Physics-Guided Deep Learning for Short-Term Probabilistic Offshore Wind Power Forecasting

To address the nonlinear amplification of numerical weather prediction (NWP) errors and the difficult trade-off between coverage and sharpness in short-term offshore wind power forecasting, this paper proposes PRWind, a physics-guided framework for short-term probabilistic forecasting. Built upon a Transformer encoder, the framework integrates horizon-adaptively weighted power-curve, persistence, and wind-speed cubic-law baselines; propagates wind-speed uncertainty into a power distribution through a probabilistic wind-speed distribution with Monte Carlo integration; and employs spatial graph convolution to model inter-turbine correlations, together with online bias correction and segmented post hoc interval calibration. The framework is validated on two offshore wind farms in coastal China (8350 kW and 10,000 kW units), with all data analyzed at hourly resolution and forecasts issued for 23 hourly steps ahead. For point forecasting, the NMAE reaches 8.74% and 8.24%, with MAE reductions of 19.1% (95% CI: 18.1–19.9%) at Wind Farm A and 7.8% (95% CI: 7.3–8.4%) at Wind Farm B relative to the strongest baselines, both statistically significant under the Diebold–Mariano test (p < 0.001). For probabilistic forecasting, the PICP remains stable at 0.89, balancing coverage and sharpness. The error of PRWind accumulates markedly more slowly with the forecast horizon than that of the baselines, and ablation studies confirm the synergistic effect of the modules. Trained and evaluated independently on the two wind farms, PRWind performs consistently, demonstrating cross-wind-farm stability (rather than zero-shot generalization) conferred by physical guidance, and offering significant engineering value for intra-day and day-ahead offshore wind power dispatch.