Physics-Informed Spatiotemporal Deep Learning for Cable Fault Detection and Localization in Smart Power Grids
TL;DR
This study proposes a physics-informed spatiotemporal deep learning framework for fault detection and localization in medium-voltage cable systems that embeds telegrapher-equation residual constraints and distributed-generation boundary consistency within a physics-informed neural network to incorporate physical knowledge into the learning process.
Abstract
Reliable detection and localization of cable faults are essential for the secure operation of modern distribution networks, particularly with increasing distributed generation integration. This study proposes a physics-informed spatiotemporal deep learning framework for fault detection and localization in medium-voltage cable systems. The approach embeds telegrapher-equation residual constraints and distributed-generation boundary consistency within a physics-informed neural network to incorporate physical knowledge into the learning process. Long short-term memory networks capture transient temporal features of fault signals, while a topology-aware graph convolutional network identifies the faulted cable section within the network. A physics-regularized regression stage then estimates the precise fault distance inside the identified section. The framework is validated using electromagnetic transient simulations of a 10 kV underground cable network developed in PSCAD/EMTDC under various fault locations, grounding conditions, and distributed generation penetration levels.
