Mooring Tendon Dynamic Tension Estimation in a 15 MW TLP-Type FOWT: A Comparison of Self-Attention, LSTM, and GRU Networks

Tension Leg Platform (TLP)-type Floating Offshore Wind Turbines (FOWTs) rely on continuously pre-tensioned mooring tendons, the integrity of which must be monitored to ensure safe operation. However, direct measurement of tendon tension at submerged locations is difficult in practice. This study investigates a virtual sensing approach in which the effective tension at multiple tendon points—fairlead, middle, and anchor—of a 15 MW TLP-type FOWT is estimated from responses measured at or near the free surface. Three deep learning architectures are comparatively evaluated: a Transformer-encoder-based self-attention network (EN-ATT), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Fully coupled time-domain simulations are used to generate the training and test data, and the models are assessed under both nominal and noisy input conditions across multiple noise levels. The EN-ATT achieves the highest accuracy in terms of RMSE, MAE, and the coefficient of determination under both nominal and noisy conditions. Feature gradient analysis indicates that the self-attention model exhibits greater sensitivity to longer input lags than the recurrent networks. While the EN-ATT does not consistently outperform the recurrent networks for extreme values, the results suggest that self-attention architectures are a promising direction for mooring tension monitoring of TLP-type FOWTs.