Modelagem da curva de potência de turbinas eólicas com processos gaussianos

Abstract

In this dissertation, the wind turbine power curve (WTPC) modeling problem is revisited with the objective of proposing and evaluating a new semi-parametric, probabilistic and data-driven modeling framework. For this purpose, Gaussian processes and their heteroscedastic and robust extensions are combined with logistic functions, resulting in models which resemble the sigmoidal shape expected for WTPCs, output probabilistic predictions properly modeling the heteroscedastic behavior of the phenomenon and are robust to outliers. The proposed modeling framework is compared to multiple modeling benchmarks found in both the technical and scientific WTPC literature, namely, the method of bins, polynomial regression, neural networks, logistic functions and standard Gaussian process regression. Using a rich dataset of 1-year of operational data of a wind turbine, all models are compared in multiple scenarios concerning the key features of the WTPC modeling problem. The results show that the proposed modeling framework has competitive results regarding deterministic metrics when compared to the evaluated benchmark models, while also exhibiting the desired probabilistic properties, which gives it the ability to properly represent uncertainties intrinsically found in WTPC modeling.