A novel hybrid method based on SVM and linear regression for short-term load forecasting was presented. It is well known that temperature information is very important for load forecasting, but the local structure of temperature sensitive information is not adopted in the literature. The proposed model adopts an integrated architecture to handle the local temperature sensitive information. Firstly, the input load data set is clustered into several temperature similar days subsets by the k-means algorithm in an unsupervised manner, Then compute the temperature correction in each subsets and split the time point (5 minutes, 288/day) into three time zone: the most temperature sensitive time zone, the less temperature sensitive time zone, and partial temperature sensitive time zone. The most temperature sensitive time zone is forecasted by the linear regression, the less temperature sensitive time zone is coded only using past load data, and then use the generic support vector regression to forecast the next day load in that time point, the partial temperature sensitive time zone is coded combining the past load and temperature information and then using the support vector regression same as the less temperature sensitive time zone. Finally, we smooth the whole forecasted load curve using linear programming. The empirical results indicate that our hybrid method results in better forecasting performance than the original generic support vector regression.