Gaussian processes with logistic mean function for modeling wind turbine power curves

Abstract

The wind turbine power curve (WTPC) is a mathematical model built to capture the input-output relationship between the generated electrical power and the wind speed. An adequately fitted WTPC aids in wind energy assessment and prediction since the actual power curve will differ from that provided by the manufacturer due to a variety of reasons, such as the topography of the wind farm, equipment aging, and multiple system faults. As such, this paper introduces a novel approach for WTPC modeling that combines Gaussian process (GP) regression, a class of probabilistic kernel-based machine learning models, and standard logistic functions. This semi-parametric approach follows a Bayesian reasoning, in the sense of maximizing the marginal likelihood to learn the parameters and hyperparameters through a variational sparse approximation to the GP model. Using real-world operational data, the proposed approach is compared with the state-of-the-art in WTPC modeling and with an alternative probabilistic approach based on generalized linear models and logistic functions. Finally, we evaluate the proposed model in its extrapolation ability for unmodelled data.