Intelligent early fault warning for offshore wind turbine generators: a hybrid deep learning model with causal condition-adaptive dynamic thresholds

Offshore wind turbine generators face harsh, variable marine conditions that heighten unexpected failure risk. Conventional early warning systems suffer from limited predictive accuracy and static thresholds that ignore operational dynamics, leading to false alarms and missed detections. To address these challenges, this study proposes a novel hybrid deep learning model named TGA (TCN–GRU–Attention) with adaptive dynamic thresholds for wind turbine fault warning. The model integrates a temporal convolutional network (TCN), a gated recurrent unit (GRU), and an attention mechanism to achieve high precision in generator temperature forecasting. Based on the TGA predictions, a condition-adaptive dynamic thresholding strategy is introduced, in which absolute prediction residuals are modeled using a causal sliding-window quantile method. This enables real-time adjustment of alarm boundaries in response to changing operating conditions. Validation on real offshore wind farm SCADA data demonstrates that the proposed TGA-based approach achieves lower prediction errors and more stable absolute residuals than representative baselines. Furthermore, the adaptive thresholding strategy effectively reduces false alarms and missed detections while improving the timeliness of fault warnings. In summary, the proposed framework offers a feasible and data-driven solution for intelligent fault early warning, with demonstrated benefits for enhancing the reliability of offshore wind turbine operations.