A New Gridded Offshore Wind Profile Product for US Coasts Using Machine Learning and Satellite Observations

Abstract. Offshore wind farms are a low-cost, efficient technology for green energy. They deliver significant economic benefits through their manufacturing and operation, and can be readily deployed at scale. Offshore wind also offers a route to opening up access to renewable energy for a global population, 40 % of whom live within 100 km of the coast. Presently, offshore wind speed data around wind turbine hub heights are fairly limited, available either through in situ observations from wind masts and floating Light Detection and Ranging (lidar) buoys at selected locations or as forecasting-model based output such as from the 2023 National Renewable Energy Laboratory (NREL) National Offshore Wind (NOW-23) and the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5). In situ wind profiles are very sparse and costly to obtain en masse, whereas satellite-derived 10 m wind speeds have vast coverage at high resolution. In this study, we show the improvement of deploying machine learning techniques, in particular random forest regression (RFR), over conventional methods for accurately estimating offshore wind speed profiles on a high-resolution (0.25°) grid at 6-hourly resolution from 1987 to 2022 using satellite-derived surface wind speeds from the National Oceanic and Atmospheric Administration (NOAA) National Centers for Environmental Information’s (NCEI) Blended Seawinds version 2.0 (NBSv2.0) product. We use 276,577 wind profiles from five publicly available lidar datasets over the Northeast US and California coasts to train and validate a RFR model to extrapolate wind speed profiles up to 200 m. A single extrapolation model applicable to the coastal regions of the contiguous US and Hawai’i is developed, instead of site-specific ones attempted in previous studies.

Our RFR outperforms conventional extrapolation methods at the five training stations under cross validation (where each station is held out from the training once and used for validation), especially under conditions of high vertical wind shear and at wind turbine hub heights (~100 m). This model is then tested on two lidar stations that were not used in the training data and profiles from six NOW-23 station locations to evaluate its performance on unseen data. The final model is applied to the NBSv2.0 data from 1987–2022 to create publicly available wind speed profiles over the coastal regions of the contiguous US and Hawai’i on a 0.25° grid, which are shown to outperform NOW-23 and ERA-5 reanalysis at 100 m using a correlated triple collocation method over five years of matchup data (2015–2019). Gridded maps of wind profiles in the marine boundary layer over US coastal waters will enable the development of a suite of wind energy resources and will help stakeholders in their decision making related to wind-based renewable energy development.