Posters

Abstract

Accurate estimation of turbine power without direct wind measurements is a key challenge for upgrade assessment and operational benchmarking. This study introduces a neighbor-aware ensemble framework that leverages operational data from adjacent turbines to estimate the target turbine’s power output. The methodology combines physics-informed feature engineering, stacked ensemble learning, and conformalized quantile regression (CQR), enabling both high-accuracy predictions and calibrated uncertainty quantification. The case study is based on the Hill of Towie Wind Farm dataset from the WeDoWind Challenge [5], using SCADA and ERA5 reanalysis data. Features were enriched with temporal lags, yaw–wind misalignment, air density adjustments, residuals from an empirical power curve, and cyclic encodings to capture diurnal and seasonal variability. An ensemble of LightGBM, XGBoost, CatBoost, and Random Forest was trained with AutoGluon [6], while conformal calibration was applied to quantile regression for distribution-free interval estimation. Results show that the proposed Ensemble-CQR (FE) model achieves the highest accuracy (R² = 0.992; MAE = 47.55 kW), outperforming single models and baseline ensembles. Importantly, it delivers well-calibrated prediction intervals with an empirical coverage of 89.9% for nominal 90% intervals, ensuring alignment between confidence levels and observed coverage. This capability overcomes the common coverage mismatches seen in quantile- or ensemble-only approaches [2, 3]. Overall, the framework demonstrates that leveraging neighbor turbine data allows reliable power estimation even in the absence of direct wind speed inputs. This offers significant practical value for turbine upgrade performance assessment and risk-aware decision-making in wind power integration and retrofitting strategies [1, 4].