Machine learning based co-optimization of carbon dioxide sequestration and oil recovery in CO2-EOR project

Abstract

This paper presents a machine learning assisted computational workflow to optimize a CO2-WAG project considering both hydrocarbon recovery and CO2 sequestration efficacies. A compositional field-scaled numerical simulation model is structured to investigate the fluid flow dynamics of an on-going CO2-EOR project in the Farnsworth Unit (Texas, US). Artificial-neural-network (ANN) based proxy models are trained to predict time-series project responses including hydrocarbon production, CO2 storage and reservoir pressure data. The outputs of the proxy model not only serve for evaluating the objective function but also provide significant physical and economic constraints to the optimization processes. In this work, the objective function considers both the oil recovery and CO2 sequestration volume. Moreover, the project net present values (NPV) and reservoir pressure are employed to screen the optimum solutions. The proposed optimization workflow couples the Particle Swarm Optimization (PSO) algorithm and the ANN proxies to maximize the prescribed objective function. The results of this work indicate that the presented workflow is a more robust approach to co-optimize the CO2-EOR projects. Results show that the optimized case can store about 94% of the purchased CO2 within Farnsworth Unit. Comparing to the baseline case, the CO2 storage amount of the found optimal case increases by 21.69%, and the oil production improves 8.74%. More importantly, the improvements in CO2 storage and hydrocarbon recovery lead to 8.74% greater project NPV and 19.79% higher overall objective function value, which confirms the success of the developed co-optimization approach for CO2 sequestration and oil recovery. The lessons and experiences earned from this work provides significant insights into the decision-making process of similar CO2-EOR cases.