Hidden faults in substation grounding grids pose a serious threat to the safe and reliable operation of power systems. Traditional fault diagnosis methods ignore the correlation between grounding grid topology structures, resulting in low fault location accuracy. To address this issue, this study proposes a high-precision fault diagnosis and localization method that integrates graph neural networks (GNNs) with multi-source information. Innovatively abstracting the physical structure of the substation grounding grid into a topological graph model, with connection points and grounding electrodes as nodes, and conductor segments and their properties as edges, accurately representing the topological relationship of the grounding grid. A multimodal graph neural network (MM-GNN) was designed based on this, which captures topological features through graph convolution operations guided by grounding grids and adaptively integrates multi-source monitoring data such as electrical, infrared thermal imaging, and electromagnetic fields. In addition, this article innovatively developed an end-to-end conductor level positioning mechanism, incorporating physical connection rules into model training using topological constraint loss functions to ensure accurate output of fault locations. The experimental results based on a 220kV substation show that the proposed method significantly improves the diagnostic accuracy and provides an effective solution for the diagnosis and localization of grounding grid faults.
Pu et al. (Wed,) studied this question.