Abstract: To better extract and study the characteristics of wind speed in the time and frequency domains, and to solve the time-domain randomness and frequency-domain complexity problems of the wind speed signal, we propose a combined short-term prediction model, WD-VMD-DLSTM-AT, which is based on wavelet decomposition and reconstruction (WD), variational mode decomposition (VMD), a long-short-term memory (LSTM) network and an attention mechanism (AT). On this basis, we propose a multi-input multiple output (MIMO) codec multi-step prediction model (MMED-AT) based on an attention mechanism. A comparison and analysis of the experiment results proves that the proposed combined forecasting model has the smallest statistical error, and can significantly improve the prediction accuracy in the short-term wind speed prediction. MMED-AT models based on the proposed combined model can obviously eliminate the cumulative error of recursive multi-step prediction and improve the stability of multi-step prediction.