WakeMeta: Physics‐Guided Graph Meta‐Adaptation for Cross‐Farm Probabilistic Wind Power Forecasting

ABSTRACT Short‐term wind power forecasting is essential for reliable grid operation, yet newly commissioned wind farms often provide only limited supervisory control and data acquisition (SCADA) histories for model adaptation. Existing cross‐farm methods face two related limitations: Data‐driven models must relearn site‐specific interturbine dependencies, while probabilistic forecasts can become poorly calibrated under layout and operating‐regime shifts. This study proposes WakeMeta, a physics‐guided graph meta‐adaptation framework. WakeMeta converts an established Bastankhah–Porté–Agel Gaussian wake relation into a directed, training‐independent graph operator determined by turbine geometry, wind direction, and fixed engineering wake parameters. The graph constrains a spatiotemporal encoder with dilated temporal convolutions and graph attention, while a probabilistic head combines Gaussian likelihood, continuous ranked probability score (CRPS), coverage calibration, and head‐level Monte Carlo (MC) dropout. A first‐order meta‐learning procedure is applied only to data‐driven parameters, and target adaptation uses validation‐gated early stopping in which the unadapted initialization is a legal candidate. Under the frozen evaluation protocol (V3‐final) with five paired nonoverlapping target windows, 4‐h forecasts, and 14 days of target data, WakeMeta achieves CRPS and calibration error (CE) values of 0.1117/0.0386 on Penmanshiel and 0.0640/0.0742 on La Haute Borne. The corresponding CRPS is 5.0% lower than seasonal autoregression on Penmanshiel and 14.1% lower than persistence on La Haute Borne, while direct source pretraining with fine‐tuning remains strongest on La Haute Borne. Robustness experiments show that backbone, predictive family, graph, and adaptation effects are target dependent, supporting a deliberately scoped interpretation centered on portable physical structure and calibrated probabilistic transfer.