A Physics-Guided Graph Transformer for Short-Term SCADA-Based Multi-Turbine Wind Power Forecasting

Short-term wind power forecasting depends on reliable Supervisory Control and Data Acquisition (SCADA) measurements and on models that can represent interactions across a turbine fleet. PG-RAGT is proposed for this setting as a physics-guided graph Transformer for 10–60 min multi-turbine power forecasting. The contribution is organized at three levels: application-oriented system integration, state-conditioned fleet interaction, and forecast-time physical-state closure. Physics-Consistent SCADA Screening acts as a nominal-regime training filter for active-power observations that depart strongly from a simplified aerodynamic reference. Axial Fleet Blocks form sample- and patch-dependent cross-turbine interaction matrices from the current SCADA representation without a predefined adjacency. Gated Physical-State Rolling generates horizon-dependent wind speed, external temperature, and mean pitch angle from the last observed state, allowing a state-dependent aerodynamic power prior to be evaluated without future SCADA measurements. On the SDWPF test set, the main seed gives RMSE = 106.69 kW, R2 = 0.9125, and MAE = 57.12 kW; across three independently seeded runs, the mean RMSE is 107.19 kW. The cross-turbine pathway shows the most consistent ablation effect, whereas the screening and rolling effects are smaller and more training-run dependent. The resulting framework provides comparatively accurate fleet-scale forecasts while retaining probabilistic outputs and a reduced-model option for lower computational cost.