You are on page 1of 5

2009 Fifth International Conference on Natural Computation

Particle Swarm Optimization Based on Cultural Algorithm for Short-term Optimal Operation of Cascade Hydropower Stations
Wei Xie, Chang-ming Ji, Xin-wu Li
Research Institute of water Resources and Hydro-Electric Engineering, Renewable Energy School North China Electric Power University Beijing, China baobaoxiewei@163.com
AbstractDifficulties likely exist in optimal reservoir operation along with discreteness and nonlinear constraints. Therefore, Particle Swarm Optimization based on Cultural Algorithm (PSO-CA) is presented in this paper for overcoming these defects. In this article the evolutionary mechanism of Particle Swarm Optimization Algorithm (PSO) is guided by Cultural Algorithm (CA). PSO-CA uses PSO in population space and guides the evolution by shape knowledge and standardization knowledge in belief space. An example is also used to show that the PSO-CA algorithm has a better applied prospect for its high reliability and fast operation speed in global optimization. Keywords-optimal operation of reservoir; particle swarm optimization; cultural algorithm;

which may realize searching optimal point on continuous space by multipoint optimization. Furthermore, aiming to avoid the precocity in the application of cascade hydropower stations and how to settle them with shape and standardization knowledge in belief space is discussed in this paper. II. ANALYSIS ON PSO-CA ALGORITHM

I.

INTRODUCTION

The short-term optimal operation of cascade hydropower stations is typical nonlinear and multi-dimensional optimization problem with unequal constraints in nature. Because of its dominating characteristics such as multivariable, highly-complexity and multi-constraint [1], it is more complicated and difficult to be solved rather than other nonlinear optimal problems. Systematic methods and algorithms have been quickly developed exemplified by dynamic programming (DP), large system decomposition coordination model (LSDC), network flow method [2], genetic algorithm (GA) [3] etc., for their easy realization and simple programme. However, they werent versatile to settle reservoir regulation, and some defects were exposed in their application, such as dimension disaster and overlong calculating time of the DP, high complexity and low convergence of LSDC caused by coordination factor, are easily plunged into local optimum of GA caused by low limits constructing of object function and constraints. Particle swarm optimization algorithm [4] (PSO) has been widely applied to the nonlinear and nondifferentiable problems of power system [5, 6] as a stochastic global optimization technology for its low dependence of object function and constraint conditions, easy use and rapid convergence. However, some questions still exist. For example, PSO algorithm is easy to fall into local optimum and lead to precocity phenomenon as its not a global convergence algorithm [7]. As previous research has not great perfection on reservoirs operation, Particle swarm optimization based on cultural algorithm (PSO-CA) is adopted to overcome the non-high precision in the paper,
978-0-7695-3736-8/09 $25.00 2009 IEEE DOI 10.1109/ICNC.2009.182 289

A. Principle of PSO The PSO algorithm is a new optimization algorithm, which derived from researching of bird and fish flocks preying. The algorithm modifies the individual behavior strategies by the information of flock and their own experiences given by evaluating the fitness of solution, and finally finds global optimal solution [8]. The main idea of PSO algorithm can be described as follows: firstly, a swarm of particles is randomly initialized with a position and velocity. The position which decides the particle is flying direction can be expressed as an ndimensional vector Xi= (xi,1, xi,2 xi,n), and the velocity can be also expressed as Vi= (vi,1, vi,2...vi,n). Each particle has a fitness value (Yi) which can be attained when put vector into objective function. And each particle do self-renewal follows two peaks. One is the optimal solution of a particle itself in flying named individual optimal solution Pik = (pki,1, pki,2pki,n). The other is the optimal solution among all the particle flying named global optimal solution Pgk = (pkg,1, pkg,2pkg,n)., where k is iteration number and k [1, K], and K is the biggest generation number in k. During each iteration, particle will regulate its velocity and then renew its position. The movement equations are as follows:

Vi k +1 = w Vi k + c1r1[ Pi k X ik ] + c2 r2 [ Pgk X ik ]

(1)

X ik +1 = X ik + Vi k +1

(2)

Where w is inertia weight coefficient and can alter the particles searching capacity by the way of adjusting its value. The value of w is linear degression in the interval [0.1, 0.9] along with the increase of iteration number in this paper; r1 and r2 are stochastic numbers between [0, 1]; c1 and c2 are two constant values, normally c1=c2=2.

