1

Effects of Multi-Objective Genetic Rule Selection on Short-term Load Forecasting for Anomalous Days
Li Feng, Ziyan Liu

Abstract-- One advantage of multi-objective genetic optimization algorithms over classical approaches is that many non-dominated solutions can be simultaneously obtained by their single run. In this paper, we proposed a fuzzy rule-based classifier for electrical load pattern classification by using multiobjective genetic algorithm and fuzzy association rule mining. Multi-objective genetic algorithm is used to automatically select the rules with better classification accuracy and interpretability, and the key concepts of fuzzy association rule mining are the bases of heuristic rule selection for improving the performance of genetic algorithm searching. Through computation experiments on a real power system, it is shown that the generated fuzzy rulebased classifier leads to high classification performance, and can supply more sufficient historical data for load forecasting of anomalous days, better performance of load forecasting is gained accordingly. Index Terms-- Association rule mining, Fuzzy rule-based classifier, Multi-objective genetic algorithm, Load forecasting

I. INTRODUCTION hort-term load forecasting plays an important role in power systems since the improvement of forecasting accuracy results in the reduction of operating costs and the reliable power system operations. Due to the simplicity of use and the excellence in performance, various types of artificial neural network (ANN) have been proposed for short-term load forecasting [1]-[3]. However, conventional ANN based short-term load forecasting methods have several drawbacks. Neural network based forecasting models, regardless of their level of sophistication, make use of only numerical data, and are unable to handle fuzzy information that are inherent in human information processing. The biggest drawback of the ANN forecasting model is its inability to accurately forecast on special days in anomalous load condition, since sufficient data sets to train the network for peculiar load profiles on these special days are usually not available. And the load patterns of special days are difficult to
Li Feng is with the Dispatch and Transaction Center, Chongqing Electrical Power Corporation, Chongqing 400014, P.R.China. (e-mail: [email protected] ). Ziyan Liu is with College of Computer Science and Communication Technology, Guizhou University (Cai Jiaguan Campus), Guiyang, Guizhou 550003, P.R.China. (e-mail: [email protected] ).

determine for the random fluctuant factors in these days. There have been several research works for developing accurate load forecasting techniques for special days. A combined approach was presented using a Kohonens Self Organizing Map and a multi-layer perceptron to forecast the loads during the anomalous load condition periods in [4],[5]. In this approach, the Kohonens SOM provides a day-type identification code for each input pattern through unsupervised learning stage. In [6] a special adjusting techniques that compensated the negative impact of holidays forecasts is applied for improving holidays forecast. A fuzzy expert system modifies the base load forecasted by an ANN model in [7], and by adding the specific fuzzy rules obtained from the knowledge and experiences of operators for holiday, the model was modified. In spite of improving the forecasting accuracy, this model did not have an efficient architecture because the knowledge base consisted of too many fuzzy rules. These methods lack the ability to handle uncertainty and common-sense knowledge frequently used by human experts to forecast load profiles. The predictive accuracy of forecasting models depends on projecting values and interrelationships of important parameters (weather, etc.) that are themselves intrinsically uncertain. Hence, if the historical data can be clustered by those important parameters and experts knowledge, it can provide appropriate training data for ANN, which is good for improving the forecast performances consequently. With the development of intelligent theory, data mining and optimization strategy, there are more effective measures to manage modes classifying, and they are applied in power systems in decades. Here we propose a hybrid short-term load forecasting method for special days in anomalous load conditions, such as public holidays, consecutive holidays, and days preceding or following holidays. Combining with associate rules mining, multi-objective genetic algorithms and fuzzy classification systems, a novel load pattern classification system are constituted according to the load characters and weather conditions, in which associate rules mining and multi-objective genetic algorithms are used to simplify and optimize fuzzy rules constructed by a fuzzy classification system. And then, the optimal load patterns are applied to select proper historical data for ANN training. The proposed method was shown to improve the forecasting

1-4244-0493-2/06/$20.00 2006 IEEE.

