Hybrid Intelligent Modeling Technique And Its Applications In Short-term Load Forecasting

For many complex systems, researchers have difficulties in identifying and predicting them due to their certain characteristics such as nonlinearity, process uncertainty, time variety, time lag, multivariable coupling, etc. The intelligent modeling techniques which are based on soft-computing methods provide an effective way for solving this kind of problems.The dissertation begins by indtroducing the concept, applications and development of intelligent modeling, mainly summarizing the hybrid intelligent modeling techniques based on fuzzy inference system and its applications in power systems. The dissertation could be divided into two parts:The first part mainly contributes to proposing several evolutionay fuzzy systems (EFSs), and the applications of them in predicting chaotic time series, identification and control of nonlinear dynamic systems and nolinear function modeling. This part includes four chapters, which can be summarized as follows:The second chapter proposes a novel hybrid algorithm EPLSE to design fuzzy rule base automatically, which is based on the combination of EP (Evolutionary Programming) and LSE (Least Squares Estimate). By utilizing the consequent parameters of the extended T-S model, the training error is decreased greatly. In the prediction for the Mackey-Glass time series, the proposed modeling methods shows its superiority compared with other methods. The idea of this method also gives inspirations to the proposal of the three-stage fuzzy modeling method in short-term load forecasting.In the third chapter, through analysing the merits and shortcomings of the two different type evolutionary algorithms: Evolutionary Programming (EP) and Particle Swarm Optimization (PSO), a novel hybrid evolutionary algorithm EPPSO is proposed. The hybrid algorithm EPPSO is applied in designing fuzzy identifies and fuzzy controllers for nonlinear dynamical systems.The fourth chapter proposes the core intelligent model of the dissertation: Adaptive Extended Fuzzy Basis Function Network (AEFBFN). The model assimilates the basic modeling idea of ANFIS. In AEFBFN, firstly, we use the Least Squares (LS) method to implement the fuzzy input space partition, and then the initialized fuzzy model will be tuned ulteriorly by the hybrid algorithm EPPSO. This idea provides a novel way to enhance the performance of the fuzzy model. As a new evolutionary fuzzy system (EFS), its effectiveness has been demonstrated in the nonlinear system modeling, chaotic time series prediction and STLF problem.The second part mainly contributes to 24 hours ahead short-term load forecasting, developing different methods for the STLF problem in the whole year. The whole work is based on the implementation of identification and revision of the outliers in the historical load data. The fith chapter to the seventh chapter constitutes this part.The fifth chapter is devoted to the data cleaning of the historical electrical load data. Based on the statistical methods, we are capable of identifying the abnormal data. And then, the artificial immune network and the graph clustering algorithm are utilized to extract the typical load curve, which is used to correct the abnormal part of the historical load data.The sixth chapter presents a three-stage weather sensitive STLF algorithm, based on a novel fuzzy modeling strategy using Least Squares (LS) method and the hybrid algorithm EPPSO. In the first stage, the LS method is used to design the Fuzzy Basis Function Network (FBFN) and thus completes the fuzzy space partition of STLF fuzzy models. At stage two, the obtained FBFN is firstly extended to a 1st-order Takagi-Sugeno (T-S) fuzzy model, and then the hybrid algorithm EPPSO is used to tune the premise parameters and learn the consequent parameters of the fuzzy model simultaneously. At the last stage, the hourly load forecasting errors using the evolved fuzzy model are regarded as a new time series to be identified by merely the weather variables, and the identification form could be expressed as a regression model and thereby identified using the LS method. This chapter has performed the STLF of Zhejiang Electrical Power Company (EPC) for all of the days in 2001 except for the holidays. For the weather insensitive days in spring and autumn, we use the STLF fuzzy model constructed based on only the former two stages. The testing results and comparisons with other methods demonstrate the effectiveness of the proposed three-stage hybrid algorithm for generating STLF fuzzy model.The seventh chapter proposes a novel STLF method for the holidays, in which the problem is divided into two questions: to forecast the scaled load curves (SLC) of holidays and to forecast the maximum and the minimum loads of holidays. For the first question, the evolutionary fuzzy system AEFBFN is used directly. While for the second question, we use the fuzzy regression method to foreast the maximum and the minimum loads of the holidays. Finally, the results of the AEFBFN and the fuzzy regression method are combined to forecast the 24 hourly load of holidays.