Topology-Aware Multi-Task Graph Learning for Fault Diagnosis in PV-Integrated Distribution Networks with Missing Measurements

Abstract

Fast and reliable fault diagnosis is a prerequisite for restoring service in active distribution networks, where distributed photovoltaic (PV) generation and sparse sensing complicate protection decisions. Existing learning-based methods degrade sharply when measurements are missing at test time: zero-filled inputs are indistinguishable from genuine deep voltage sags, and graph models built on the full feeder dilute informative readings across many unmeasured buses. This paper presents TMGN, a topology-aware multi-task graph network that performs fault-type identification and section localization with one model; it propagates measurements over a Kron-style reduced measurement graph weighted by electrical distance, and combines a learnable missing token with mask-randomized training. Because every graph node carries a measurement and missingness is encoded explicitly, the learned representation degrades gracefully as observability drops. On a simulated IEEE 33-bus feeder with three PV plants, ten fault types, and eight sections, TMGN matches the strongest single-task baselines at the nominal 20% missing rate (localization accuracy within 0.6 points) and retains $\mathbf{5 2 . 7 6 \%}$ localization accuracy at $\mathbf{6 0 \%}$ missing, 4.7 points above the best baseline, across five random seeds.