Prediction of Coral South FLNG Side-by-Side Offloading Operability Using Radial Basis Function Artificial Neural Networks

Abstract

Coral South Floating Liquefied Natural Gas (FLNG) unit is designed to offload its product to LNG Carriers (LNGC) moored in a Side-by-Side (SBS) configuration, using Marine Loading Arms (MLA) technology. With such a method, multiple design aspects require accurate hydrodynamic simulations at an early stage of engineering phase for a large number of environmental conditions, including wind sea and swell: sizing of FLNG mooring outfitting, design of the MLAs for the candidate LNGC fleet and availability of LNG offloading. Complex phenomena like multi-body hydrodynamic coupling and LNG sloshing in partially filled tanks must be accounted for, and non-linear berthing characteristics require time-domain simulations. However the computing time for all the environmental conditions described in the available 23-year hindcast databases combined with all the possibilities of FLNG heading obtained with thruster assistance, for multiple LNGC and loading conditions, is not compatible with the project engineering phase timeframe, so the simulations must be first performed on a suitable-size sample of environmental conditions, creating a database which can then be used for predicting the data related to any environment. The selected sample of environmental conditions must meet the constraints of computing time while keeping a sufficient resolution to be reliable. In other projects, the time-domain quantities subject to operability criteria, derived from this limited number of simulations, were used to predict the behavior for any unknown environment by interpolation. This approach presented some limitations like the overfitting of maxima dependent on the wave realization (seed), which is seen as a noise, and was not best suited for generalizing the results to non-simulated environments out of the sample. In this paper, the improved methodology used for the Coral South FLNG project is presented. A Radial Basis Function (RBF) Artificial Neural Network (ANN) is used to model the variables impacting the SBS offloading operability. The ANN learns from the results of simulations performed on a sample defined with a K-means clustering algorithm. The RBF is modified to be adapted to the specifics of the driving parameters, of which some are periodic (wave direction) and the rest non-periodic. A proper smoothing of seed-dependent maxima and accurate estimations for unknown environments (generalizations) are achieved. The learning process does not require significant computing time and fewer preliminary time-domain simulations are needed. This design methodology represents a significant improvement for the calculations performed during the project's engineering phase, but it may also be applied later once offshore, to assist in decision making relative to the weekly forecasted SBS LNG offloading operations. When Coral South FLNG operates, the learning database may be completed with on-site measurements to further improve its accuracy. The principle may also be extended to other offshore operations constrained by environmental conditions.