B. Principle of CA In human society, culture can be viewed as a vehicle for the storage of information that is potentially accessible to all the members of the society, and used in guiding their problem solving activities. CA, developed by Robert G.Reynlods in 1994 [9], is used to model the evolution of the cultural component of an evolutionary computational system over time as it accumulates experience in solving a given set of problems. Culture guides the populations evolution with a certain speed which transcends that only by genes. The knowledge, collected from identifying patterns in the populations problem solving experiences, is used to influence the generation of candidate solutions, either to promote more instances of desirable candidates or to reduce the number of less desirable candidates in the population. The framework of CA is depicted in Figure 1. Different from other evolutionary algorithms, CA contains two evolutionary spaces. They are population space that evolutionary computation methods have and belief space composed by individuals knowledge and experience. Evolution can be viewed as an optimization and as a learning process through a supporting communication mechanism that between the two components. Individual Experiences selected from the population space are used to generate problem solving knowledge that resides in the belief space. The belief space stores and manipulates the knowledge and in turn influences the evolution of the population component. In this way, the population space and the belief space interact with each other and support each other.
Belief Space Update ()

1) Definition and Update of Belief Space: Shape knowledge has n+1 elements (n is the size of species), the last element is bg which consists of the history optimal position (X), the corresponding fitness value (Y) and continuous not update iteration number (nStaCount) of species. The other n elements described as b1~bn, and bi is composed of the history optimal position (Xi), the corresponding fitness value (Yi) and continuous not update iteration number (nStaCount) of the ith particle. Data structure of shape knowledge shown as Figure 2:
b1 b2 b3

bn

bg

x1

x2

xm

nStaCount

Figure 2. Data structure of shape knowledge

Acceptance ()

Influence ()

Population Space Evolution () Objective ()

Figure 1. Cultural Algorithm framework

When initializing shape knowledge, bi consists of initial position, the corresponding fitness value of the ith particle and nStaCount is equal to 0; bg also should use initial position, the corresponding fitness value of the particle which has the largest fitness value in the initial particle swarm and nStaCount also equals 0. Shape knowledge is updated by Acceptance () function during each iteration. When the current fitness value of a particle is higher than that which stores in shape knowledge, replaces the latter with the former. At the same time, the corresponding nStaCount is equal to 0 otherwise add 1. So does bg. Standardization knowledge represented as N = <N1, N2 Nm> means the range of decision variable (Xi) and m is the number of that Ni expresses as <Ii, Li, Ui>. And I = [li, ui] = {Xi| li Xi ui} which indicates the range of Xi. Li is the fitness value which can be obtained when li is put into the objective function and its initial value is + . The meaning of Ui is as the same as Li. Then standardization knowledge is renewed according to the probability conditions in (3) under the assumption that the lower limit of the jth particle is affected by the ith particle and the upper limit is affected by the dth particle.

C. The PSO-CA Algorithm CA provides an efficient algorithm framework and each kind of species evolutionary algorithm can provide the species for population space, such as GA, PSO, EC, etc.. The definition of belief space usually adopts shape knowledge (SK) and standardization knowledge (NK), namely <S, N> structure. In this paper, species are generated by PSO in population space and the evolution of species is leaded by <S, N> structure in belief space. Although PSO has virtues advanced above, it has shortcomings on lower search precision and tending to partial optimization. SO shape knowledge is used to improve the capabilities of research and standardization knowledge to enhance the computational efficiencies of the algorithm.

