PiGT-2F: Physics-Informed Graph Transformer for Robust Short-Term Distribution Forecasting

Abstract

High-resolution node-level forecasts of net load and distributed photovoltaic (PV) injection are increasingly essential for modern distribution system operations and planning. Yet most feeders remain sparsely metered and subject to intermittent telemetry loss, degrading purely statistical forecasting pipelines. We present PiGT-2F, a physics-informed graph Transformer that (i) learns jointly across electrically coupled feeders, (ii) modulates spatial attention by branch admittance, and (iii) embeds soft penalties that encourage compliance with Kirchhoff’s Current Law and ampacity limits. Evaluated on a multi-feeder benchmark constructed from public feeder models, national load profiles, PV telemetry, and realistic random-plus-structured telemetry gaps, PiGT-2F lowers system-average nRMSE by 15–34% across 5- to 60-minute horizons and cuts physics-violation metrics by up to 80% relative to strong deep learning baselines. The archite­cture’s near-linear temporal attention cost enables multi-day history windows, and its physics regular-ization provides resilience when the grid is partially blind. Evaluation follows a feeder-level holdout to assess cross-topology generalization, and includes an anonymized real-telemetry micro-validation; code, configuration files, a benchmark generator, and a containerized reproduction package accompany the paper as described in the “Code and Data Availability” section.