Towards Robust Machine Learning Models for the Estimation of Mooring-Line Tension in Floating Wind Turbines under Unknown Sea Conditions

Given the practical limitations of direct measurement in harsh offshore environments, the indirect estimation of mooring-line tension in Floating Offshore Wind Turbines (FOWTs) constitutes a necessary approach for Structural Health Monitoring (SHM) and Remaining Useful Life (RUL) assessment. This study investigates the performance and robustness of several Machine Learning (ML)-based regression models for estimating mooring tensions from floater motion measurements under both known (in-distribution) and unknown (out-of-distribution, OOD) sea states. The analysis is based on wave-tank experiments performed on a scaled model of a typical floating structure with a catenary station-keeping system representative of a FOWT, namely the HarshLab floating laboratory installed at the Biscay Marine Energy Platform (BiMEP). A Multi-Layer Perceptron (MLP) augmented with temporal context outperforms not only traditional tree-based ensemble models but also the more complex sequence-learning architecture evaluated in this study, namely the Long Short-Term Memory network (LSTM). Our results show that all models achieve high accuracy under sea conditions represented in the training data, while their performance degrades as test conditions increasingly deviate from the training distribution, particularly under more severe sea states. In light of this observed performance degradation under OOD conditions, the ensemble variance of a tree-based ensemble regressor is further evaluated as a model degradation indicator (MDI), demonstrating its effectiveness in detecting global estimation-quality deterioration and potentially signaling when model retraining is required to incorporate the effects of previously unseen sea-state dynamics on mooring-line tension.