Fuzzy support vector regressions for short-term load forecasting

The accurate short-term point and probabilistic load forecasts are critically important for efficient operation of power systems and electricity bargain in the market. Fuzzy systems achieved limited success in electric load forecasting. On the other hand, support vector regression models have seldom been part of a winning solution of the electric load forecasting competitions during the last decade. In this paper, we propose a methodology to integrate the fuzzy memberships with support vector regression (SVR) and support vector quantile regression (SVQR) models for short-term point and probabilistic load forecasting, respectively. One fuzzy membership function is proposed to efficiently calculate the relative importance of the observations in the load history. Three SVR and one SVQR models, including L1-norm based SVR, L2-norm based SVR, kernel-free quadratic surface SVR and SVQR models, are utilized to demonstrate the effectiveness of the proposed methodology. For point load forecasting, we compare the proposed fuzzy SVR models with a multiple linear regression, a feed-forward neural network, a fuzzy interaction regression, and four SVR models. For probabilistic load forecasting, the proposed fuzzy SVQR model is compared with a quantile regression model, a quantile regression neural network, and a SVQR model. The results on the data of global energy forecasting competition 2012, demonstrate that the proposed fuzzy component can improve the underlying SVR and SVQR models to outperform their counterparts and commonly-used models for point and probabilistic load forecasting, respectively.