Improved spatio-temporal offshore wind forecasting with coastal upwelling information

Accurate short-term wind forecasts are critical for the reliable operation and integration of wind energy into the electric grid. For offshore wind farms, additional environmental uncertainties introduced by the oceanic environment can complicate the task of obtaining high-quality forecasts. A relevant example is the physical phenomenon of coastal upwelling which is a common oceanographic process wherein persistent along-coast winds drive the colder, deeper waters upwards, affecting the vertical wind profile and consequently, the power output of offshore wind turbines. This work introduces a spatio-temporal wind forecasting model which utilizes upwelling information derived from satellite imagery in order to improve short-term offshore wind speed and power predictions. Rooted in regime-switching modeling, the proposed approach learns regime-specific features of the offshore wind field, including relevant offshore weather effects and space–time correlations. Forecast evaluations using real-world data from the United States Mid-Atlantic—a region with frequent upwelling events and significant offshore wind energy activity—show that the intra-day and day-ahead forecasts from the upwelling-informed model are of significantly higher accuracy than those from baseline models that overlook such physically relevant information. Average forecast errors are reduced by 3.76% relative to state-of-the-art space–time methods, and by up to 27.53% when compared to classical time-series approaches. The results attest to the merit of formulating offshore-specific wind forecast models that are tailored to the unique physics of the offshore environment. • A regime-switching model for offshore wind speed and power forecasting is proposed. • The model integrates the effect of coastal upwelling on hub-height offshore winds. • Regimes are defined based on upwelling information derived from satellite imagery. • Regime-specific offshore weather effects and space–time correlations are modeled. • Real-world data from the U.S. East Coast confirm the merit of the proposed approach.