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

The Short-term Load Forecasting (STLF) plays an important role both in production planning and routine operation style of a power system. It is an important means to insure electric network safety and stability. With strict mathematics theoretical basis, Support Vector Machine (SVM), a kind of machine learning method which is recently advanced, can achieve more excellent properties. In this paper, SVM is introduced to STLF. Experimental results are discussed to illustrate the proposed method.After reviewing the present condition of research on STLF, the common methods of STLF are concluded. Then the basis of SVM is introduced and the detailed procedures of using SVM for regressive problem are given. More discussions about forecasting process, the method of sample construction, the election of parameters, and training algorithms are presented in this assay. After that, the concrete application of the method in STLF is introduced.Considering that weather have great effects on STLF, a sample construction method is proposed in this paper, in which temperature, humility, sunlight and air pressure are calculated. Compared with the method which only based on historical data, the proposed method has better performance. The experiment results on predicting load of the typical day validate the above conclusion.According to the characteristics of electric power system load, in this paper an improved SVM method is presented. By inducting two variables, empirical error of the training sample is treated differently based on the time span between the forecasting sample and the training sample. The results of the examples indicate that this algorithm can improve accuracy of STLF.In the end, some researches on the construction of kernels are made. In order to obtain a flexible kernel function, a hybrid kernel function is developed. Experimental results are discussed to illustrate the proposed method and show that the SVM with the hybrid kernel outperforms that with a single common kernel in terms of generalization power.