Forecasting of wind speed has a big influence on the rational dispatch and safety operation of electric power system. Aiming at the problem that the nonlinear characteristics of wind speed time series cause its poor prediction accuracy, a combined model based on complementary ensemble empirical mode decomposition and a support vector regression machine optimized by Gray Wolf Optimization (GWO) algorithm is used to predict ultra-short-term wind speed. First, the non-stationary wind speed time series is decomposed into a series of relatively stationary components by CEEMD. Then, each component is predicted by SVR optimized by GWO. Finally, the prediction values of each sequence are superimposed as the final prediction of wind speed. The results show that the prediction model is more accurate compared with other intelligent algorithm benchmark models and has superiority in wind speed prediction. This work is supported by National Natural Science Foundation of China (No. 41571016).