Abstract: A short-term load forecasting method based on feature selection and combination model is proposed. At first, the method divides the feature vectors into two sets according to the individual characteristics. Spearman rank-order correlation coefficient and max-relevance & min-redundancy algorithm are individually employed for selection. Bayesian information criterion is used to get the dimension of the optimal feature vector. And then, three different simple-kernel based support vector regression models are built using three kernel functions respectively and complete prediction. Finally, a neural network is set up for experimental analysis. The simulation results show that the proposed combination model has a great high forecasting accuracy and robustness.