Direct measurement of structural loads on offshore wind turbines is often limited by cost and practical constraints, motivating the use of model-based approaches as an alternative to physical sensing. This study investigates the prediction of blade root flapwise and tower bottom fore-aft damage equivalent loads using SCADA data from the Lillgrund offshore wind farm, using a data-driven approach. The objectives are to compare the performance of different supervised regression models and to evaluate the impact of feature engineering strategies on predictive accuracy. The analysis is based on a three-year dataset. Three modelling approaches are assessed: eXtreme Gradient Boosting, Artificial Neural Networks, and Gaussian Process Regression. The results show that all models achieve high predictive accuracy, with XGBoost demonstrating slightly superior predictive performance and substantially higher computational efficiency. Feature engineering is found to play a critical role, with optimal configurations varying across turbines and load targets. In addition, a simplified fatigue damage estimation example is presented using the predicted DELs. Overall, the findings highlight the potential of data-driven models as virtual sensors to support fatigue monitoring, maintenance planning, and improved operational decision-making in offshore wind farms.
Link to paper: https://zenodo.org/records/22030432