Presentations

Abstract

Low-level jets (LLJs) are a recurrent atmospheric phenomenon in offshore environments and represent a significant, yet often under-recognised, challenge for wind turbine and wind farm operation. LLJs are characterised by a local maximum in wind speed within the lower atmosphere, introducing strong vertical shear, and temporal variability across the rotor. These conditions affect power production, structural loading, and controller behaviour [1–3]. As offshore turbines continue to increase in size and hub height, it becomes increasingly important for industry to understand, detect, and mitigate the impacts of LLJs. This contribution presents an integrated framework that combines long-term observations, aeroelastic simulations, and machine-learning-based detection to quantify and manage the effects of LLJs on wind turbines. The study is based on Doppler LiDAR measurements installed at multiple locations in the Belgian and Dutch parts of the North Sea. The application of combined relative and absolute wind speed fall-off criteria [4] identifies LLJs in approximately 11% of the observed wind profiles, demonstrating their operational relevance. The mean LLJ core height is found between 117 m and 129 m above sea level and is frequently located below the hub height of next-generation turbines, such as the IEA fifteen-megawatt reference turbine. This positioning results in negative vertical shear across the rotor, leading to inflow conditions that differ from conventional synoptic wind profiles. Turbine-level impacts are quantified by applying representative LLJ profiles, derived from LiDAR observations, as inflow conditions in OpenFAST simulations. The results show that LLJs systematically reduce harvested power at a given mean wind speed and significantly increase control variability. Aerodynamic forcing is both amplified and highly unsteady, propagating through the turbine control system. Pronounced fluctuations in generator torque and increased blade pitch activity are observed, indicating enhanced dynamic loading. These effects have implications not only for energy yield, but also for fatigue loading, component wear, and maintenance planning, particularly as turbines operate closer to their design limits. Building on these findings, a machine-learning-based LLJ detection framework is introduced to enable scalable deployment at wind farm level. A Temporal Convolutional Network is trained using LiDAR-based LLJ labels together with sparse, conventional meteorological and oceanographic measurements. Once trained, the model operates without LiDAR input, enabling LLJ detection from data streams typically available offshore. On an independent test set, the model achieves 92% correct predictions, with a Symmetric Extremal Dependence Index of 0.942. Saliency analysis reveals that wind direction and temporal context dominate predictions close to LLJ onset, while broader physical variables contribute to predictive skill up to eighty to ninety minutes in advance. Overall, the presented framework demonstrates how machine-learning-enabled LLJ detection can bridge the gap between atmospheric complexity and operational decision-making. By linking observed inflow phenomena to turbine response and providing early warning capabilities using standard measurements, the approach supports the development of LLJ-aware control strategies, and forecasting tools. This contributes to more resilient, efficient, and cost-effective offshore wind farm operation and is directly relevant for operators, OEMs, and developers seeking to reduce uncertainty and improve reliability at scale.