Abstract: Wind power prediction (WPP) technology is a key factor in power system scheduling and safe operation. In order to better improve the accuracy of the WPP technology, this paper proposes a multi-integrated cluster short-term WPP method based on ensemble learning, which includes four steps. The first step is to decompose the original wind power sequence into multiple sub-sequences by using variational mode decomposition, empirical mode decomposition and wavelet transform. The second step is to construct multiple stacked denoising autoencoders (SDAEs) based on subsequences for deep learning. The third step is to randomly divide the results of the second step into several sets, and integrate each set using the support vector machine (SVM). The fourth step is to randomly divide the integration results of the third step into several sets, integrate each set with SVM, and repeat the above steps until the final integration prediction result is obtained. The results show that the average normalized root mean square error (RMSE) of the first 96 h prediction results obtained by multiple integration learning is reduced by 0.010 1 (9.01%) compared with that obtained by single integration learning. Compared with SDAE, it decreases by 0.015 1, and the percentage is 13.54%. Compared with SVM, it decreases by 0.017 5, a percentage of 14.66%. This paper can provide reference for short term power prediction of wind power cluster based on deep learning and ensemble learning.