accuracy of special days by the simulation results. II. CLASSIFICATION SYSTEMS BASED ON FUZZY RULES Fuzzy systems are particularly suitable for modeling, control and classification problems [8]-[10], as a human expert is able to analyze and comprehend the knowledge stored in the form of linguistic variables and rules. Classifier induction is a specific data mining task [11] that is part of the knowledge discovery process whose aim is the extraction of accurate, comprehensible and interesting knowledge [11], [12], thus suggesting that several objectives have to be optimized while inducing a classifier. Let us assume that our pattern classification problem is an n-dimensional problem with K classes. We also assume that we have M given training patterns xp={xp1,xp2,,xpn} p=1,2,,M. The number of the samples for the k-th pattern is Mk, obviously M = M k . Without loss of generality, each
k =1 K

attribute of the given training patterns is normalized into a unit interval [0,1]. In this paper, we use fuzzy if-then rules of the following type in our fuzzy rule-based classification system: Rule Rj: if x1 is Aj1 and and xn is Ajn Then Class Cj with CFj, j=1,2,,N (1) where Rj is a label of the j-th fuzzy if-then rule, Aj1,, Ajn are antecedent fuzzy sets on the unit interval [0,1], Cj is the consequent class (Cj=1,,K), CFj is the grade of certainty of the fuzzy if-then rule Rj, and N is the total number of fuzzy ifthen rules. The consequent class Cj and the rule weight CFj are determined in a heuristic manner from compatible training patterns with the antecedent part of Rj. For more details of Cj and CFj one can refer to [13]. Now, in the researches of fuzzy classification systems more emphasis has been mainly placed on the classification accuracy rather than the comprehensibility of fuzzy rules. Thus the performance maximization has been the primary objective. However, in real world the rules with good comprehensibility are easily wieldy and modifiable for decision maker. Therefore how to enhance the comprehensibility of fuzzy rules becomes an important problem of fuzzy classification systems. III. MULTI-OBJECTIVE GENETIC ALGORITHM In real world most problems are composed by many objectives, and these various objectives often conflict mutually in the high-dimensional spaces. Multi-objective optimization is to provide a set of evenly distributed possible solutions and select the satisfactory solutions for the decision maker [18]. This is known as the concept of Pareto optimality, and the solution set is known as the Pareto-optimal set. Pareto-optimal solutions are also called non-dominated solutions. The simultaneous optimization of competing objective functions deviates from the single function optimization in that it seldom admits a single, perfect solution. Single

objective optimization has only one objective function, a global optimum is found in general after the optimization. Instead, multi-objective optimization problems tend to be characterized by a family of alternatives that must be considered equivalent in the absence of information concerning the relevance of each objective relative to the others. Evolutionary algorithms seem particularly desirable to solve multi-objective optimization problems, because they have good searching abilities in complex spaces, and deal with a set of possible solutions which allows to find an entire set of Pareto optimal solutions in a single run of the algorithms, instead of having to perform a series of separate runs as in the case of the traditional mathematical programming techniques. Since the Vector Evaluated Genetic Algorithm (VEGA) proposed by Schaffer (1985) [19], genetic algorithms have been successfully applied in multi-objective problems [20][22]. Let us consider the following n-objective optimization problem: (2) Minimize F=( f1(x), f2(x),, fn(x)) where F is the objective vector, x is the decision vector, and X is the feasible region in a decision space. If the following solution x* is found in the feasible region X, x* can be said as the optimal solution for the n-objective optimization problem: (3) f i ( x* ) f f i ( x) for i {1, 2,..., n} and x X Thus the task of evolutionary multi-objective algorithms is not to find a single final solution but to find all solutions that are not dominated by any other solutions. Let a and b be two decision vectors ( a, b X ). Then b is said to be dominated by a (i.e., a p b ), if and only if the following two conditions hold: fi (a) fi (b) for i {1, 2,..., n} (4) fi (a) < fi (b) for i {1, 2,..., n} when b is not dominated by any other solutions in X, b is said to be a Pareto-optimal solution. That is, b is a Pareto-optimal solution when there is no solution a in X that satisfies the above two conditions.

