Predicting power and fatigue loads of yawed large offshore wind turbine based on cross-scale transfer learning

With the offshore wind turbine becoming larger, predicting its power and fatigue loads has become increasingly complex due to multi-physics coupling effects, especially considering yaw control. Although high-fidelity simulations can provide accurate results, they are extremely time-consuming, which hinders the wind turbine design and wind farm optimization. Meanwhile, significant data of medium wind turbines accumulated in industry has not been effectively utilized. Transfer learning provides a way to accelerate modeling large wind turbines by using those existing data. However, related studies on predicting wind turbine power and fatigue loads remain largely unexplored. Hereafter two wind turbine datasets (5 MW and 15 MW, each has two tandem-arrayed turbines) were generated using FAST. Farm and cross-scale transfer learning (CSTL) framework is proposed: a model is pre-trained on 5 MW data to learn general physical patterns, then fine-tuned with limited 15 MW data to adapt to large-turbine characteristics, effectively reducing target-domain data requirements. Results show that CSTL achieves equivalent accuracy with only half target-domain data compared to the best baseline, XGBoost, and with the same data volume, accuracy can be improved noticeably. SHAP analysis confirms that the model primarily relies on key features, i.e., wind speed, yaw angle, and turbulence intensity, which is consistent with wind turbine aerodynamics.