In neural network based short-term load forecasting, complexity and redundancy of input data have a negative effect on network training efficiency and forecasting precision. Focusing on solving this problem, a multiple method of data processing is developed. Firstly a method called input variable contribution analysis is applied, which divides input variables into primary variables and minor variables according to their contribution to network output. Minor variables are tossed out. Then principal component analysis is applied to primary variables to eliminate linear correlation among them, thus reduce the variable dimension. Based on this method, the main components are gotten, and then simplified network structure is designed. The result shows that after data processing, the training time is reduced noticeably and forecasting precision is enhanced.