Because the actual wind speed and wind power sequences are fluctuating, intermittent and the hidden node number of RBF neural network is unchangeable after the structure of RBF neural network is confirmed, a method of ultra-short-term wind power prediction based on wavelet and minimum resource allocation network is proposed. Firstly the historical wind speed and wind power sequences are denoised and multi frequency decomposed by wavelet transform, several high frequency signals and a low frequency signals are obtained. Then neural network prediction models of different frequency signals are built respectively to predict the wind power in the next 4 hours. Finally, the final ultra-short-term wind power prediction result is obtained from wavelet reconstruction of different components. The experimental results show that this method can effectively improve the prediction accuracy. This work is supported by Sino-Dan International S & T Cooperation Program (No. 2014DFG62530), National Natural Science Foundation of China (No. 51507053), and Fundamental Research Funds for the Central Universities (No. 2017B42314).