Study On Short-term Load Forecasting Based On Support Vector Machine

Short term load forecasting (STLF) is the precondition of economic and secure operation of power system, and because the power systems are getting more and more marketable, STLF with high quality is getting more and more important and exigent. Support vector machine (SVM) is a novel learning machine with many merits such as fast solving and strongly generalizing abinity. In this paper, some researches are developed for STLF using SVM in several parts.Model parameters influence the performance of SVM evidently, but structuring method for parameters selecting is lacking and parameters are generally determined by crossing test. Particle swarm optimization (PSO) algorithm is a fast and simple intelligent algorithm, In this paper, it is studied to use PSO algorithm to optimize the parameter selection of SVM, and then the forecasting model is constructured, which is called PSO-SVM model.Based on the history data from a practicle power system in Zhengzhou city, the characteristics of load data and influencing factors are analysed. Based on statistics, "disorder data" can be removed, and then accurate and effective load forecasting can be ensured. The influencing factors are processed by normalization, and the conception of "similar degree" of characteristic variables is used for training data selection.Based on load curves are very close when the day characteristics are similar, it is researched to use PSO-SVM model and similar day method for next day load forecasting. Because of the poor meteorological sensibility of SVM and fuzzy logic specializes in processing uncertain information, the fuzzy rules are constructed to forecast stochastic load influenced by weather factors, which is used to correct the forecasted load curve of PSO-SVM, and then the final results are gained.It is a hard work for special day load forecasting. The PSO-SVM model and fuzzy reference are combined for special day load forecasting here. The work of special day load forecasting is devided into two parts. Fuzzy inference method is used to forecast the day maximal load and day minimal load, and PSO-SVM is used to forecast the load scaled curve. The special day load is forecasted by combining the forecasting results of above two methods.