For the low prediction accuracy caused by the inherent vibration and nonlinear characteristics of wind speed signal, a combined model of the ensemble empirical mode decomposition (EEMD) and the gated recurrent unit (GRU) is proposed to predict the wind speed. Firstly, the model normalizes the data and removes outliers by isolated forest. The wind speed is then resolved into signals of different scales by EEMD to obtain stable component signals with the non-stationary data removed. The component signals are trained by the GRU model, from which the predictions are accumulated to obtain the final wind speed. The data collected in the field are applied for the experiment. The results show that the EEMD-GRU method has a significant improvement in the prediction accuracy compared with the dominated EEMD-LSTM and EMD-LSTM methods.