A Data-Mechanism Fusion Method for Distribution Network Fault Diagnosis Based on Digital Twin

Abstract

: Accurate fault diagnosis in distribution networks is essential for ensuring reliability and reducing outage time, particularly as power systems undergo digital transformation. Existing digital twin-based methods struggle to deeply integrate data-driven adaptability with physics-based interpretability, which limits their robustness and accuracy in complex scenarios. To address this challenge, this paper proposes a data-mechanism fusion fault diagnosis method within a digital twin framework. The data model employs a super-resolution graph convolutional network to reconstruct full-network features from sparse μPMU measurements, followed by a graph attention network for fault classification and location. The mechanism model applies wavelet modulus maxima-based single-ended traveling wave ranging to generate a set of physically constrained candidate fault nodes. An adaptive soft-voting fusion strategy is then used to combine the probabilistic outputs of both models, enhancing diagnostic accuracy and reliability. Tests on the IEEE 37-node system show that the proposed method achieves 99.2% fault classification accuracy and 97.1% zero-hop location accuracy, outperforming standalone models. Comprehensive robustness tests under varying fault resistances, noise levels, topology changes, and distributed generation integration confirm strong resilience to real-world operational uncertainties. This work provides a practical digital-twin-enabled solution for distribution network fault diagnosis that balances diagnostic precision with economic