In order to improve the forecasting accuracy of ultra-short-term wind power, the improved wavelet-BP neural network method is applied. To solve the widespread delay problems of the prediction model, the original signal is decomposed into high and low frequency signal by the discrete wavelet transform. Moreover, genetic algorithm is used to optimize the BP neural network model separately. Finally, the summation of all the prediction results is gotten. As the wind speed and power series have chaos characteristics, the C-C method is used to optimize parameters of phase space reconstruction and the embedded dimension is taken as the input layer’s node number of neural network. It is applied in a wind farm, in Shandong Province, and the simulation results show that it has higher prediction accuracy than BP neural network model in forecasting wind speed and power. With the conversion of wind speed prediction results by the measured power curve, the power prediction accuracy goes bad.