Physics-Informed Graph Attention Network with Topology Masking for Probabilistic Load Forecasting in Active Distribution Networks

The integration of distributed photovoltaics (PV) introduces time-varying electrical coupling in active distribution networks, limiting the efficacy of conventional forecasting methods that rely on incomplete topological information and static physical models. This paper proposes a physics-informed spatio-temporal graph attention network (PI-STGAT) for probabilistic load forecasting under highly fluctuating conditions. A condition-adaptive correlation blending mechanism, derived from voltage–power sensitivity principles, fuses physical priors with statistical correlations using a PV-weighted strategy to capture time-varying electrical connectivity. An impedance-weighted continuous physical gating architecture maps voltage correlation coefficients into continuous attention biases, reflecting the spatial continuity of electrical distances while suppressing long-range noise. An uncertainty-aware adaptive physical constraint strategy dynamically modulates physical loss weights based on prediction variance and PV penetration, balancing fitting accuracy against physical consistency. Validation on real-world distribution network data demonstrates that, over a 24 h day-ahead horizon, PI-STGAT achieves a MAPE of 5.50%, a 3.7% relative reduction compared with LSTM. The model further attains a prediction interval coverage probability of 97.9%, confirming reliable uncertainty estimates under complex conditions.