Robust optimization of flue gas-WAG injection using data-driven proxy models in egg oil reservoir

Enhanced oil recovery (EOR) using carbon dioxide (CO₂), particularly in the form of flue gas without costly separation processes, presents a dual benefit: increased oil production and long-term CO₂ storage. However, optimizing flue gas-based water alternating gas (WAG) injection is computationally expensive due to the need for numerous reservoir simulations. This study addresses this challenge by developing fast and accurate machine learning-based surrogate models namely, generalized regression neural network (GRNN) and cascaded forward neural network (CFNN) models to estimate the net present value (NPV) of the flue gas-WAG process in the Egg reservoir model. The GRNN model demonstrated exceptional predictive performance, with an R² of 0.999 and an average absolute percentage relative error (AAPRE) of 2.546% across the dataset. Optimization of injection parameters was performed using four metaheuristic algorithms: grey wolf optimizer (GWO), particle swarm optimization (PSO), ant colony optimization (ACO), and genetic algorithm (GA). Among them, GWO achieved the best results, identifying an optimal NPV of $77.16 million with superior convergence speed and solution accuracy. The novelty of this work lies in the integration of adaptive machine learning proxies with metaheuristic optimization to significantly reduce computational cost while maintaining accuracy. Unlike previous studies that focused on CO₂-EOR with separated CO₂ streams, this research demonstrates the viability of directly using flue gas as an injection medium, offering a more practical and cost-effective solution for field-scale applications.