Prelude FLNG Free Weathervaning Heading Prediction and Uncertainties, Based on Machine Learning Model

Abstract

Abstract Prelude Floating Liquefied Natural Gas (FLNG) facility is moored with an internal turret allowing it to free weathervane (FW), i.e. by leaving the unit to rotate according to environmental loads. During the engineering phase, the FLNG FW heading is estimated by the heading analysis (i.e. physics-based approach), and results are then used as input for other studies. Therefore, a good estimation of the various environmental effects (waves, current and wind) and their contributions in terms of loads on the FLNG is critical to ensure a correct prediction of the FW heading. For the predominant contributions (wind and current), the force coefficients have been initially derived from wind tunnel tests during the engineering phase. However, Prelude FLNG being now installed on-site, measurements over recent years have shown slight discrepancies with the numerical predictions by the heading analysis. Preliminary investigations were carried out and were aimed to improve some parameters of the numerical model. Nevertheless, it appeared that even with these improvements, discrepancies between numerical predictions and measurements were not always resolved. These discrepancies may have several origins, such as inadequacy of the numerical model, variability of the metocean data, uncertainties in measurements, etc. In order to overcome the aforementioned uncertainties and unknowns, it has been decided to set-up a machine learning model (i.e. data-based approach). This machine learning model (RBF ANN - Radial Basis Function Artificial Neural Network) was trained with the recorded metocean data (input) and measured FLNG FW heading (output). Considering the amount of the measured data available (two years with a time step of 10 minutes), the necessity to optimize the model’s hyperparameters and the computer capability, a stepwise approach has been applied to ensure an accurate model can be built in a reasonable timeframe. Finally, the machine learning model calculation shows a significant improvement in the prediction capability when compared to the measured FLNG FW heading. The resulting surrogate model is hence used to predict the FW heading and to derive the associated prediction intervals, which define the range of error with certain probability (for instance 95%). This paper describes the machine learning model used, the methodology and challenges of the approach, and discusses the results. The main conclusions and lessons learnt are also shared.