Abstract: Most statistics prediction methods for short-term wind speed are based on univariate wind speed time series,they have limited prediction accuracy;however there is no specific method for selecting variables of multivariate prediction.This paper presented a multivariate local predictor for short-term wind speed prediction of wind farm.It sifted multivariate time series by correlation principle to reconstruct multivariate phase space,and searched the neighborhood of the prediction state points to build the support vector regression models.The data of real-world collected from a wind farm was applied to verify the conclusions.The example results show that the proposed method can improve the searching efficiency of local predictor that can find much more similar neighbor points.And combining with support vector regression（SVR） model that could provide good capability of nonlinear fitness,it can effectively improve the accuracy of short-term wind speed prediction.