Abstract
Reliable fault localization is essential for fast restoration and resilient operation in distribution networks, yet the practical performance of learning-based fault localization is fundamentally constrained by sparse and heterogeneous measurement infrastructure. Under limited observability, different fault locations can yield highly similar current-magnitude patterns at available monitoring points, which makes magnitude-only classifiers ambiguous and expands confusion-prone sections, particularly in long unmonitored spans and laterals. This paper investigates fault-location predictability across multiple measurement infrastructure levels and proposes a phase-aware machine-learning framework tailored to realistic distribution conditions with missing phases and single-phase sections. The proposed algorithm constructs infrastructure-consistent feature sets using only observable quantities, aligns heterogeneous phase channels into a fixed input template, embeds current phase information through a trigonometric mapping to mitigate angle wrap-around discontinuities while preserving location-dependent phase signatures, and applies a training-time reweighting strategy to reduce performance degradation in data-scarce single-phase regions. Simulation studies on two distribution systems with different topological characteristics further show that confusion is spatially concentrated where monitoring boundaries are weak, and that phase-informed features provide additional separability precisely in those under-constrained regions where magnitude-based signatures alone are insufficient.
Funding Information
  • research grants from Daegu Catholic University, in 2026

This publication has 16 references indexed in Scilit: