Abstract: This paper proposes a short-term wind power combination forecasting method based on two-mode decomposition,two-channel convolutional neural network（CNN）and long short-term memory neural network（LSTM）to improve the forecasting accuracy. Firstly,the wind speed time series clustered by the PAM method is decomposed into two modes of multi-scale subsequences by the singular spectral decomposition（SSD）and variational mode decomposition（VMD）. The multi-scale subsequences of different modes can reduce the complexity and nonstationay of the original wind speed,and achieve the complementarity of the two modes.Secondly,the wind speed subsequences obtained by the two decomposition methods are combined into a matrix,which is input to the dual-channel CNN for waveform feature depth extraction. Finally,LSTM network is used to establish the historical wind speed time series dependency relationship,and the final wind speed prediction result is obtained on the basis of spatiotemporal correlation analysis. Experiment results show that the combined prediction model based on dual mode decomposition-dual channel CNN-LSTM can effectively improve the accuracy of short-term wind speed forecasting.