Interpretable statistical power curve models for improving onshore wind turbine condition and performance monitoring

Accurate power curve modelling plays a critical role in the continuous assessment of wind turbine (WT) performance and condition monitoring. This study proposes explainable data-driven models that use data fusion by integrating operational data from Supervisory Control and Data Acquisition (SCADA) systems with environmental measurements (e.g., wind speed and direction from met mast data) to model wind turbine power curves for fault detection, performance monitoring and condition monitoring purposes. Three advanced non-parametric and explainable statistical techniques, namely, Bayesian Additive Regression Trees (BART), Generalized Additive Models for Location, Scale, and Shape (GAMLSS), and Generalized Additive Mixed Models (GAMMs) will be used and compared to identify the most accurate approach for power curve modelling. The modelled power curves are built using historical operational and met mast datasets of operational three-bladed pitch-regulated wind turbines sourced from Whitelee Wind Farm located in Scotland. Model performance is evaluated using standard error metrics to identify the most effective approach. It is found that a GAMMs-based power curve model returned the lowest error metrics; however, the margin between GAMMs and GAMLSS is below 0.2% in RMSE and therefore within the variability expected from a single train–validation partition. The two techniques are consequently regarded as practically equivalent in point-prediction accuracy, and GAMMs is preferred on interpretability, residual-behaviour and confidence-interval-consistency grounds rather than on a demonstrated accuracy advantage. This study highlights the advantages and limitations of the proposed explainable statistical techniques in developing a robust fault detection algorithm for wind turbines using power curves. The primary model inputs comprise hub-height wind speed, atmospheric pressure, air temperature, and turbulence intensity, all derived from integrated SCADA and met mast measurements. The study employs data from a Siemens SWT-2.3-108 onshore turbine at Whitelee Wind Farm, Scotland, and discusses the transferability of the proposed framework to broader multi-turbine operational contexts.