Probabilistic evaluation for early wind turbine yaw misalignment detection

Abstract

Nowadays, one of the biggest challenges for wind turbines is to reduce operation and maintenance costs. Therefore, it is essential to develop predictive maintenance, anticipating failures early and thus avoiding unnecessary actions on the wind turbine. In this way, the uptime and performance of the turbine are maximized, and its useful life is extended. This work describes a general methodology for fault detection based on probabilistic models and its evaluation. This methodology combines a fault detection method based on the Fisher Test and the development of probabilistic models of wind turbine power curves. Several probabilistic models of power curves have been evaluated: Gaussian mixture model (GMM), Frank copula model, Gaussian mixture copula model (GMCM), Gaussian process regression (GPR) and epsilon-insensitive loss function support vector regression (ε-SVR). The results indicate that the Gaussian mixture copula model is the most efficient in terms of accuracy and computational cost. The detection of a wind turbine orientation misalignment error has been tested as a use case. It is shown how with this probabilistic approach it is possible to detect the fault in a short period of time from its appearance, 10–30 times faster than other techniques found in the literature with which it has been compared.