Machine learning-based icing prediction on wind turbines

In cold regions like Northern Europe or Northern America, formation of ice on wind turbines not only reduces the power output of the system, but also leads to wind turbine breakdowns and poses a risk to nearby vehicles and people. Economical solutions for anti-icing systems such as preventive heating are required. Purely meteorological icing predictions are not sufficient to accurately predict icing on wind turbines. Therefore, this paper proposes an icing prediction approach that uses historical weather data and data from a supervisory control and data acquisition (SCADA) system plus methods from supervised machine learning to predict the risk of icing. The first results of the prediction model show its capability to predict most of the icing events. The discussion shows that the model should consider more meteorological data that describe the icing risk such as humidity or liquid water content to achieve better prediction results. The final aim is a prediction system that will use real-time data to determine icing risks hours beforehand, allowing the operation of existing anti-icing systems more efficiently.