l k +1 = j Lkj+1 = u
k +1 j

xik, j lk j obj ( xi ) Lkj


k xd , j

if xik, j l k or F ( xik ) < Lkj j otherwise if xik, j l k or F ( xik ) < Lkj j otherwise
k k if xd , j u k or F ( x d ) < U k j j

(3)

u tj obj ( xd ) Uk j

otherwise
k k if x d , j u k or F ( xd ) < U k j j

U k +1 = j

otherwise

290

2) Influence on Population Space by the Belief Space: To change the default, adjust the template as follows. Population space obtains the information from belief space through Influence () function and these influences include the following two aspects: Firstly, influence of shape knowledge. It is clear that the particle will be the changed one when nStaCount is greater than staMax, where staMax is a given value. When the quantity of the variable particles reaches , in that case, a randomized crossover and mutation will be carried out to those ones, where is a supposed ratio. And the corresponding information in the belief space also should be renewed at one time. If the given individuals of a and b should be changed, the laws are given as:
k k X a +1 = s X a + (1 s ) X bk

E = max

Pi ,t t

t =1 i =1

(6)

Pi ,t = K i Qi ,t H i ,t
Where E is the total generated energy; Pi,t is the average output; t is the length of time interval, and T is the time interval number; N is the number of cascade hydropower stations; Ki means the output coefficient of power station; Qi,t expresses the power flow; Hi,t shows the water head of time interval t which is the water level difference between upstream and downstream. B. Main Constraints a) Water Balance Constraint

X bk +1 = s X bk + (1 s ) X ak

(4)

Vi,t +1 = Vi,t + (I i,t Qi,t Si,t )t I i,t = Qi1,t i + qi,t

(i = 1,2,...N; t = 1,2,...T )

(7)

Where s is the parameter that obtained by the experience of experiment. Secondly, influence of standardization knowledge. Different particle has different position and velocity and the influence for the next flying is also not the same. The next velocity is determined by the particles current position, so (1) can be replaced by the (5).
k w k vik, j + c1 r1 pik, j xik, j + c 2 r2 p g , j xik, j

[
k j

] ] ]
(5)

Where Vi,t and Vi,t+1 are respectively the reservoir storage in the time t and t+1; Ii,t expresses the reservoir inflow in the time interval t while Qi,t is the reservoir outflow in the same time interval; Si,t shows the surplus flow in time interval t while qi,t is the region inflow in the same period; Qi-1,t- i means the outflow of the i-1th power station and i is the water flow time between the i-1th and ith power station. b) Water Level Constraint

if x vik, +1 = j

k i, j

< l and x

k w k vik, j c1 r1 pik, j xik, j c 2 r2 p g , j xik, j

k i, j

<p

k g, j

Z i , min Z i ,t Z i , max (i = 1,2,...N ; t = 1,2,...T )

(8)

if x

k i, j

> u and x

k w k vik, j c1 r1 pik, j xik, j c 2 r2 p g , j xik, j

k j

k i, j

>p

k g, j

otherwise
Where means may using + or - simultaneity III. MATHEMATICAL MODEL OF OPTIMAL OPERATION OF HYDROPOWER STATIONS

Where Zi,t expresses the water level in the time interval t; Zi,min and Zi,max respectively means the minimum and maximum limited water level during the whole scheduling period. c) Output Constraint

Pi , min Pi ,t Pi ,max (i = 1,2,... N ; t = 1,2,...T )

(9)

Generally speaking, the optimal operation of cascade hydropower stations has two criterions: the maximum energy output under the condition of guarantee irrigating, flood control security, water supply and ice prevention; the minimum water consumption of reservoirs under the condition of meeting the load demand. This paper researches on the former. A. Objective Function The initial water level and total water consumption during the period of scheduling have been known. The objective function can be expressed as the following equation

Where Pi,t is the power station output in time interval t, while Pi,min is the minimum allowable output in the same period, which is often the guaranteed output; Pi,max means the maximum allowable output in the time interval t, which is often the installed capacity. d) Outflow Constraint

Qi , min Qi ,t Qi , max (i = 1,2,... N ; t = 1,2,...T )

(10)

Where Qi,t indicates the reservoir outflow in the time interval t; Qi,min expresses the minimum outflow according to the demand of water resource department in time interval t, and Qi,max shows the permitted maximum flow to guarantee the flood control security of downstream region.

