Large-scale synthetic dataset and machine learning benchmark for fault localization in distribution systems

Rapid and accurate fault localization is essential to ensure the reliable and uninterrupted operation of modern power distribution networks. While machine learning (ML) techniques have shown immense potential for efficient, data-driven fault localization, their real-world deployment is fundamentally hindered by a generalization gap. Constrained by scarce operational data, current ML models frequently overfit to narrow, homogeneous benchmark datasets and fail when applied to unseen network topologies. To bridge this gap, this study develops an OpenDSS-based simulation framework to generate a large-scale dataset of over 80,000 diverse fault scenarios, spanning multiple large-scale network topologies, varying operational conditions, and diverse fault types. By providing the critical scale and heterogeneity previously lacking in power system ML research, this framework enables the rigorous training and benchmarking of complex ML models. Through empirical validation, we demonstrate that a Gated Graph Neural Network (GGNN) trained on this expansive dataset mitigates the overfitting observed in prior work, demonstrates improved generalization relative to the baseline, and is able to narrow the candidate fault region under completely unseen network topologies. Furthermore, we conduct a systematic feature analysis to uncover how observable parameters – such as nodal current and voltage values – vary with fault location and type, providing key insights into feature importance.