A new approach to ultra-short-term wind speed forecasting is presented. It is based on the phase space reconstruction technique and the local prediction method， to search the neighbors in the optimal neighborhood in the phase space and to build support vector regression（SVR） models. This approach finds the optimum neighborhood by considering of proportion of false neighbors， which guarantees the high similarity between neighbors and prediction state points， and the SVR model has a good capability of nonlinear fitness. The results have demonstrated the accuracy of the proposed method in comparison with that by other prediction methods.