Early fault prediction for wind turbines based on deep learning

This research focuses on the predictive maintenance of wind turbines, using operational data of 31 wind turbines located in Taiwan's Changbin Industrial Zone, for a total of five years from 2015 to 2019. A hybrid method fault prediction mechanism for wind turbines is developed using machine learning and deep learning methods. The random forest method is applied to identify features that are highly correlated with faults, and to eliminate low-correlation features to maximize prediction model efficiency. Long short-term memory (LSTM) deep learning methods are then applied to handle the time series data, analyze historical pre-failure information, use the dynamic weight loss function to address data imbalance, and finally predict the future wind turbine health status. The resulting fault prediction model produces average prediction accuracy, precision and recall rates of 99%, 70% and 77%, respectively for predictions of one to six hours ahead, indicating that the proposed model can effectively predict wind turbine failures in advance, thus providing increased time for fault response and effectively improving the wind turbine lifespan.