Estimating High-Resolution Wind Speed Profiles from a Reanalysis Dataset using TabNet: A Transfer Learning Approach

Technological advancements have enabled the deployment of massively tall wind turbines in offshore regions. For example, the MySE 16-260 has a 16 MW capacity, a rotor diameter of 260 meters, and a hub height of 152 meters, reaching a total height of over 280 meters. The so-called Rotor Equivalent Wind Speed (REWS) based power curve approach offers more accurate power production estimates than conventional hub-height-based power curves for these tall turbines. However, applying the REWS approach requires vertical wind speed profiles spanning several hundred meters within the offshore boundary layer. It is well-known that specific meteorological conditions result in distinct wind speed profiles, such as low-level jets (LLJs) during stably stratified conditions, well-mixed profiles during unstable conditions, and Ekman profiles during neutral conditions. Wind shear and turbulence intensity associated with these diverse wind profiles significantly influence power production. Beyond power output, wind profiles under different stability conditions also significantly impact turbine loads. These facts underscore the importance of accurately estimating wind profiles for wind resource assessments and turbine designs.

However, characterizing wind speed profiles across the rotor-swept area has been hindered by the scarcity of high-altitude observations, as deploying wind measurement tools such as offshore met masts and floating lidars is generally cost-prohibitive. As a poor man's alternative, numerous empirical equations and formulations based on similarity theory exist for wind speed extrapolation, such as the power law and the Monin-Obukhov similarity theory. Yet, some of these equations are only valid within the surface layer, while the swept areas of contemporary turbines extend well beyond this layer, complicating their application. Additionally, some formulations often require extra information (e.g., turbulent flux, boundary layer height) that near-surface stations (e.g., buoys) typically do not measure. Currently, the most affordable and practical approach for estimating offshore wind speed profiles is to use mesoscale models. In recent years, several mesoscale model-based datasets have been generated, such as the Copernicus Regional Reanalysis for Europe (CERRA), New European Wind Atlas (NEWA), Dutch Offshore Wind Atlas (DOWA), Winds of the North Sea in 2050 (WINS50), and Solar and Wind at Gray-Zone resolution (SWaGZ). However, mesoscale model simulations require significant computational resources and are sensitive to initial and boundary conditions as well as physical parameterizations.

Recently, we proposed a proof-of-concept deep learning framework, henceforth WindProfileNet, to address these challenges. We selected TabNet, a state-of-the-art attention-based sequential deep learning model, for our approach. The model was trained using coarse-resolution meteorological features extracted from the ERA5 reanalysis as inputs, with high-resolution wind speed profiles from the CERRA dataset as targets. We approximated wind speed profiles using Chebyshev polynomials to ensure our method could be generalized across different datasets with varying vertical resolutions. Emulating the traditional measure-correlate-predict (MCP) approach, we trained ten ensembles of TabNet models on one year's data and predicted for a different year. The TabNet models accurately predicted wind speed profiles across various wind regimes, including high shear, well-mixed, low-level jet, and strong winds.

In the present study, we extend the WindProfileNet framework to lidar-based wind speed profiles. However, the observations are available for limited periods (only one year in our case), making it challenging for WindProfileNet to predict infrequent wind regimes, such as low-level jets, accurately. To address this challenge, we developed a transfer-learning approach.

In the first phase, we trained TabNet models using the Chebyshev coefficients that approximate vertical wind speed profiles from the CERRA reanalysis as targets over 17 years. Although the simulated data were not perfect, this initial training enabled TabNet models to capture diverse wind patterns. In the second phase, we retrained the TabNet models using the lidar data-based Chebyshev coefficients. This process allowed for the transfer of information across all wind patterns through learned weights. We emphasize that even though the vertical spacing of the lidar data and the CERRA data were different, the Chebyshev polynomial-based approximation allowed us to combine them seamlessly.