6) (2001) 565588. [29] D.A. Van Veldhuizen, Multiobjective Evolutionary Algorithms: Classications, Analyzes, and New Innovation, Ph.D. dissertation, Dept. Elec. Comput. Eng., Air Force Inst. Technol., Wright-Patterson AFB, OH, May 1999. [30] P. Venkatesh, R. Gnanadass, P. Narayana Prasad, Comparison and application of evolutionary programming techniques to combined economic emission dispatch with line ow constraints, IEEE Trans. Power Syst. 18 (2) (2003) 688697.

[9] A. Farag, S. Al-Baiyat, T.C. Cheng, Economic load dispatch multiobjective optimization procedures using linear programming techniques, IEEE Trans. Power Syst. 10 (1995) 731738. [10] J.E. Fieldsend, R.M. Everson, S. Singh, Using unconstrained elite archives for multiobjective optimization, IEEE Trans. Evol. Comput. 7 (3) (2003) 305323. [11] Z.-L. Gaing, Particle swarm optimization to solving the economic dispatch considering the generator constraints, IEEE Trans. Power Syst. 18 (3) (2003) 11871195. [12] R.P. Guerrero, Differential Evolution Based Power Dispatch Algorithms, M.S. thesis, Electrical Engineering Department, University of Puerto Rico, Mayaguez Campus, 2004. [13] P.K. Hota, S.K. Dash, Multiobjective generation dispatch through a neurofuzzy technique, Electric Power Components Syst. 32 (2004) 11911206. [14] T. Jayabarathi, K. Jayabarathi, D.N. Jeyakumar, et al., Evolutionary programming techniques for different kinds of economic dispatch problems, Electric Power Syst. Res. 73 (2005) 169176. [15] J. Kennedy, R. Eberhart, Particle swarm optimization, IEEE Proceedings of the International Conference on Neural Networks, Perth, Australia, 1995, pp. 19421948. [16] J. Kennedy, R. Eberhart, Swarm Intelligence, Morgan Kauf-mann Publishers, San Francisco, 2001. [17] E.F. Khor, K.C. Tan, T.H. Lee, C.K. Goh, A study on distribution preservation mechanism in evolutionary multi-objective optimization, Artif. Intell. Rev. 23 (2005) 3156. [18] K.Y. Lee, J.-B. Park, Application of particle swarm optimization to economic dispatch problem: advantages and disadvantages, IEEE Proceedings of Power Systems Conference and Exposition (PSCE), Atlanta, GA, 2006, pp. 188192. [19] K.Y. Lee, A.S. Yome, J.H. Park, Adaptive Hopeld neural networks for economic load dispatch, IEEE Trans. Power Syst. 13 (2) (1998) 519526. [20] C.E. Lin, G.L. Viviani, Hierarchical economic dispatch for piecewise quadratic cost functions, IEEE Trans. Power Apparatus Syst. 103 (6) (1984) 11701175.