Graph-Attention Long-Short-Term-Memory Network (GAT+LSTM) for Mooring Line Load-Field Estimation Across Variable Topologies

Abstract

Reliable prediction of mooring-line motion and tension underpins the station-keeping safety of every floating offshore platform, whether built for wind energy, wave energy, or oil and gas production. High-fidelity finite-element solvers resolve the nonlinear dynamics that govern this safety, including snap loads and peak tensions, but their computational cost makes repeated evaluation across design iterations or environmental conditions impractical. Deep-learning approaches promise to overcome this cost. Yet existing mooring-line surrogates either frame the problem as failure classification or reduce the line to a lumped fairlead reaction force, so none reconstructs the full spatial field of motion and tension along lines of different discretisations with a single trained model, nor biases its predictions toward the conservative peak estimates that structural design requires.

This thesis develops a surrogate that couples a graph attention network (GAT) spatial encoder to a long short-term memory (LSTM) temporal encoder, denoted GAT+LSTM, and that addresses this gap. A mooring line is represented as a chain graph whose nodes and edges carry static geometric and mechanical features alongside dynamic motion, tension, and contact features. A two-layer GATv2 encoder captures the coupling between neighbouring nodes, a per-node LSTM encoder captures each node's temporal evolution, and a compact prediction head returns the future motion and tension of the whole line in one non-autoregressive pass. The network is trained under a mixed-topology scheme on finite-element simulations spanning ten offshore locations and node counts from N = 4 to N = 21, with two locations withheld to test generalisation, under two compared training objectives, a fixed-weight loss combining data fidelity with soft physical penalty terms in physical units, and a Kendall multi-task uncertainty-weighting scheme intended to learn those weights automatically.

One set of trained weights serves every discretisation from N = 4 to N = 21 nodes and both withheld locations, with test accuracy meeting or exceeding validation accuracy at every node count. The surrogate reconstructs the spatial field with a global tension R^2 of 0.98. A variance decomposition shows this accuracy certifies the response's spatial profile and per-sea-state level, together at least 98.5% of its variance, not the wave-frequency fluctuation, which the model returns as a near-flat mean. Under the fixed-weight objective this yields a positive mean peak-tension bias of +197.5 N and an underprediction rate of 17.6%. The uncertainty-weighted alternative matches this safety margin only once its one-sided penalty is reformulated as a two-sided quantile loss, and even then it does not surpass it.

The thesis therefore contributes a validated, topology-general load-field surrogate, a safety-aware training methodology with the diagnosis of its failure mode, and a rigorous characterisation of what the surrogate's accuracy does and does not certify. The surrogate is deliberately not a wave-by-wave forecaster of mooring tension. Its value lies in reproducing the spatial load field, across topologies and locations, at a fraction of the finite-element solver's computational cost.