Physics-Informed Feature-Based Convolutional Network for Fault Classification and Section Localization in Distribution Feeders

Abstract

This paper presents a compact physics-informed convolutional neural network (CNN) for post-fault classification and feeder-section localization in medium-voltage distribution networks using only three-phase voltages and currents measured at the primary substation. Instead of relying on raw waveforms, each post-fault record is represented by features extracted from a one-cycle window at 50 Hz and sampled at 10 kHz. Fundamental-frequency phasors are estimated through a single-cycle discrete Fourier transform, and three feature families are evaluated: phasor magnitudes, phasor magnitudes with phase-difference features, and an extended representation including symmetrical-component indicators. The proposed multi-task CNN jointly predicts the faulted feeder section among six candidate line sections and the fault category among four classes: line-to-ground (LG), line-to-line (LL), double line-to-ground (LLG), and three-phase (LLL). The method is validated on a simulated 20 kV two-feeder radial distribution network under an unseen fault resistance of 15Ω, which is excluded from training. Results show that phase information improves single-window localization, while the inclusion of symmetrical-component indicators further enhances performance, reaching 94.00% location accuracy and 98.65% fault-type accuracy. A record-level decision strategy based on the fusion of consecutive post-fault windows improves event-level robustness, achieving 96.44% Top-1 location accuracy, 100.00% fault-type accuracy, and 99.92% Top-2 location accuracy. Residual localization errors are mainly associated with electrically adjacent sections, which is consistent with the observability limits of single-ended measurements.