Advancing Wake Loss Prediction in Wind Farms through Analytical Machine Learning Hybridization

Wake models are still important for wind farm evaluation because they are computationally efficient when predicting turbine interactions and performance of the wind farm layout. Although the Jensen model is the simplest of the wake models and is often used, its assumption of linear wake expansion limits its accuracy when modelling wake interactions in complex terrain. To overcome the limitation of the Jensen model, this work introduces a hybrid framework that combines the analytical Jensen model with Extreme Gradient Boosting (XGBoost). The proposed framework utilizes wake predictions based on physics from the Jensen model combined with operational SCADA data, with XGBoost serving as a residual correction layer that learns the errors of the Jensen model and corrects them to improve downstream wind speed predictions. The results of this framework indicate considerable improvements from the Jensen model the R² increased from 0.8228 to 0.9237, RMSE decreased from 1.4011 m/s to 0.9201 m/s, MAE decreased from 0.9699 m/s to 0.6759 m/s. These results demonstrate the efficacy and novelty of hybridized physics machine learning approaches as a scalable, interpretable solution for understanding wake loss in wind energy applications.