Volume 64

Graph Neural Network–Based Security Assessment for Power Grids with Interpretable Feature Attribution Yuanhao Dai, Yunchao Sun, Wei Hu

Abstract

The power grid is the central hub of modern energy systems. Traditional security analysis methods struggle to cope with the rapid changes in operating states arising from high renewable energy penetration. Enhancing real-time security assessment capability is therefore of great importance. This paper proposes a graph neural network (GNN)–based framework for power grid security analysis, which fully exploits the inherent graph structure of power systems and incorporates operating state information to directly learn security-related patterns, enabling fast assessment of grid security. Furthermore, gradient-based attribution is applied to the trained GNN to help quantify and highlight critical buses, transmission lines, and power flow features that influence grid security. Case studies demonstrate that the proposed method can significantly reduce computational cost compared with conventional simulation-based analysis while accurately identifying key security characteristics under different operating conditions.

Keywords power system security analysis, graph neural networks, model interpretability, intelligent energy systems