Harish Baki

Estimating high-resolution profiles of wind speeds from a global reanalysis dataset using TabNet

24-Apr-24

The escalating demand for global wind power production, driven by the imperative need for sustainable energy sources, necessitates accurate estimation of vertical wind profiles for efficient wind turbine performance assessment. Traditional methods relying on empirical equations or similarity theory face limitations due to their applicability beyond the surface layer. Recent studies explore Machine Learning (ML) techniques to extrapolate wind speeds, but often focus on single levels, lacking a comprehensive approach to predict entire wind profiles. This study proposes a proof-of-concept in addressing the challenge, utilizing TabNet, an attention-based sequential deep learning model, to predict the entire wind profiles, provided by large-scale meteorological features from reanalysis. To make the methodology generic across datasets, the Chebyshev polynomials are used to approximate the wind profiles with Chebyshev coefficients. Trained on the meteorological features as inputs and the coefficients as targets, the TabNet better predicts unseen wind profiles for different wind conditions, such as high shear, low shear/well mixed, low level jet, and high wind, with a good accuracy. The methodology also addresses the correlation of wind profiles with associated atmospheric conditions by assessing the feature importance. The model demonstrates the feasibility of predicting wind profiles from large-scale meteorological variables, providing a valuable alternative to conventional methods.