291

IV. PSO-CA APPLICATIONS ON SHORT-TERM OPTIMAL OPERATION OF CASCADE HYDROPOWER STATIONS Reservoir optimal operation is a continuous and dynamic problem with multi-time intervals and unequal constraint in nature. Reservoir operating rule is signified as power discharge sequence in this paper. But in the programming process the power discharge sequence may usually transformed into water level change sequence in each storage time interval. The paper unified PSO-CA theory and reservoir operation together, reservoir water levels corresponding to particles, amount of scheduling period to dimension of searching space, generation object function to fitness function. In that way, one particle can be indicated as mn matrix (Xmn) where m is the number of hydropower stations and n is the quantity of time interval. Thus the component xit(i=1,2,,m; t=1,2,,n)of the matrix means the water level in tth time interval of the ith station. The only problem to be solved is to dispose restriction condition by punishment function, which shows as the follow [10]:

of the cascade hydropower stations is annual regulating mode while the other (shows as B) is daily. Here, the dead water level of A is 172m and B is 69m, the maximum restriction water levels are not higher than 210m and 90m respectively. The output of A is in the range of 55~1220MW and B is 20~300MW. The outflow of A is in the range of 70~1200m3/s and B is 20~860 m3/s. Note that the scheduling period is 24h and the time interval is 1h. A is 60km away from B thus we set the water flow time i is 2h. Here, the process of inflow and the initial water level of the two hydropower stations respectively 185m and 80m have been known.
Begin Initialization particle swarm Initialization belief space Initialization belief space

Calculation the fitness value of each particle Updating particle swarm Belief space

F ( x ) = E Rk

k k =1

[hk ( x)]2 R j

j j =1

max[0, g j ( x)]2

(11)
Producing a new generation of groups

Updating particle swarm

Where F(x) is fitness value function and E is the objective function of reservoir optimal operation; hk(x) expresses equation restriction and gi(x) indicates inequation constraint; Rj and Rk are weight coefficient. Reservoir optimal operation calculation steps based on PSO-CA algorithm are represented as the followings: Step 1: Initialization particle swarm. To shorten searching time of the algorithm, the initial particles and their corresponding velocities are randomly generated in constraint. Step 2: Ascertaining the parameters: constant c1 and c2, inertia weight coefficient w and the largest number of evolutionary iterations K. Step 3: Initialization the corresponding information of shape knowledge and standardization knowledge according to the initial swarm in belief space. Step 4: Evaluation fitness value of each particle in population space. Step 5: Updating the belief space by means of Update () function which gains the messages through Acceptance () function. Step 6: Producing a new generation of groups by Evolution () function based on the guidance of belief space. Step 7: Judgment. Estimating whether the program meets the convergence condition or reaches the maximum iteration number or not. If it can satisfy the condition, then withdraw it, otherwise return to the step 4. To understand the process of the program better please look at Figure 3. V. EXPERIMENT SETUP AND RESULT ANALYSIS An experiment is designed mainly to verify the validity and reliability of PSO-CA algorithm. This algorithm has been used in two hydropower stations. One (expresses as A)

Whether satisfied the condition of change?