IV. ELECTRICAL LOAD PATTERNS CLASSIFICATION SYSTEM Generally, the load shapes of anomalous days are distinct from those of the weekday and weekend. And even for the same holiday the load shapes varied with the variation of weather and load levels of every year. Since it is difficult to obtain sufficient data set for learning those peculiar load profiles, the accuracy of load forecasting of the artificial neural networks based model depressed for the deficient historical data. So in the paper we classified the historical data, according to the correlative factors of electrical load, into K separate classes. A. Generating Fuzzy If-Then Rules The maximum and minimum value of the electrical loads (LDmax and LDmin) of each day are normalized to the same range of values between 0 and 1 by xscale (t ) = ( xactual (t ) xmin ) ( xmax xmin ) , where xmax and xmin

denote the maximum and minimum value of the vector xscale respectively, t=1,,24, and the normalized load are clustered into K classes that are defined as a set of load patterns. The weather conditions, including maximum temperature (Tmax), minimum temperature (Tmin), wind speed (Ws), humidity (Hu), Rainfall (Rf), and the change of the loads of two adjacent days are defined as input vectors of a classification system. A single fuzzy partition was used for all inputs that are divided into five fuzzy subsets [Very Low (VL), Low (L), Normal (NM), High (H), Very High (VH)], as shown in Fig.1. The fuzzy rules are generated automatically in the form of (1).

where Aq and Class h are the antecedent part and consequent class of an association rule A q Class h . The confidence c
indicates the grade of the validity of A q Class h . The support s indicates the grade of the coverage by A q Class h . The compatibility grade Aq ( X p ) is defined
by the product operator as: A ( X p ) = A ( X p1 ) A ( X p 2 ) K A ( X pn )
q q1 q2 qn

(7)

where Aqi () is the membership function of the antecedent fuzzy set Aqi. h=1,,K, p=1,,M. Generally it cannot be affirmed that the rules constructed by the minimum confidence and support are interesting for users. There may exist many redundant and nonsense rules which have passive impact on classification. Hence, another criteria interest is introduced. interest measures the attention degree of the users on the association rules with the consideration of novelty, availability and comprehensibility of knowledge. interest is defined as 1 s(Class h) (8) interest (A q Class h) = (1 s (A q )) (1 s(A q Class h)) Here we define a criterion fcr to extract candidate rules from the generated fuzzy rules that are divided into K groups according to their consequent classes. (9) f cr = 1 s c + 2 interest where 1 and 2 are nonnegative weights between 0 and 1. Fuzzy rules in each group are sorted in a descending order of the criterion fcr. For selecting N candidate rules, the first N rules are chosen form each of the M groups. In this manner a pre-specified number of candidate rules are chose as candidate rules in our genetic algorithm-based rule selection method. With those novel and concise rules, the searching spaces of genetic algorithm are reduced effectively, and the searching ability of genetic algorithm is enhanced accordingly. C. Genetic Rule Selection In fuzzy rule system more rules are needed for higher classification accuracy, while the comprehensibility of fuzzy rules decreases, and the complexity of the classification system enhanced accordingly. Considering the balancing relations between the classification accuracy and comprehensibility, we propose a multi-objective genetic algorithm to find the Pareto-optimal rules with high classification accuracy and high comprehensibility in the paper. Our rule selection problem is formulated as the following three-objective optimization problem: (10) f ( S ) = w1 f1 ( S ) w2 f 2 ( S ) w3 f 3 ( S ) where f1(S) is the number of correctly classified training patterns by the fuzzy rule sets S, f2(S) is the number of fuzzy rules in S, and f3(S) is the total number of the antecedents of fuzzy rules in S. wi is the weight of the target fi (i=1,2,3) satisfying wi [0,1] and w1 + w2 + w3 = 1 . Because we want to find the rules with high classification accuracy and high comprehensibility, the searching process of

