Feeder-Level Ground Fault Localization in GFN-Compensated Distribution Networks: Harmonic Feature Engineering and Event-Level Evaluation on Utility Field Data

Abstract

Identifying which distribution feeder carries a ground fault is a hard problem in networks compensated by a Ground Fault Neutralizer (GFN): the compensation that limits fault energy also suppresses the zero-sequence signatures that conventional protection relies on, so a fault on one feeder can de-energize an entire substation. High impedance faults (HIFs) are harder still; their small, distorted currents can escape detection while posing electrocution and wildfire risks. Existing data-driven localization methods are validated almost exclusively on simulation. To the best of our knowledge, this paper presents the first feeder-level ground fault localization framework in GFN-compensated distribution networks developed and validated entirely on utility field data, using 183 staged fault tests recorded by Southern California Edison on two live 12 kV, three-feeder substations under varying load. The task is formulated as multi-class classification over per-cycle phasor features of the zero-sequence voltage and feeder currents: fundamental and higher-order harmonic magnitudes and phases, bus–feeder phase differences, per-feeder admittances, total harmonic distortion (THD), and discrete wavelet transform (DWT) coefficients, organized into six progressively richer feature sets. Each configuration is trained separately for low impedance faults (LIFs) and HIFs and evaluated under an event-held-out protocol with multi-seed averaging and event-level majority voting, which scores each fault event as a unit so that correlated frames from one event never span the train and test split and inflate accuracy; on the strongest feature set for each task, recurrent, feedforward, and tree-ensemble classifiers of matched capacity are compared. Low-order harmonic features localize LIFs at $99.8 \pm 0.1\%$ frame-level accuracy with every held-out event correct. For HIFs, frame-level accuracy plateaus near 90% when 20–40 harmonics are retained, but event-level reliability does not follow frame accuracy: a compact 93-feature harmonic-plus-wavelet configuration localizes all 79 HIF events in event-level cross-validation, because its frame errors scatter within events instead of overturning majority votes. Classifier choice is secondary: feedforward, recurrent, and tree-ensemble models of matched capacity perform comparably, so the contributions are the per-fault-type harmonic and relational feature engineering and the event-level evaluation protocol. Feeding multi-frame sequences to the recurrent models does not improve on per-frame classification, so a per-frame feeder label is available one cycle after the analysis window fills. Trained on one substation and tested on another, localization transfers for LIFs but not for HIFs, which remain substation-specific and require local training data. The framework offers a data-driven alternative to fixed-threshold protection for compensated distribution networks.