Advanced Machine Learning and Optimization Techniques for Real-Time CO2 Membrane Operations in the Gulf of Thailand

Abstract This project aims to optimize the real-time performance of the Gulf of Thailand Offshore CO2 removal membrane system, overcoming challenges posed by interdependent parameters. This optimization system enables improvements in separation efficiency, minimizes hydrocarbon loss, and ensures compliance with sales gas specifications. The scope includes developing predictive models to uncover hidden correlations, optimizing key process variables, and providing actionable recommendations that enable operators to achieve optimal performance under dynamic conditions. The development of this optimization system comprises two components: Prediction Model Development and Optimization Model Development. The prediction models were developed by integrating machine learning with theoretical physics to improve accuracy. The development process involved parameter identification, data exploration, and model evaluation using non–tree-based techniques. The optimization phase aims to reduce hydrocarbon loss while ensuring operational efficiency. By leveraging predictive insights and real-time plant conditions, the optimizer refines key adjustable parameters, providing actionable recommendations for optimal performance. This data-driven approach overcomes traditional modeling limitations. Prediction models and optimization algorithms have been successfully developed for two offshore processing platforms, Arthit Processing Platform (APP) and Greater Bongkot South Processing Platform (PPS) with implementation currently in progress. The first development phase focused on prediction models, which achieved strong performance. For example, at one platform, the model for CO2 in the combined permeate stream yielded a Mean Absolute Error (MAE) of 1.95% (mean value: 75%) given its role in monitoring GHG reduction in flare gas, while CO2 in the combined raffinate stream achieved an MAE of 0.21% (mean value: 22%), demonstrating required precision for controlling sales gas specifications. These models support the production team by enabling accurate output predictions under varying input conditions. The second phase involved developing an optimization algorithm that ensures gas nomination targets, sales gas specifications, and membrane constraints. The algorithm was tested using historical operational data and showed that the optimized decision variables aligned with theoretical expectations. Field testing of this optimization system has been conducted on PPS, achieving an average hydrocarbon loss reduction of 558,000 SCFD, and is expected to reduce 500,000 SCFD at APP. These improvements are projected to yield savings of approximately 2.14 MMUSD, based on increased hydrocarbon reserve and reduced CO2 emissions. This paper presents a novel application of machine learning to model CO2 membrane separation, a process that remained a black box—even to simulation software—due to its complex behavior. By enabling machine learning based performance prediction, it helps close modeling gap in the industry. Combined with real-time optimization, it reduces non-value-adding tasks for operators, lowers reliance on experience-based decisions. Finally, it supports GHG emission reduction through more efficient gas processing and reduced hydrocarbon losses.