Abstract: To improve the accuracy of power load forecasting, it is necessary to more effectively extract the hidden features and information in the massive data of power load. Based on the fact that the load has the characteristics of strong nonlinear, non-stationary and temporal, we proposed a short-term power load forecasting method based on the hybrid model of empirical mode decomposition（EMD）, convolutional neural network（CNN） and long-term and short-term memory network（LSTM）. In the process of forecasting, the massive past load data, temperature and the historical information of electricity price are constructed as a series feature vector and taken as input with time sliding window. First, we used EMD to reconstruct the data into multiple components, and superimposed and combined the high, medium and low frequency components. Then we used CNN to extract the hidden features of the high and medium components to reduce the weight. The components were used as input data for the LSTM network in the form of feature vectors for load prediction. Finally, the prediction results of each component were superimposed to obtain the final value of load prediction. The experimental results show that this model has higher load forecasting accuracy than BP neural network（Back Propagation Neural Network）, support vector machine（SVM）, long-term and short-term memory network and EMD-LSTM models.