Accurate fault location in distribution networks is essential for grid reliability and rapid service restoration. As distributed energy resources—solar, wind, and battery systems—proliferate in low- and medium-voltage grids, the task becomes increasingly complex. High DER penetration introduces intermittent generation, bidirectional power flows, and altered fault current signatures that degrade traditional protection schemes and machine learning models trained on historical data.
This research evaluates whether spatio-temporal graph attention networks (STGATv2) can reliably locate faults across varying DER penetration levels. Using a modified IEEE 123-bus distribution feeder with multiple DER injection points, researchers systematically compared STGATv2 against purely temporal models (GRU), purely spatial models (GATv2), and conventional machine learning baselines.
Results demonstrated clear advantages for the spatio-temporal approach. STGATv2 achieved 92–94% macro F1-score under in-distribution conditions and exhibited asymmetric generalization behavior: models trained at 50% DER penetration maintained near-baseline performance when tested at 10% penetration, but training at 10% showed significant degradation at 50% penetration. Critically, STGATv2 retained 81–84% F1-score under this drastic scenario shift—substantially outperforming GATv2 (69–74%) and GRU (73–75%).
When realistic measurement noise was introduced, STGATv2 maintained over 85% accuracy, whereas GRU performance collapsed to as low as 33.5%, underscoring the importance of topological awareness in dynamic grid conditions.
The findings highlight that jointly modeling spatial network topology and temporal voltage and current dynamics enables robust fault location despite high DER penetration and measurement uncertainty. This topological awareness allows the model to exploit structural grid information that pure temporal approaches cannot access. As distribution networks continue adding renewable resources and microgrids, deployment of such physics-informed learning methods may become essential for maintaining protection coordination and rapid fault isolation in active distribution systems.
