Comparison of advanced non-parametric models for wind turbine power curves

To continuously assess the performance of a wind turbine (WT), accurate power curve modelling is essential. Various statistical methods have been used to fit power curves to performance measurements; these are broadly classified into parametric and non-parametric methods. In this study, three advanced non-parametric approaches, namely: Gaussian Process (GP); Random Forest (RF); and Support Vector Machine (SVM) are assessed for WT power curve modelling. The modelled power curves are constructed using historical WT supervisory control and data acquisition, data obtained from operational three bladed pitch regulated WTs. The modelled power curve fitting performance is then compared using suitable performance, error metrics to identify the most accurate approach. It is found that a power curve based on a GP has the highest fitting accuracy, whereas the SVM approach gives poorer but acceptable results, over a restricted wind speed range. Power curves based on a GP or SVM provide smooth and continuous curves, whereas power curves based on the RF technique are neither smooth nor continuous. This study highlights the strengths and weaknesses of the proposed non-parametric techniques to construct a robust fault detection algorithm for WTs based on power curves.