Leveraging Machine Learning to Optimize CO2-WAG Flooding for Enhanced Oil Recovery and Carbon Storage

Abstract The utilization of CO2-carbonated water-alternating-gas (CWAG) has been extensively studied as an EOR technology, demonstrating significant promise in carbon capture, utilization, and storage (CCUS). The objective of our work is to create a reliable and precise machine learning model and optimization workflow that optimizes operation parameters and accurately predicts production levels, thereby enhancing the efficiency of the CWAG process. Following the development of the reservoir numerical model with typical numerical simulation software, this study proceeded to build up 5000 distinct numerical models. Each of these models was characterized by varying reservoir geology parameters, fluid parameters, initial conditions, as well as operation parameters. To assess the performance of the reservoir numerical models, cumulative oil production and carbon storage were calculated as prediction and evaluation indicators. A training set comprising 70% of the available data was utilized, while the remaining 30% served as the evaluation set. This division ensured a robust evaluation of the machine learning regression methods employed in the study. Then seven regression machine learning methods was applied to conduct the analysis, enabling a thorough examination of the dataset, and facilitating the training and evaluation processes. Following the evaluation process, the XGBoost algorithm achieved the best prediction performance among all methods. Building upon this finding, an efficient and reliable optimization workflow was developed by integrating the trained XGBoost model with a hybrid genetic algorithm. The workflow aimed to enhance the efficiency and accuracy in optimizing various parameters related to the CO2-WAG operation process. This coupling workflow facilitates the swift optimization of parameters pertaining to CO2-WAG flooding in field development design, as well as the prediction of production outcomes under diverse geological conditions. The objective is to achieve high oil recovery and carbon storage efficiency simultaneously. By incorporating a broader range of factors into consideration, the XGBoost model demonstrates an impressive accuracy rate of 98%. This significant improvement in accuracy highlights the superiority of the XGBoost model over traditional numerical simulation methods. The research workflow and optimization workflow developed in this study exhibit good extensibility, making them applicable to a wide range of EOR methods beyond CO2-WAG. They can be effectively utilized for the prediction and optimization of water flooding, chemical flooding, and other EOR techniques.