Machine Learning-Based Operability Assessment of Side-by-Side Offloading

Abstract

Operability assessment of multibody offshore operations is commonly performed using large numbers of high-fidelity time-domain simulations. When environmental conditions are discretised across realistic ranges of wave, wind, and current parameters, simulation campaigns can involve up to 10⁵ individual three-hour runs, resulting in substantial computational cost. This challenge is shared across a broad class of coupled floating-body operations, including offshore hydrocarbon offloading, CO₂ ship-to-ship transfer, and floating wind installation.

This thesis investigates the use of data-driven prediction models to support simulation-driven operability assessment while retaining response-based decision criteria, using an FLNG–LNGC side-by-side offloading configuration as a case study. The study is based on a simulation dataset produced by the Maritime Research Institute Netherlands (MARIN) for Shell, consisting of time-domain simulations of a spread-moored FLNG with an LNG carrier in side-by-side configuration. From this dataset, a sea-wave subset of 8317 simulations is used for model development and evaluation. Each simulation includes summary response extrema as well as full time-series records of wave elevation, vessel motions, mooring line tensions, and fender loads.

Two complementary machine learning (ML) approaches are considered. The first is a summary-statistics-based learning (SSBL) approach, which predicts operability-limiting response extrema directly from static environmental descriptors using feed-forward regression models. The second is a time-series-based learning (TSBL) approach operating on full response time series. Within this framework, a segment-based forecasting model based on long short-term memory (LSTM) networks and a full time-series model based on transformer architectures with cross-attention heads are employed. For time-series modelling, response signals are decomposed into wave-frequency and low-frequency components to distinguish first-order wave-induced behaviour from second-order drift-dominated dynamics.

At a training fraction of approximately 70%, wave-frequency models achieve mean coefficients of determination R² between 0.6 and 0.8, while low-frequency LSTM models attain mean R² values of approximately 0.8. To reduce training requirements while preserving operability-critical behaviour, a compact training dataset of 1500 simulations is derived using a two-stage selection strategy combining stratified sampling over key sea-state parameters with K-means clustering in response space.

For operability classification based on response extrema across the full sea-wave dataset, a multilayer perceptron trained on the reduced dataset achieves an overall accuracy of 98.22% for 6816 unseen sea states. Time-series-based models achieve an overall classification accuracy of 86.7% and enable reconstruction of complete multibody response histories when the training data sufficiently represent dominant physical regimes, including head-sea, beam, and near-beam wave conditions.

From a computational perspective, the original MARIN sea-only time-domain simulations require approximately 1428 core-hours to evaluate 8317 sea states. In contrast, transformer-based time-series models require approximately 18.9 GPU-hours for one-time training on 1500 sea states and 2.85 GPU-hours for prediction over the remaining sea states, while extrema-based model prediction is effectively instantaneous. These results demonstrate that data-driven models can substantially reduce the computational burden of simulation-driven operability assessment for coupled floating-body systems more broadly.