A Graph Neural Network-Based Approach for Power System Protection

This paper proposes a novel approach to power system protection using graph neural networks (GNNs), which can detect and locate faults on lines in power grids. For each relay, the GNN processes local measurements as well as measurements from adjacent lines to predict both the faulty line and the exact location of the fault. The GNN is trained on a diverse set of randomly generated, realistic grid structures, enabling generalization to previously unseen grids without retraining.The results of a case study demonstrate the effectiveness of the proposed concept. The GNN-based fault localization outperforms conventional distance protection and communication-based fault locators under noise. Additionally, the concept achieves full line coverage of the primary protection zone of each relay while being robust to missing measurements. Therefore, the proposed GNN-based protection system has the potential to improve the security and dependability of power system protection, particularly in future multivariate grids.