A domain-adaptive deep neural network for real-time mooring tension prediction in floating wind turbines

A domain-adaptive deep learning framework that learns sea-state-invariant features under varying operational domains is proposed. This allows real-time prediction of mooring line tension in floating offshore wind turbines, which is essential for health monitoring under uncertain sea states and is an improvement over traditional physics-based approaches, which are limited to generalising across different environmental conditions. Specifically, a Domain-Adaptive Convolutional Neural Network–Long Short-Term Memory–Attention (DACLA) model is developed, integrating convolutional layers for spatial feature extraction, recurrent units for temporal modelling, and an attention mechanism for dynamic feature weighting. To enhance transferability, a gradient reversal layer-based adversarial training strategy is incorporated, aligning feature distributions between source and target domains. The IEA 15 MW wind turbine serves as the case study. Extensive experiments demonstrate that the DACLA model significantly improves prediction accuracy across wave and wind variations compared to conventional architectures. Furthermore, frequency-domain power spectral density analysis is introduced to interpret the cross-domain behaviour, offering insights into spectral fidelity and dynamic consistency. Finally, robustness tests under Gaussian noise and wind-wave misalignment validate the model's resilience to sensor-level disturbances. These findings confirm that the proposed DACLA framework offers a scalable and adaptive solution for real-time mooring tension forecasting in offshore floating wind applications. • A domain-adaptive deep learning framework is proposed that learns sea-state-invariant features across varying domains. • This enables the real-time prediction of mooring line tension in floating offshore wind turbines. • A Domain-Adaptive Convolutional Neural Network–Long Short-Term Memory–Attention (DACLA) model is developed. • To enhance transferability, a gradient reversal layer-based adversarial training strategy is employed. • Extensive experiments demonstrate that the DACLA model significantly improves prediction accuracy.