Short-Term Wind Power Prediction Based on Multi-Parameters Similarity Wind Process Matching and Weighed-Voting-Based Deep Learning Model Selection

Abstract

Sufficiently accurate short-term wind power prediction is important for the grid dispatch of the power system. To improve the accuracy by selecting suitable model for each piece of wind processes, this paper presents a short-term wind power prediction method based on multi-parameters similarity wind process matching and weighed-voting-based deep learning model selection. First, a novel multi-parameters similarity wind process matching method is presented to match each forecast target sample with groups of highly similar historical wind processes, in which each 96h-time-scale sample is divided into multiple wind processes by a tumbling time window, and a combinational similarity matching algorithm that consider four similarity indexes is proposed to judge the similarity among wind processes. Second, a weighed-voting-based deep learning model selection method, in which the matched highly similar historical wind processes are introduced to vote the optimal candidate deep learning model, is proposed to select the optimal model from LSTM, BLSTM, CNN, CNN-LSTM, CNN-BLSTM, and SDAE for each forecast target wind process. Case studies are presented to verify the effectiveness and superiority of the proposed method. Based on this new method, the 24h-day-ahead and 96h-short-term prediction RMSE can be decreased by 0.69% to 1.7% and 1.15% to 2.2% respectively compared to single deep learning model, which demonstrates the effectiveness of the proposed method.