Aiming at the problems caused by the current indirect methods for ultra-short-term wind speed prediction which used the same model for each frequency sequence, this paper proposed a combined ultra-short-term wind speed prediction method using a hybrid model based on wavelet decomposition. The proposed method is based on three basic methods, which are autoregressive differential moving average model, back propagation (BP) neural network and support vector machine. The appropriate method is selected and the corresponding model is created according to the characteristics of each frequency sequence after wavelet decomposition, then the results of ultra-short-term wind speed prediction can be obtained through wavelet reconstruction. The proposed method can consider the difference and predictability among each frequency sequence fundamentally, therefore the prediction accuracy can be improved. The proposed method has higher prediction accuracy under different time horizons. When mean absolute error is used as the evaluation index, the prediction accuracy of the proposed method can be improved by 64.2% and 61.4% under 1 h and 4 h time horizons respectively compared with the persistence method, 7.2% and 5.7% under 1 h and 4 h time horizons respectively compared with the minimum prediction error method in traditional single model combination prediction method.