Abstract: In order to improve the accuracy of ultra-short-term wind power multi-step prediction and solve the problem that the gradient correction learning algorithm can easily fall into local optimum when the adaptive wavelet neural network（AWNN）trained by random initialization parameters. An improved model based on the adaptive wavelet neural network（AWNN）and improved PSO-DE algorithm is put forward in this article. The IPSO-DE algorithm is the combination of the improved particle swarm optimization（IPSO）algorithm and the differential evolution（DE）algorithm. It is used to optimize the random initialization parameters of the AWNN. The improved model, which is the IAWNN, is applied to the ultra-short-term wind power multi-step forecasting. The simulation results show the IPSODE algorithm is better than gradient correction learning algorithm, IPSO algorithm and DE algorithm in optimizing the random initialization parameters of the AWNN. The proposed model estimated performance is obviously superior to AWNN model,persistence method（PM）model and BP neural network（BPNN）model.