Abstract: This paper proposes a combination forecasting method based on the multi-mode decomposition. the sparrow search algorithm(SSA), the residual neural network (ResNet) and the gated recurrent units (GRU) network to improve the accuracy of wind speed prediction. Firstly, the wind speed time series clustered by the fuzzy C-mean method is decomposed into three modes of multi-scale subsequences by the wavelet transform, the variational mode decomposition and the complete ensemble empirical mode decomposition with adaptive noise. The wind speed subsequences obtained by the three decomposition methods are combined into a matrix, which is input to the convolutional network. Then the multi-scale subsequences of the three different modes achieve the complementarity of the fluctuation patterns. Subsequently, the improved residual module is added into the traditional convolution network for the feature extraction of the multi-modal decomposition components, which enhances the deep features significantly. Finally, in order to further extract the features of wind speed components in time series, the features are combined and input into the GRU module. Besides, the key parameters are optimized in the Res-GRU with the SSA. By applying these steps, the wind speed prediction is achieved. Experimental shows that the combined prediction model proposed in this paper effectively improves the accuracy of wind speed prediction compared with the traditional model.