An early warning method for oil temperature in offshore wind power gearbox based on TCN-LSTM-ECA

Abstract

This paper introduces a novel digital twin-based method for early warning of oil temperature anomalies in offshore wind turbine gearboxes, leveraging a hybrid Temporal Convolutional Network (TCN), Long Short-Term Memory (LSTM), and Efficient Channel Attention (ECA) framework to analyze Supervisory Control and Data Acquisition (SCADA) data from an operational wind farm. The TCN-LSTM-ECA model integrates TCNs for capturing local temporal dependencies, LSTMs for modeling long-range dependencies, and ECA for prioritizing critical features, achieving an RMSE of 0.005956 and MAE of 0.004709, significantly outperforming baseline TCN, LSTM, and TCN-LSTM models, and providing advance warnings of up to 27 h and 58 min. The TCN-LSTM-ECA model significantly enhances prediction accuracy and timeliness compared to baseline models, offering substantial improvements in RMSE and MAE. By integrating XGBoost for feature selection and employing advanced data preprocessing techniques, this approach predicts potential gearbox issues well in advance and contributes to operational reliability and maintenance cost reduction. The results underscore the efficacy of combining temporal convolutional learning with attention mechanisms for robust predictive maintenance in the complex environment of offshore wind energy systems.