An ultra-short-term wind power prediction method based on EMD-SA-SVR is proposed for the high randomness and strong fluctuation. Empirical Mode Decomposition (EMD) is used to extract different features lying in wind power sequence. By decomposing the original sequence into several more regular modes, the accurate prediction model can be built easily for each mode to eliminate the coupling between different features in the original sequence. The prediction models are constructed with Support Vector Regression (SVR) for its fine generalization performance. The Simulated Annealing (SA) algorithm is used to optimize the SVR parameters to solve the multi-extreme complex nonlinear problem of model selection, and the optimized model can be obtained for each mode. The wind power prediction value can be gained by summarizing the results of every single mode. The comparative analyses on a wind farm history data show that the EMD-SA-SVR model can effectively improve the accuracy of ultra-short-term wind power prediction. This work is supported by National Natural Science Foundation of China (No. U1504617).