No

Yes Crossover and mutation Guide the evolution

Produce a new generation of particle

Whether to meet the end condition? No Yes End

Figure 3. Algorithm process of PSO-CA

Setting parameters: the number of particle swarm is 25, c1=c2=2, StaMax=5, =0.3, nStaCount=10, K=500. In the experiment the problem has been solved with PSO-CA under the condition of Visual C++6.0. The results are given in table I. It is obvious that better result can be gotten by PSO-CA algorithm compared with PSO algorithm. The growth rate of output of A and B, respectively, is 2.7% and 5.42%. In other words, the average growth rate of the cascade hydropower stations is 3.36%. Figure 4 illustrates the convergence of algorithm during the iterative process. It is clear that PSO quickly falls into local optimum although its convergence speed is faster than that of PSO-CA. Nevertheless the belief space stores and manipulates the knowledge of species and in turn influences the evolution of the population. In that way, PSO-CA can

292

jump out the local optimum and eventually reach the global optimum. In terms of the relative diversity which represented as Figure 5, it is higher of PSO-CA than PSO before the iteration is end. This means that the relative diversity of PSO-CA keeps the high level before finding the optimal value through the crossover and mutation of population.
TABLE I. Hydropower station A B Total Total water inflow per day /km3 54324.0 72948.4 127272.4 RESULTS OF OPERATION OPTIMIZATION Total water consumption per day /km3 72948.4 71380.3 144328.6 Output by PSOCA /MWh 18099.0 5990.2 24089.2 Output by PSO /MWh 17623.0 5682.2 23305.2 Output growth rate /% 2.70 5.42 3.36

VI.

CONCLUSION

There is a variety of mechanisms which guide evolutionary search within population space. The goal of this paper is to use a CA to acquire knowledge which is necessary to control the evolution. In belief space, saving the information of outstanding individuals and ensuring the diversity of particles by shape knowledge to prevent the algorithm into local optimization; guiding the evolution of particle swarm with different selection by standardization knowledge to increase the speed of convergence. The experiment has verified that PSO-CA is a very important technique for solving the discreteness and nonlinear problems. It is testified that PSO-CA is a new and effective optimal method and better solution can be gained. So going to investigate the performance of PSO-CA in tracking the optimum on more complicated problems such as long-term optimal operation of cascade hydropower stations is suggested. REFERENCES
[1] Oliveira R, Loucks DP. Operating rules for multireservoir systems. Water Resources Research, 1997, 33(4), pp839-852. [2] Chang Jianxia, Huang Qiang, Wang Yimin. Optimal operation of hydropower station reservoir by using an improved genetic algorithm. Journal of hydroelectric engineering, 2001, (3), pp85-90. [3] Howson H R, Sancho NCF.A new algorithm for the solution of multistate dynamic programming problems. Math program, 1975, 8(1), pp104-116. [4] Kennedy J , Eberhart R . Particle Swarm Optimization. Proceedings of IEEE conference on Neural Networks. Perth, Australia, 1995, 4, pp1942-1948. [5] Wang Ni, Wang Shao-bo, Xie Jiangcang. Modified Particle Swarm Optimization And Its Application on Optimal Operation of Hydropower station. Third International Conference on Natural Computation. Haikou, China: IEEE, 2007, pp549-553. [6] Huang Qiang, Zhang Hong-bo, Chen Xiao-nan, Lv Yu-jie. Application of Particle Swarm Optimization algorithm to Reservoir Operation.Third International Conference on Natural Computation. Haikou, China: IEEE, 2007, pp595-600. [7] Van den BerghF. Analysis of Particle Swarm Optimizers. South Africa: University of Pretoria, 2001. [8] Zeng Jianchao, Jie Jing, Cui Zhihua. Particle Swarm Algorithm.Beijing: Science publishing company.2004. [9] Robert G. Reynolds. An introduction to cultural algorithms. Proceeding of the Third Annual Conference on Evolutionary Programming. New Jersey: World Scientific, 1994, pp131-139. [10] Zhang shuang-hu, Huang Qiang, Sun Ting-rong. Study on the optimal operation of hydropower station based on parallel recombination simulated annealing algorithms. Journal of hydroelectric engineering, 2004,23(4), pp16-19.

Figure 4. Convergence comparison

Figure 5. Comparison of the relative diversity

293

You might also like