A residual graph attention network for smart grid stability prediction with incomplete data

Abstract

Accurate and reliable stability prediction is an important direction in smart grid, especially when data is incomplete. Traditional machine learning methods usually need complete feature inputs and rely on interpolation techniques, which will distort the relationships between variables. To address this problem, we proposed a graph-based framework that can directly receive incomplete samples without the need for data interpolation. In the proposed method, each sample is regarded as a graph, where the nodes correspond to the available features and the edges represent the relationships between them. The Pearson correlation coefficient is used as the weight of the edges to introduce prior statistical information. Besides, a residual graph attention network (GAT) is developed to learn complex feature interactions while reducing the oversmoothing existing in deep graph models. We carried out some experiments on an open smart grid stability dataset. The results show that the proposed method achieve better performance than other baseline models, including GAT, GraphSAGE, Graph Transformer, and traditional machine learning methods. This model achieves a recall rate of up to 0.9340, which demonstrates a strong ability to detect unstable grid states. Other analysis including sensitivity experiments and ablation experiments further confirmed the robustness of the proposed method in different missing rates and noise levels.