Ultra-Short-Term Prediction of Wind Farm Cluster Power Based on Embedded Graph Structure Learning With Spatiotemporal Information Gain

Abstract

Ultra-short-term prediction of wind farm cluster power (UPWFCP) is of great significance for the development of intra-day power generation plan, and the power prediction accuracy is difficult to be further improved due to the chaotic effect of the weather system and the incompleteness of the information. In this regard, this paper proposes an embedded graph structure learning method for wind farm cluster (WFC) that incorporates spatiotemporal information gain (STIG) theory. The graph structure describing the spatiotemporal evolution relationship of information between wind farms (WFs) is constructed based on the spatiotemporal transfer relationship of power waveforms between WFs. An embedded graph attention network (EGAN) is proposed to learn STIG adjacency relationship among WFs, and a dynamic grouping scheme of redundant nodes in WFs based on STIG distance is constructed to reduce the modeling complexity. The proposed method is applied to the WFC of Inner Mongolia, China, and the results show that the NRMSE, NMAE, and MAPE of the proposed method are on average 2.63%, 2.09%, and 20.95% lower, and the R ² and Pr are on average 7.66% and 6.64% higher, respectively, compared with the rest of the comparison methods at all time scales.