Machine Learning Applications in Vertical Offshore Wind Profile Estimation

Abstract Accurate wind speed estimates are vital for the offshore wind farm industry. LiDAR measurements provide excellent wind data but are often limited in duration. Long-term wind time series are typically derived from either cost-effective but limited theoretical profiles or more accurate but costly atmospheric models. This paper aims to combine machine learning (ML) with theoretical profiles and observations to offer a cost-effective approach in generating wind speed time series at various heights. Two methods were used: the random forest regression (RFR), a supervised learning method, and the artificial neural network (ANN), a subset of deep learning which uses layered structures to model and capture patterns in the data. These models were developed using wind profiles, along with atmospheric and oceanic parameters, measured in different offshore lease areas along the U.S. coast. The modeled wind speeds were compared to estimates derived using the Power Law, an industry-standard wind profile parameterization. Results indicate that the industry-standard power law profile with a constant power (α) could not adequately capture the variability of the wind profile conditions. By allowing α to vary in time, the ML models produced more accurate wind speed profiles when compared with measured datasets. ML models trained based on the wind profile and meteorological measurements by LiDAR buoys can predict offshore wind profiles based on the parameters measured near-surface. Reliable offshore wind profiles estimations can strongly facilitate design of offshore wind farms by improving the estimations regarding mass flow through the turbine, performance at low and high wind speeds, extreme turbine operating conditions and the lifetime of the turbine.