Abstract: For the purpose of promotion the precision of wind power forecasting, an ultra-short-term wind power forecasting model based on adaptive quadratic mode decomposition, convolutional neural networks and bidirectional long-short term memory network is proposed.In view of the fluctuation of wind power, using the improved fully adaptive noise integrated empirical mode decomposition method to decompose the wind power data. The sparrow search algorithm is introduced to optimize the decomposition number and penalty factor of variational mode decomposition, so that VMD has adaptability. The high-frequency component I1 obtained by decomposition of ICEEMDAN is decomposed secondarily by SSA-VMD to reduce the sequence instability. At the same time, the CNN network containing two pooling layers is constructed for feature extraction with the ultra-short-term prediction model of BiLSTM network, and the prediction results of each subsequence are superimposed to obtain the final wind power output prediction results. Experiments conducted through the analysis of arithmetic examples show that the prediction accuracy of the proposed wind power prediction method is better than other models, which verifies the superiority of the prediction model.