Posters

Abstract

In the pursuit for effective predictive maintenance capable of prevention of downtime events of wind turbine generators (WTGs), well-functioning prediction models are essential. According to literature, several methods and models are available for prediction of these downtime events [1-3], from icing event prediction [4] to gearbox failures [5], but there seem to be considerable challenges bringing these solutions to market. One of the challenges seem to be that the available publications on these machine learning models are often calibrated to a very small population size, e.g., single WTGs or single wind sites. The performance of such models decreases to a significant degree as the underlying population of WTGs are expanded, as this will introduce increasingly heterogeneous data distributions in the training data [6]. This drift in input data come from the different designs of the WTGs which impact the behavior significantly and from the different meteorological characteristics specific to the region of the WTG. From the perspective of an OEM, this is an issue of special concern. As the service provider to several thousand WTGs globally, the machine learning models employed are desired to be as scalable and widely applicable as possible, since it is an impossible task to develop a large set of individually specialized models. To balance the scalability of the models with the detrimental effects of increasing the level of heterogeneous input data, we have developed a WTG clustering method designed to create training datasets with the least heterogeneous data possible. With this method, we are capable of developing effective predictive models that are applicable to the broadest selection of WTGs