Wind Turbine Fault Prediction Based On A Novel Gated Recurrent Neural Network Model

Due to the harsh environmental issue and hard accessibility, offshore wind turbines (WTs) have more challenges for operation and maintenance (O&M). Thus, it is crucial to develop effective condition monitoring (CM) methods for WT fault prediction to detect incipient faults before their occurrences, thus preventing durable downtimes. In this paper, eight specific faults are classified for fault prediction using status information from Supervisory Control and Data Acquisition (SCADA) data. The classification steps are based on fault prediction from 10 to 210 minutes prior to faults. By embedding a model-agnostic vector representation for time, Time2Vec (T2V), into Gated Recurrent Unit (GRU), a novel deep learning neural network model, T2V-GRU, is applied for fault classifications. As a result, T2V-GRU successfully predicts over 84.62% of faults and outperforms its counterpart, vanilla GRU, in both overall and individual fault predictions in terms of accuracy, recall scores and F-scores.