Wind Power Low-output Event Identification and Accurate Prediction Strategy Based on Global Heterogeneous Information Dynamic Fusion with Dynamic Dual Graph Neural Networks

The frequent occurrence of wind power low-output events (LOE) poses a serious challenge to the dispatching safety and operation stability of power systems. The existing research lacks a systematic identification and prediction scheme for such events. So, this paper proposes a dynamic dual graph networks to achieve accurate capture of LOE. In the identification graph, used a dynamic graph convolutional network (DGCN) that integrates the time-domain encoding convolutional attention mechanism (TDECAM) to integrate the global meteorological and power information of the wind farm cluster (WFC) to accurately identify the LOE of the target wind farm. By constructing the conditional weighted distance (CWD), the multi-dimensional feature dynamic correlation between wind farms is depicted. In the prediction graph, established a dynamic spatiotemporal heterogeneous graph attention network (STHGAT) that integrates the delay effect of wind speed (DEWS) to characterize the spatiotemporal heterogeneous connections between wind farms of different output types, so as to achieve accurate prediction of power evolution process under LOE. The proposed method was applied to a WFC in Inner Mongolia, China for validation. Compared to traditional models, the accuracy, precision, recall, and F1 score for LOE identification were on average 10.34%, 10.29%, 10.34%, and 10.35% higher, respectively, with identification accuracy exceeding 96%. In power prediction for low-output scenarios, compared to traditional models, the IRMSE and IMAE were reduced by an average of only 4.16% and 3.45%, respectively, while R2 increased by an average of 24.56%. At the wind farm level, overall prediction accuracy improved to over 95%.