Abstract
The roll motion of a Floating Production Storage and Offloading (FPSO) unit has complex interactions that are difficult to be modeled precisely: non-linear damping depending on the motion amplitude and velocity and influence of the mooring lines forces, for example. In order to improve the estimation of the roll motion of an FPSO, a new model is proposed combining the already established physics model with a regression neural network developed to calculate the residue between the physics model and the vessel’s actual measured motions. More specifically, the regression neural network corrects the maximum roll amplitude and its standard deviation obtained from a time domain simulation of the physics model.
This study considers a typical spread-moored FPSO located offshore on the Brazilian coast. The physics model considers the 6 degrees of freedom dynamics of a floating body subject to different drafts, mooring lines and environmental forces (wind, current, waves). A conventional Multi-Layer Perceptron (MLP) neural network is trained with data from more than 44,000 samples taken from actual environmental conditions and corresponding measurements of the platform roll motion between the years 2007 and 2014. The resulting hybrid physics-MLP model receives as input the environmental conditions and the FPSO draft, executes a 1-hour simulation of the physics model, derives the maximum roll amplitude and its standard deviation, and then the MLP compensates for the expected residue of the physics model. By adding information from real measurements to the mathematical model, the previously non-modeled effects are now considered.
The present paper concludes that the hybrid architecture combining a physics-based model and an MLP has improved the estimation of the roll motion, correcting the overdamping previously observed.
