Accurately predicting wind speed is of key importance for large scale wind power connecting to the grid. To improve the short-term wind speed forecasting accuracy, a least squares support vector machine wind speed prediction model based on wavelet decomposition and improved firefly algorithm is proposed. Firstly, the actual wind speed series is decomposed and reconstructed to approximate series and detail series, then the series are separately predicted by LSSVM optimized by chaotic firefly algorithm, at last the separate prediction series are superposed as the ultimate prediction wind speed. To verify the proposed model, two different time scale actual wind speed data are applied to simulation. The results show that the proposed model has higher prediction accuracy than classical model like CV-LSSVM, IPSO-LSSVM, WD-DE-LSSVM and BP neural networks, showing its validity and superiority.