Fig.1 Membership degree

Since the number of fuzzy rules increases exponentially along with the increasing of the input vectors in a fuzzy system, there will be a dimensional disaster by using the traditional methods to generate fuzzy rules. Redundant fuzzy sets and rules would not only increase the complexities of a system, but make it difficult to describe a system in natural languages. And the optimum resolution is difficult to find, even does not exist in high-dimensional input spaces. Moreover, the maximization of any accuracy measure does not always mean the maximization of their actual performance, because it often leads to the overfitting to training patterns, which degrades the actual performance of classification systems on test patterns. Recently the tradeoff between the accuracy and the comprehensibility was discussed in some studies [14],[16],[17],[23]. In the paper, multi-objective genetic algorithms are applied to manage the problem. In order to gain the rules with good classification performances, we suggest the use of a data mining technique for extracting a pre-specified number of candidate rules in a heuristic manner. B. Heuristic Rule Extraction In the field of data mining, association rules are often evaluated by two rule evaluation criteria: support and confidence. The fuzzy version of the two evaluation criteria [24] is given below. s(A q Class h) = s(A q Class h) (5) 1 A ( X p ) = M X Class h
q p

c (A q Class h) = s (A q Class h) s (A q ) =
X p Class h

A ( X p )
q

p =1

Aq

(X p)

(6)

genetic algorithm is to maximize the first item and minimize the other two items. The last two items can measure the complexity of classification system, and avoid the overfitting of training sets. In order to avoid immersing in some local optimum and obtain the evenly distributed Pareto-optimal solutions, we use NSGA-II [25] to enhance the global searching ability of genetic algorithms. It has many merits on keeping population diversities effectively, finding the global optimal solutions and retaining the individuals near Pareto-optimal solutions by elitisms strategy. The processes of the multi-objective genetic algorithm are given below. Step1 Population initialization Let us assume that N fuzzy rules have been extracted as candidate rules using the fcr criterion. A subset S of the N candidate rules is handled as an individual in the algorithms, and represented by a binary string of the length N as S=s1s2sN where sj=1 and sj=0 mean that the j-th candidate rule is included in S and excluded from S, respectively. Step 2 Selection Generate parent strings from the current population using binary tournament selection. Step3 Crossover and mutation A standard crossover operator is applied over two strings. For finding some simple rule sets, we use two uniformed mutation probability where a larger probability is assigned to the mutation from 1 to 0 than that from 0 to 1. This is for efficiently decreasing the number of fuzzy rules in each rule set. Step4 Evaluating the chromosome Calculate the three objectives of each of the newly generated strings. We remove those fuzzy rules that cannot degrade the first objective after we calculate the first item of the fitness function, which decreases the complexity for calculating the other two objectives and improves the performance of the second and the third objective. Evaluate the fitness of each individual according to the given fitness function in (10). The weights wi in (10) are specified randomly whenever a pair of parent strings is selected. It has an advantage to search for a variety of non-dominated rule sets in the three-dimensional objective space. Step5 Elitist Strategy Since the Pareto-optimal solutions in each population have the same importance. In order to find the optimal solution, we sort the parents and children population by the nondomination ranks. The individuals with high rank are saved and the inferior individuals are eliminated. If there are some individuals that belong to the same front, then the crowding distance computation [25] is applied. The individual with lesser-crowded region is selected as the optimal solution, which can preserve the population diversity effectively. Step 6 Termination Test If the stopping condition is not satisfied, return to Step 2. Otherwise terminate the execution of the algorithm.

