Accurate prediction of wind power is of great significance to promote large-scale integration of wind power for the grid. Most of the existing researches focus on single-step prediction in the ultra-short-term range. In order to achieve multi-step prediction of wind power which is closer to the engineering application, this paper proposes a multi-step prediction method of wind power based on ensemble empirical mode decomposition and encoder-decoder. First, the k-means cluster algorithm is employed to group wind turbines, and then the ensemble empirical mode decomposition algorithm is used to decompose the power sequence of each group of units. In this way, the temporal and spatial distribution characteristics of wind power are extracted. Advanced multi-step prediction of the wind power is achieved through the encoder-decoder prediction network based on the gated recurrent unit. Finally, by reconstructing the prediction values of all subsequences the total power in a wind farm will be obtained. Taking the data of a real wind farm as example, the simulation results show that the performance of the proposed algorithm is better than other traditional models in different application scenarios such as 1~6 hours ahead, and the prediction accuracy is improved by 6.45%~13.56%.