The selection of input feature vector is the first important step in the establishment of wind power prediction model, but due to the excessive monitor items, the correlation between partial monitor items and wind power is not obvious or even irrelevant, and the redundancy information causes the selection of input vector set is not reasonable, so the accuracy of the power prediction model is affected. In order to solve this problem, three effective data mining algorithms for feature selection are studied synthetically, and a new method of selecting input vectors of wind power prediction model with better comprehensive performance is proposed, and the characteristics and application range of other methods are analyzed. Finally, the proposed method is validated based on the least squares support vector regression algorithm using the actual operation data of the turbine. The simulation results show that this method can effectively reduce the complexity of the model by reducing the input vectors, which not only speeds up the prediction speed, but also improves the learning ability of the model.