Machine Learning Based FPSO Topsides Weight Estimation for a Project on an Early Stage

Abstract

Abstract The objective of this study is to develop a data-driven machine learning based tool to estimate the FPSO topsides weight. The data were collected from public sources including IHS, news and magazines, covering world-wide active FPSO geographic locations, topsides weights, and their production throughput. One of the challenges is that the size of the dataset is less than 200 data points, largely due to the limited total number of FPSOs worldwide. Another challenge is that there are missing values for gas production, as such, imputation of missing values becomes necessary. In this study, data imputation was conducted by incorporating geographic information and physics guided feature engineering, through which the imputation is more accurate compared to simple imputers. For machine learning algorithms, polynomial regression was first evaluated as the baseline model and various machine learning models were built and compared with the baseline, such as Gaussian process regressor, random forest, neural network, and natural gradient boosting, with the purpose of identifying the most accurate one. To solve the overfitting issue caused by the small size of the dataset, several strategies have been investigated and compared, such as k-fold cross validation, regularization and extensive hyper-parameter tuning via Bayesian optimization algorithm based on the Hyperopt library. Among all the machine learning models, it is found that the natural gradient boosting method is the best performer with a mean absolute percentage error (MAPE) of 24% on the blind testing data, which is 35% lower than the baseline model. Shapley Additive exPlanations (SHAP) analysis was also implemented for model interpretation and gas production was found to be the most influential feature. The trained gradient boosting model was deployed to an internal web application in which users could get a quick estimation of FPSO topsides weight by providing three features: gas production, oil production and water depth. The 2D and 3D cross plots with historic data and predicted value are also provided in the web-app for better results visualizations. The novelty of this paper is to develop a data-driven machine learning tool for FPSO topsides weight estimation on an early stage of a project, which can serve as an independent alternative to the traditional empirical based approaches to help pre-design the facilities and estimate the cost. In the back-end, the best machine learning model was identified, along with the best imputation strategy based on a physics guided feature engineering approach. In the front-end, a web application was developed for an interactive estimation of FPSO topsides weight. With continuous enrichment and validation of the collected data, the machine learning approach can serve as a trustworthy fast and early estimation for FPSO topsides weight.