A wind speed forecasting method based on wavelet packet decomposition (WPD), improve bat algorithm (IBA), phase space reconstruction (PSR) and extreme learning machine (ELM) was developed. Aiming at the problems of poor initial population traversal and easily falling into local optimum in the iteration process of BA, Chaos initialization population and the updating mechanism of genetic algorithm (GA) is introduced into Ba and obtain IBA. The forecasting method firstly utilizes WPD to decompose the original wind speed into a series of sub-sequences with smaller amplitude and volatility. Then, each sub-sequence is reconstructed by PSR for ELM to make forecasting. The number of WPD decomposition layers, the reconstruction input matrix dimensions is optimized by IBA-based simultaneous optimization. Finally, the final forecasting results are yielded by accumulating the prediction results of each sub-sequence. The effectiveness of the new developed model is verified by the measured wind speed data from a wind farm in Inner Mongolia, China.