It is very important to forecast short-term wind farm output for the security and stability of the power grid. Due to its non-steady and non-periodic characteristic， the wind power time series is decomposed into the random component and trend component by using the empirical mode decomposition（EMD） theory. Chaotic prediction is made for the random components and trend components using neural network with radical basis function and using least squares support vector machine， respectively， thus the final consequence can be obtained by combining the prediction result of each component. The power output of a wind farm in Yunnan is used as the case study for the model proposed. The outcome shows that the prediction model has high accuracy compared with the traditional artificial neural prediction model and provides a reference for the wind power forecasting.