Abstract: The depth model achieves complex function approximation by learning a deep nonlinear network structure,which has strong adaptive perception ability. In order to improve the prediction accuracy of power load, a deep learning prediction method based on stacked auto-encoder neural network is proposed in the paper. A multi-input single-output prediction model is built by combing the auto-encoder with the logic regression classifier, such data as the reconstructed historical load, meteorological elements and so on are all input into prediction model, and the load characteristics is extracted through the hierarchical learning of the stacked auto-encoder. Finally, the short-term load prediction is realized by using the logical regression model at the top of the network. Case analysis shows that the proposed model can effectively characterize the daily load change law with strong generalization performance, and its prediction accuracy can reach 96.2%, which is higher than that of two shallow learning models based on support vector regression and fuzzy neural network respectively.