A method for short-term load forecasting based on kernel principal component analysis and BP neural network is proposed .Firstly, original input variables of neural network were got fully considering the load factors. Then, the dimension of original input variables was reduced by Kernel Principal Component Analysis method to get input variables of neural network. Finally, improved training algorithm of neural network was established to predict the load for every moment. Load of a region in Gansu Province in 2014 was forecasted by using the proposed method in the paper. Comparison with other two methods shows that the method in the paper can improve the accuracy of load forecasting.