D. Neural network forecasting In this method, the entire load data are classified into four different day types. The classified historical load data are applied to anomalous load forecasting in the following section. The ANN models with three layers are trained using the load data of 2000-2001, and tested with the data of from March to Oct. of 2002. In order to avoid the generalization ability decreasing of the neural networks by the mixed load data of different patterns, different ANN models for each day type are implemented. 1) Input vectors The input vectors of a neural network for weekend load forecasting are the historical load and weather data of four nearest days that have the same load patterns as that of the day we want to have a forecast. And for holidays and their adjacent days forecast they are the data of four nearest days with the same load patterns that are selected from the data within half a month before the forecasting day every year. All the input vectors are fuzzified as described in subsection IV. 2) Output vectors The output vectors are defuzzified by , which are the final xactual = xscale ( xmax xmin ) + xmin forecasting load. V. SIMULATION RESULTS We apply the proposed method to the historical data from 1999 to 2002 that include the historical load data and the weather conditions in classification processes. All the attribute values are normalized into a real number in the unit interval [0,1]. According to the analysis in Subsection IV, we specify the class K=4, which refers to four patterns weekday, weekend, holiday and the day adjacent to holidays. The load data set is handled as a 4-class pattern classification problem in the 6-dimensional unit hypercube for 6 weather and load conditions. Fuzzy rules are constructed by a heuristic method. The total number of possible combinations of antecedent fuzzy sets is 56. Among those combinations, a pre-specified number of candidate rules are generated in a heuristic manner by using criterion fcr. A. Performance Analysis of the Pattern classification Calculate fcr of each fuzzy rule in the heuristic rule extraction. Fuzzy rules in each group are sorted in a descending order of the criterion fcr, and the first 100 rules are chosen from each of the K groups as the candidate rules for genetic rule selection. Then NSGA is applied to the candidate rules using the following parameter specifications. Population size: 50. Crossover probability: pc=0.85. Mutation probability: pm1=0.1, pm2=0.005. Stopping conditions: 3000 generations. For evaluating the generalization ability of obtained rule sets, we divided each data randomly into 5 subsets of the same size. One subset was used as test patterns while the other four subsets were used as training patterns. The generalization ability of obtained rule sets was evaluated by

classifying test patterns. The computer simulations were iterated 5 times for each data set so that all the five subsets were used as test patterns. Pareto-optimal rule sets by which the patterns of each object can be identified correctly were gained in iterations. In this paper each pattern was independently classified by each individual non-dominated rule set. Then the majority class was chosen as the final classification result of that pattern by the classifier. Table 1 shows the classification performance of two Pareto rule sets for training sets and test sets. From Table 1, we can see that our rule selection method found various rule sets with different classification rates and different sizes, and the two given classification rule sets both have good classification performances. The maximum and minimum error for training sets is 13.21% and 0 respectively, and the maximum and minimum error for test sets is 15.32% and 1.37% respectively. In the second rule set, the classification performance for training sets reaches 100% when the number of the rule is 58, at the same time the classification error for test sets touches bottom. In the first rule set, error rates on test patterns tend to increase with the number of fuzzy rules while error rates on training patterns monotonically decrease. The overfitting to training patterns is due to the increase of fuzzy rules. Hence, individual classifiers with low error rates on training patterns did not always have low error rates on test patterns. This makes it very difficult to choose a single appropriate classifier from the multiple alternative ones. The aggregation of many classifiers using a voting scheme avoids this difficult task. In Table 2, the maximum and minimum errors of test sets for individual classifier are 15.32% and 1.08% respectively, while for the classifier ensemble they are 11.68% and 1.37% respectively. We can see that the performance of our classifier ensembles was better than many individual classifiers.
TABLE 1 PERFORMANCE OF THE FUZZY RULE SETS

Here we use the classifier ensemble to determine the load patterns of each day. A part of the correlative days under the same load pattern of five anonymous days are listed in Table 3. The number in a bracket behind a date refers to the day in a week, that is, 1 represents Monday, 2 represents Tuesday, etc. the date having the same load pattern as 10th, Mar. in 2002 are 2nd, 3rd, 4th, 5th, 9th. It matches the cases in real world. From Table 3 we can observe that holidays have the similar characters, and the day under the same load pattern may have a different day type in a week. It is highly advantageous to us to use the limited historical data of holiday effectively. Moreover the holidays have an impact on the load of the day preceding or succeeding holidays, especially the consecutive holidays. Then when regarding the load forecasting of the days preceding or succeeding consecutive holidays, the loads in those days are considered as a separate load pattern when forecasting in the paper. By using the proposed fuzzy classification system those days with the similar load characters are found, and the appropriate data are supplied to neural networks for forecasting accordingly.
TABLE 2 COMPARISON OF THE MAXIMUM AND MINIMUM ERRORS

