Robust Fault Location in Distribution Networks Under Noisy and Incomplete Measurements Using Physics-Aware Decoupled Inference

Fault location in distribution networks is often unreliable when measurements are noisy or incomplete. In actual feeders, synchronized data may be missing or distorted because of unstable edge communication. The proposed method uses physics-aware decoupled inference to locate line faults. The method works on single-time snapshots that capture voltages, currents, power flows, and zero-sequence components. These quantities are organized into an ordered hybrid tensor representing the feeder state at that instant. A one-dimensional convolutional encoder extracts spatial context from the tensor. Node measurements are handled separately and fused at the two terminals of each candidate line. This structure removes dependence on recursive graph message passing and confines the effect of local noise. The method is evaluated on the IEEE 33-bus test system under multiple noise levels, random masking of node features, and different fault resistances. With additive noise (σ=0.3) and 50% random node loss, the model achieves 92.8% localization accuracy. Average inference time per event is 0.62 ms on the tested GPU. The current implementation assumes a fixed feeder topology and synchronized aggregated measurements at the feeder level.