Short-term Power Load Forecasting Based On Data Mining Technology

Short-term load forecasting (STLF) is the precondition of economic and secure operation of power systems. With power systems getting more and more marketable, STLF with high quality is getting more and more important and exigent. Based on various data mining technologies, such as RS, SVM and NN , the author aims at each stage of STLF and has done deep research on the characteristic of load sequence, pre-process of historical load data, process of weather condition, establishment of forecasting model and its input parameters mining. All these work has laid a solid foundation for hi-accuracy STLF software development.Input parameters choosing is a big problem in SVM modeling which imposes great influence on the accuracy of forecasting. Rough Set reduction algorithm in data mining is used to solve input parameters choosing which can enable the confirmation of key property combination and structure construction and ensure the rationality and accuracy of input parameters.Model parameters influence the performance of SVM evidently, but the methods for parameters selecting always base on experiment. In this paper, it is studied to use Grid search to optimize the parameters selection of SVM, and then the forecasting model is constructed. The experimental results show that, when compared against BP neural network method, the proposed method based on RS and SVM can forecast more accurate results while shortening the training time.Finally, the paper also proposes a short term load forecasting (STLF) method by using clustering based support vector regression model. The proposed method is first based on self-organizing feature map (SOFM) that can discover the similar input data and cluster them into several subsets in an unsupervised strategy. Then, several SVR models are constructed in corresponding to the subsets; each SVR model is trained with its corresponding subset. Due to the similarity in training data and the reduction of the amount of training data for each SVR model, the proposed method can forecast with more accurate results while enhancing the training speed compared against BP neural network method and SVM.