TABLE 3 SIMILAR DAYS IN THE SAME LOAD PATTERNS

B. Simulation on Load Forecasting We trained the ANN models using the classified load data

of 2000-2001, and tested with the data of Mar. Oct. 2002. In the paper the proposed forecasting algorithm is compare with a traditional method Similar Day Algorithm (SDA) in which the historical data of those similar days to the forecasting day are selected by their similarity degree of the average temperature. Electrical loads and Weather factors [wind speed (Ws), maximum and minimum temperature (Tmax and Tmin), mean humidity (Hu) and rainfall (Rf)] of the correlated day are chosen as the input vectors of an ANN that is trained by using an error backpropagation algorithm with an adaptive learning rate and momentum. And then the test result was compared with the forecasting result of SDA. The forecasted load curves of the proposed algorithm based on fuzzy classification system (AFCS) and the traditional similar day algorithm (SDA) on Spring Festival and 5th Oct. compared with the actual load curve are illustrated in Fig.5 (a) and (b). 5th Oct. is a weekend behind National Day. In China National Day and Spring Festival is the most important consecutive holiday of a year. In general their load values are lowered to levels than other special days, and the load forecasting on those days are always more difficult than other special days. From this figure, we can observe clearly that the load curve of AFCS is closer to the actual loads than traditional SDA, and the mean error of MAPE for AFCS on the two days are 3.176% and 1.532% respectively; and for SDA they are 3.973% and 2.056% respectively. Then better prediction accuracy is obtained by AFCS. The comparison of our results with SDAs results is illustrated in Table 4, where the numbers represent the mean errors of MAPE (mean absolute percentage error). The performance of the maximum prediction error of SDA is 5.12% on National Day 2002, however AFCS has yielded a better performance of prediction error 4.681%. And the minimum prediction error of AFCS on 5th, Oct. is 1.532%, while SDA prediction error on that day is 2.056%. So the forecast was improved by the proposed method to select proper historical data. VI. CONCLUSION In order to improve the forecast of anomalous days, sufficient and accurate historical data are needed. In the paper combined with associate rule mining and multi-objective genetic algorithms, a pattern classification system based on fuzzy rules was constituted. Associate rule mining was used to select interesting candidate rules, and genetic algorithm was applied to simplify and optimize fuzzy rules. The system can provide appropriate and comprehensive historical data for load forecasting of anomalous days, as results the limited information are utilized sufficiently, and good classification performances are gained. It can provides a more accurate and effective forecast for load forecasting of anomalous days. VII. REFERENCES
[1] Chow T W S, et al. Neural network based short-term load forecasting using weather compensation, IEEE Trans. on Power Systems, vol.11, no.4,

[2]

[3]

1996, pp.1736-1742. D.C. Park, M.A. El-Sharkawi, and R. J. Mark II, Electric load forecasting using an artificial neural network, IEEE Trans. on Power systems, vol.6, no.2, pp.442-449, May 1991. K.H. Kim, J.K. Park, K.J. Hwang, and S.H. Kim, Implementation of hybrid short-term load forecasting system using artificial neural networks and fuzzy expert systems, IEEE Trans. on Power Systems, vol. 10, no. 3, pp.1534-1539, Aug. 1995.

(a)

(b) Fig. 5. Comparison between actual and forecasting load curves of AFCS and SDA TABLE 4 FORECASTING ERROR IN MAPE

[4]

[5]

R.Lamedica, A.Prudenzi, M.Sforna. A neural network based technique for short-term forecasting of anomalous load periods, IEEE Trans. on Power Systems, vol.11, no.4, 1996, pp.1749-1756. G. Chicco, R. Napoli, F. Piglione. Load pattern clustering for short-term load forecasting of anomalous days, In Proc. of IEEE Porto Power Tech, vol.2, Sep. 2001, paper AIT2-377.

