Abstract:  Accurate wind power prediction is crucial for the efficient and safe operation of the power system. To improve the accuracy of wind power prediction, a short-term mixture model for wind power prediction was proposed based on the combination of improved complete ensemble empirical modal decomposition with adaptive white noise (ICEEMDAN), permutation entropy (PE), improved chimp optimization algorithm (ICHOA), least squares support vector regression (LSSVR) and bi-directional long short memory (BiLSTM) network. Firstly, the non-stationary original wind power sequence is decomposed into relatively stationary modal components through ICEEMDAN, and PE aggregation is used to reduce computational complexity. Secondly, the BiLSTM model and LSSVR model are applied to predict high-frequency and low-frequency components, respectively. ICHOA is used to optimize the parameters of the model. Finally, the final prediction result is obtained by overlaying the values of each predicted component. Through the analysis of specific examples, the proposed LSSVR-BiLSTM dual scale deep learning model is compared with other models, which can better fit the wind power data and has higher prediction accuracy and feasibility.