To overcome the defect of conventional BP neural network with low prediction accuracy for high-frequency component and weak generalization ability, this paper presents a hybrid technique combining wavelet transform and crisscross optimization (CSO) to optimize artificial neural network for short-term load forecasting. Wavelet transform is used to decompose the load series into different scales, after which, the neural network optimized by CSO is employed to forecast the load sub-sequences obtained by single reconstruction, and then, the values of all sub-sequences are added to get the actual forecasting results. A test for practical power system shows that the new model has stronger generalization ability and can grasp the change regulation of impact burr perfectly and improve the precision of forecasting with plenty of shock load effectively.