Short-term Load Forecasting Methods And Applied Research Based On Least Squares Support Vector Machine

Short-term load forecasting is the precondition of operation, dispatch and decision-making of power system. Accurate short term load forecasting has a significant impact on control of power system, so this research is valuable.This dissertation sums up the characteristics of short-term load forecasting, and concludes its common selection of input variables. Based on statistics, the historical data are preprocessed such as disorder data are removed and data are normalized. The input vector is usually selected with human experience in least squares support vector machine(LS-SVM) forecasting model. This makes the adaptability of the model not good. In this dissertation, rough sets are used to analyze the condition attributes, and the attributes closed to load can be selected from the candidates set which contains irrelevant and redundant variables automatically, which are then applied to the LS-SVM as the effective input vector to forecast load. Meanwhile binary genetic algorithm is used to reduce the attributes. So this method can realize the selection of input variables optimization, reduce the dependence on experience in the course of prediction model established and enhance the adaptability of the model.Based on this, two important parameters of LS-SVM model are analyzed inducing that model parameters influence the performance of LS-SVM evidently. But at present parameters are generally determined by experience or crossing test. So this dissertation proposes to use floating genetic algorithm for adaptively optimizing the parameters of LS-SVM, and establish the forecasting model.Integrating the above research, for short-term load forecasting problem, an effective model and algorithm of LS-SVM combining rough sets and genetic algorithm are proposed and programmed. In this model and algorithm, the historical data are preprocessed by rough sets to analyze the condition attributes and obtain the factors closely related with load, which are then applied to the LS-SVM as the input vector to forecast load. During the model training process, it also uses floating genetic algorithm for adaptively optimizing the parameters of LS-SVM to improve the load forecasting accuracy. Shandong power grid is analyzed to exhibit the effectiveness of the proposed approach.