7 [6] M. A. Aboul-Magd, Emad El-Din El-Sayed Ahmed Ibrahim. An artificial neural network model for electrical daily peak load forecasting with an adjustment for holidays, in Proc. of 2001 IEEE Large Engineering Systems Conference on Power Engineering, pp.105-113. J.K. Park, K.H. Kim, K.J. Hwang, G.W. Kim, and S.H. Kim, Development of a short-term load forecasting expert system, in Proc. of the international conference on Intelligent System Application on Power Systems, July, 1997, pp.508-513. M.Sugeno, An introductory survey of fuzzy control, Information Science, vol.30, no.1/2, pp.59-83, 1985. C.C. Lee, Fuzzy logic in control systems: fuzzy logic controller part I and part II, IEEE Trans. Syst., Man, Cybern., vol.20, pp.404-435,1990. Z. Chi, H. Yan, T. Pham, Fuzzy algorithms with applications to image processing and pattern recognition, Singapore: World Scientific, 1996. A. Freitas. Data mining and knowledge discovery with evolutionary algorithms, Berlin: Springer Verlag, 2002. U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth. From data mining to knowledge discovery in databases, in Advances in knowledge discovery and data mining, U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth, Ed. Menlo Park: AAAI/MIT pp.1-36, 1996. Nakashima, T., Nakai, G., Ishibuchi, H. A fuzzy rule-based system for ensembling classification systems, in Proc. 2002 IEEE International Conference on Fuzzy systems, May 12-17, Honolulu, USA, pp.1432-1437. Jin, Y. Fuzzy modeling of high-dimensional systems: complexity reduction and interpretability improvement, IEEE Trans. on Fuzzy Systems No.8, 2000, pp.212-221. Roubos, H., and Setnes, M. Compact and transparent fuzzy models and classifiers through iterative complexity reduction, IEEE Trans. on Fuzzy Systems, no.9, 2001,pp.516-524. Setnes, M., Babuska, R., Kaymak, U., and Van Nauta Lemke, H. R. Similarity measures in fuzzy rule base simlificaiton, IEEE Trans. on Sytems, Man, and Cybernetics Part B: Cybernetics, no.28, 1998, pp.376386. Ishibuchi, H., and Yamamoto, T. Fuzzy rule selection by data mining criteria and genetic algorithms, in Proc. 2002 Genetic and evolutionary computation conference, pp.399-406. Carlos M. Fonseca, Peter J. Fleming. An Overview of Evolutionary Algorithms in Multiobjective Optimization, Evolutionary Computation, vol.3, no,1, 1995, pp. 1-16. Schaffer J D. Multiple objective optimization with vector evaluated genetic algorithms, in Proc. 1985 the 1st International Conference on Genetic Algorithms, pp.93-100. Fonseca, C.M., Fleming, P.J. Genetic algorithm for multiobjective: formulation, discussion and generalization, in Proc. 1993 the 5th International Conference on Genetic Algorithms, pp.416-423. Carlos A., Coello Coello. A Comprehensive Survey of EvolutionaryBased Multiobjective Optimization Techniques, Knowledge and Information Systems, vol. 1, no. 3, August 1999, pp. 269-308. Andrzej Osyczka. Multicriteria optimization for engineering design, in Design Optimization, John S. Gero, Ed. New York: Academic Press, 1985, pp.193-227. Jin, Y. Fuzzy modeling of high-dimensional systems: complexity reduction and interpretability improvement, IEEE Trans. on Fuzzy Systems, no.8, 2000, pp.212-221. R. Agrawal et al. Fast discovery of association rules, in Advances in knowledge discovery & data mining, U. M. Fayyad et at. Ed. Menlo Park: AAAI, 1996, pp.307-328. Deb K., Pratap A., Agarwal S., Meyarivan T. A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE Trans. on Evolutionary Computation, vol.6, no.2, 2002, pp.182-197. now with College of Computer Science and Communication Technology, Guizhou University. Her current research interests are communication engine, intelligence control, microcomputer control and data mining.

[7]

[8] [9] [10] [11] [12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

VIII. BIOGRAPHIES
Li Feng received her B.S. and M.S. degrees in electrical engineering from Guizhou University of technology, China, in 1998 and 2001 respectively, and her Ph.D. degree in electrical engineering from Zhejiang University, China, in 2005. She is now with the Dispatch and Transaction Center, Chongqing Electrical Power Corporation. Her current research interests are data mining, electrical market analysis, intelligent control in power system. Ziyan Liu received her B.S. and M.S. degrees in electrical engineering from Guizhou University of technology, China, in 1997 and 2000 respectively. She is