Wind turbine power curves play a key role in various aspects during the life of a wind farm. Typical uses range from wind power forecasting to wind turbine condition monitoring. This paper addresses the identification of probabilistic models of wind power curves from observed data. The main challenge is the need to handle a statistical distribution of wind energy whose shape not only may be highly skewed, but can also change with wind speed. To address these issues, we resort to the framework of Generalized Linear Models (GLMs), proposing a Beta regression approach, with constant or variable dispersion and an appropriate preconditioning step. The proposed methodology was tested on three real SCADA measurements retrieved from public datasets, including a comparison with Quantile Regression Forests (QRFs), also in terms of robustness to outliers. The results suggest that Beta regression can be a valuable resource in the development of probabilistic models for wind energy, as it provides a high degree of flexibility while preserving an interpretable structure.
