Short-term load forecasting is affected by many factors. Euclidean distance in the traditional clustering analysis can"t measure the similarity between the load curves well. Therefore, Euclidean distance mixed with the cosine similarity is used to cluster the load curves of all kinds of users. Then information theory method is used to select the optimal combination in 9 kinds of associated factors .The user load and its associated factors whose kind is the same with the user who we want to predict its load are taken as the training sample data sets to establish the support vector machine forecasting model. Through the analysis of the actual sample data in a certain area of Shanghai, results proved that the average relative error of this method is 1.46% and 78.79% of the relative errors are below 1%. It has a better practicability.