Abstract: With rapid growth of wind power integration into the modern power grid, Wind Power Prediction (WPP) plays an increasingly import role in the planning and operation of electric power system. However, the wind power time series always exhibits nonlinear and non-stationary characteristics, which is still with great challenge to be predicted accurately. To overcome the challenge, a Stacked Denoising Auto Encoders (SDAE) based deep learning approach based WPP is proposed in the paper. Firstly, SDAE with three hidden layers is designed to capture the nonlinear and complex characteristics from the reference data sequence, and the optimal initial connection weights of the deep neural network is obtained by the layer-wise pre-training process. Secondly, the back propagation algorithm is applied to fine-tune the weights of the whole network. To achieve the optimal network architecture, the grid search method is adopted to identify the number of neurons of the hidden layers and the learning rate of each denoising auto encoder. Finally, the proposed method is evaluated by the data from 52 wind farms and compared with the Back-propagation Neural Network (BPNN) and the Support Vector Machine (SVM). The results show that SDAE is with the ability to learn the nonlinear and non-stationary characteristics of wind power data and is with better accuracy than BPNN and SVM, which is applicable for practical applications of WPP in electric power system.