Multi-State Probability Prediction of Wind Power based on Deep Learning Model

Both value prediction and interval prediction of wind power ignore the probability of prediction results, To solve the problem, this paper proposes a multi-state probability prediction model for wind power based on a deep learning model which is composed by auto-encoder and softmax discriminant classifier. This model can extract hidden features of historical WP effectively, and output multi-state probability distribution of future WP. To train the proposed model, an effective training algorithm is presented based on greedy algorithm which can avoid local optimum. In addition, two indispensable parts of the proposed method, i.e. probability training samples and accuracy evaluation index, are given and the rolling prediction based on the proposed method explained in detail. Simulation results show that the prediction model is reasonable and effective.