Accurate wind power prediction can improve the safety and reliability of power grid operation. In this study, a novel deep learning network stacked by independent recurrent autoencoder (IRAE) is designed according to the characteristics of ultra-short-term wind power data, hereafter called SIRAE (staked independently recurrent autoencoder). This approach accommodates a sheer volume of data in the smart energy era and overcomes the effects of random changes in the natural environment. First, the original sequence is decomposed into sub-sequences through variational mode decomposition techniques. Second, each IRAE extracts the structural features and deep connections of the data through unsupervised pre-training and finds an appropriate initial state. Finally, all IRAEs are stacked into SIRAE and the parameters of each layer for SIRAE are adjusted through supervised training. The results of two comparative experiments show that SIRAE model outperforms the existing popular models. In the extended application, compared with the persistence model, SIRAE shows 18.46%, 31.16%, 9.06% and 34.24% improvements in terms of root mean square error for March, June, September, and December. Therefore, SIRAE is an appropriate tool for ultra-short-term wind power prediction on account of its effective and stable forecasting performance. (c) 2020 Elsevier Ltd. All rights reserved.