Wind profile nowcasting and forecasting using machine learning

Accurate knowledge of the vertical wind profile is essential for the energy industry and structural engineering. Specifically, it is crucial for enhancing wind turbine efficiency and enabling active systems through digital twins, which require precise prediction methods to optimize and forecast system behaviors effectively. This study introduces a robust framework that employs machine learning (ML) techniques to nowcast and forecast non-dimensional wind profiles. The problem is formulated as a regression task where the parameters of non-dimensional polynomial wind profiles are predicted based on historical data. Nowcasting is specifically aimed at estimating the shape of the wind profile at the current moment using only meteorological weather information at ground level from local weather stations, utilizing XGBoost for real-time predictions. Forecasting, in contrast, employs both current and historical wind profile data obtained from anemometric towers or wind lidar systems to predict future wind profiles, using Long Short-Term Memory (LSTM) networks. We apply these methodologies using data from the Cabauw experimental site for the period February 2018 to June 2020. The calibration of the models, including hyperparameter tuning and the determination of the optimal time window for past data utilization, is discussed in detail. This research demonstrates the potential of ML in providing reliable predictions of the wind profile height distribution, which can significantly benefit wind energy optimization and structural engineering applications. • Non-dimensional polynomial wind profiles are defined and parameterized. • XGBoost (nowcasting) & LSTM (forecasting) predict wind profile parameters. • Nowcasting uses local ground weather data for real-time profile estimation. • Forecasting predicts future profiles using historical ones & local weather data. • Models are trained and validated with a comprehensive dataset from Cabauw.