In order to solve the problems such as variable redundancy and model complexity in ultra-short-term wind power prediction based on dynamic neural network (DNN), a novel method is proposed by combine the mean impact value (MIV) and principal component analysis (PCA) to optimize the predicting model constructed by DNN method. MIV method calculates the influencing degree from the input variables to the output and obtain the most important input variables to simplify the predicting model. However, its information utilization is low. PCA method extracts the principal components from the rest of the input variables. The information utilization can be greatly improved by adding a small number of principal components to make up for the deficiency of MIV method. It is verified by the data analysis and experiment that the optimized predicting model can assure the high contribution of the input variables and reduce the model complexity, which preserves the important information of the original system greatly, reduces the risk of introducing noise to the model, and makes the predicting result being improved significantly.