A meteorology-free wind power forecasting using a hybrid physics-aware, SCADA-only ensemble residual learning

This study introduces two hybrid forecasting approaches, bi-hybrid (bi-HM) and triple-hybrid models (triple-HM), that integrate physics-informed learning with machine learning (ML) using only SCADA data. The bi-HM amalgamates a Physics-Informed Neural Network (PINN) with a Bayesian-optimized CatBoostRegressor through adaptive weighting, while the triple-HM extends this by incorporating an XGBoost-based stacked ensemble. Both methods are trained and validated on 10-minute SCADA records from a Nordex N117/3600 turbine at the Esenköy Wind Farm (Turkey) over one year. Experimental results demonstrate that the triple-HM achieves superior predictive performance, with an R² of 0.9724 and RMSE of 0.0576, outperforming standalone PINN (R² = 0.9564), CatBoost (R² = 0.9577), and XGBoost (R² = 0.9551). Compared to recent studies, the triple-HM surpasses the best reported LSTM model (2023, R² = 0.9574) and the best recorded XGBoost model (2022, R² = 0.9600). Furthermore, inference time remains below 0.2 s, ensuring real-time feasibility. A SHAP-based analysis highlights the complementary roles of physics-guided corrections and ML-driven generalization, enhancing interpretability and physical plausibility. The proposed hybrid models (bi-HM, triple-HM) have demonstrated high predictive accuracy across geographically distinct datasets, outperforming conventional regression methods on the Texas wind turbine dataset. The triple-HM model outperformed conventional regressors of Ridge, LR, and BR (R² of approximately 0.962), achieving an R² of 0.9999 and comparable R² of 0.9998 respectively. Moreover, the ablation results showed that while PINN, CatBoost, and XGBoost individually achieve strong accuracy, their integration in hybrid architectures yields significantly improved accuracy and generalization, especially with the triple-HM.