Two-Stage Probabilistic Short-Term Wind Power Prediction Using Neural Network with MC Dropout and Control Information

Abstract:
This paper proposes a method for probabilistic short-term wind power prediction for a wind farm one day in advance. The proposed method uses the information from the control system, such as the cut-in and rated speeds of the turbines. The wind speeds are divided into three segments based on this information. The proposed method probabilistically predicts the segment each turbine belongs to at the first stage using a neural network (NN). Subsequently, the method generates samples representing the predicted power outputs for the three segments independently and integrates them according to the predicted probability of each segment at the second stage. An NN with Monte Carlo dropout is used for the generation of prediction samples in the segment from the cut-in to rated speeds, whereas constant values based on the control information of each turbine are used for the predicted power outputs in other segments. We present a case study in which the proposed method is applied to real wind farm and weather forecast data. The prediction accuracy of the proposed method is demonstrated in terms of the prediction interval as well as the expected prediction value.