Condition Monitoring of Wind Turbine Systems by Explainable Artificial Intelligence Techniques.

The performance evaluation of wind turbines operating in real-world environments typically relies on analyzing the power curve, which shows the relationship between wind speed and power output. However, conventional univariate models that consider only wind speed as an input variable often fail to fully explain the observed performance of wind turbines, as power output depends on multiple variables, including working parameters and ambient conditions. To overcome this limitation, the use of multivariate power curves that consider multiple input variables needs to be explored. Therefore, this study advocates for the application of explainable artificial intelligence (XAI) methods in constructing data-driven power curve models that incorporate multiple input variables for condition monitoring purposes. The proposed workflow aims to establish a reproducible method for identifying the most appropriate input variables from a more comprehensive set than is usually considered in the literature. Initially, a sequential feature selection approach is employed to minimize the root-mean-square error between measurements and model estimates. Subsequently, Shapley coefficients are computed for the selected input variables to estimate their contribution towards explaining the average error. Two real-world data sets, representing wind turbines with different technologies, are discussed to illustrate the application of the proposed method. The experimental results of this study validate the effectiveness of the proposed methodology in detecting hidden anomalies. The methodology successfully identifies a new set of highly explanatory variables linked to the mechanical or electrical control of the rotor and blade pitch, which have not been previously explored in the literature. These findings highlight the novel insights provided by the methodology in uncovering crucial variables that significantly contribute to anomaly detection.