A Novel Methodology to Warn Pre-icing Events for Wind Turbines

Around one-third of wind turbines are deployed in cold climates. It incurs risks of blade icing and potential emergency stops of wind turbines in operation. An accurate prediction of pre-icing events is a solution to mitigate these risks. However, amid the noise and data unbalance challenges, there is a lack of methods that have the mass-scale potential to be applied in industry applications. To overcome this challenge, this paper develops a novel prediction methodology for pre-icing event detections. This methodology involves a new structure, including the two-stage data rebalancing step and the classification step. First, it adopts classic under-sampling methods to rebalance the original dataset with both normal and pre-icing event data. Then, clustering is adopted to further rebalance the compressed dataset. The third step trained a classification model on top of the rebalanced dataset for making pre-icing predictions. This methodology has mass-scale application potential in terms of involving classic algorithms with low tuning difficulties. Through validation using real industry data, the overall prediction precision is over 99% and the recall rate is over 98% half an hour before the icing-induced emergency stops of wind turbines. To promote the application for different pre-icing datasets, this paper customized an algorithm tuning principle to find the optimal combination of methods at different stages.