Enhancing Wind Turbine Power Curve Monitoring with eXplainable Artificial Intelligence Techniques

The power of a wind turbine has a multivariate dependence on environmental conditions and working parameters. Hence, traditional univariate models which take as input solely the wind speed may fail in effectively monitoring the performance of wind turbines. To address this limitation, the use of multivariate wind power curves, which consider multiple variables and output power, should be investigated. For this reason, the Authors of this paper advocate the role of eXplainable Artificial Intelligence (XAI) methods in the construction of multivariate data-driven power curve models. A sequential features selection is applied and the Shapley coefficients for the various input sensors are computed on a real-world data set, which is composed of a set of covariates larger than the state of the art in the literature. The experimental results of this work lead to include in the model highly explanatory variables related to the mechanical or electrical control, such as the rotational speed and the blade pitch pressures, which use has not been contemplated in the literature before.