In order to improve the wind load forecasting precision and ensure the effective use of wind power resources, a method based on improved least square support vector machine (LSSVM) combined with error forecasting is proposed. Firstly, lifting wavelet transform (LWT) decomposition of the original data is introduced to effectively extract the main features, with which the randomness of the wind is overcome; secondly LSSVM is used for the prediction of decomposed signals to ensure accuracy; then, error forecasting (EF) is added to reduce the large error points and improve the stability of the results. Finally, experimental results using real wind farm data show that the forecasting model is better in both generalization performance and predictive accuracy, and may provide an effective and practical way for the short-term wind load forecasting.