It is important to study the fluctuation of wind speed in numerical weather prediction to improve the prediction accuracy of wind power. At first, a deep self-adaptive filtering framework is proposed. For numerical weather prediction of wind speed, the variational mode decomposition algorithm with kullback-leibler divergence is adopted. After multiple modal components are generated by decomposition, the noise components are filtered based on the non-local means algorithm. Then, the de-noised wind speed sequence is obtained by reconstruction with the effective component. On the basis, the example data is classified by season, and the wind speed sequence of the numerical weather prediction after denoising is taken as the input. In the alternative model base, the most suitable wind speed-power conversion model in this season is selected by the verification set, and the wind power prediction is carried out for the test set. A wind farm in northeast China is selected for example analysis, compared with other decomposition algorithms, the prediction accuracy of the proposed method in different seasons can be increased by 0.25% to 1.58%, that means, the depth adaptive filtering framework under seasonal typing can effectively improve the prediction accuracy of wind power.