Chenyu Wu

Date of Award

May 2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Petroleum Engineering

First Advisor

Kegang K. Ling

Abstract

CO2 Enhanced Oil Recovery (EOR) is increasingly recognized worldwide for its dual benefits of enhancing oil production and addressing environmental concerns related to greenhouse gas emissions. However, traditional methods for assessing the feasibility of CO2 EOR in a specific field, such as pilot tests and simulations, are often time-consuming and labor-intensive. A critical factor in determining the suitability of a field for CO2 EOR is the CO2 Minimum Miscibility Pressure (MMP), as miscible flooding is more feasible when the reservoir pressure exceeds the CO2 MMP. Traditional determination of CO2 MMP relies heavily on extensive experimentation, which, despite its accuracy, incurs significant time and labor costs.The advent of machine learning offers a promising avenue to streamline the CO2 EOR screening process and enhance CO2 MMP predictions' efficiency. This study presents a new CO2 EOR screening tool that incorporates a CO2 EOR scoring system alongside machine learning regression models. Additionally, for CO2 MMP predictions, the study goes beyond developing various machine-learning models by introducing a novel subgrouping method. This method categorizes the dataset based on the distinct impacts of compositional variables in injection gas and crude oil, creating distinct datasets for modeling, thereby improving model performance and predictive accuracy. For CO2 EOR screening, traditional screening criteria can be overly arbitrary, making it difficult to accurately assess whether a field is suitable for CO2 EOR—failing to meet a single requirement may result in an unsuitable judgment. Existing intelligent screening tools primarily focus on classification algorithms to develop predictive models for CO2 EOR feasibility. While these tools are accurate for shallow-depth reservoirs, they struggle to predict CO2 EOR in mid- and deep-depth reservoirs accurately. In this study, an EOR database was created based on 464 EOR projects from 18 countries. Reservoir data from the Williston Basin, TORIS (Tertiary Oil Recovery Information System), and Alberta Basin were also collected to further validate the feasibility of the proposed approach. A new screening criterion was developed based on the boxplot analysis results of the collected worldwide EOR projects and existing CO2 EOR screening guidelines. Weight factors for parameters were determined using the importance permutation technique and the proposed classification algorithms to minimize bias. An innovative CO2 EOR scoring approach was developed using membership functions, composite screening scores, and the six machine learning algorithms. The results showed high prediction accuracy for the worldwide EOR projects database with R-squared values ranging from 0.91 to 0.98. The proposed screening system was further employed to evaluate the prediction accuracy for the mid- and deep-depth reservoirs. The results showed prediction rates ranging from 86% to 92% compared to the analytical solutions. Among these six regression models, random forest outperformed the others with the most stable performance in both the testing phase and case studies with the R-squared value and root mean square error of 0.97 and 4.8, respectively. The proposed screening tool can be further applied to provide recommendations on the feasibility of CO2 EOR in mid- and deep-depth reservoirs. For CO2 MMP predictions, MMP is one of the most important parameters for designing CO2 EOR and associated storage in depleted oil reservoirs. The injection gas stream often contains a certain concentration of impurities such as N2, H2S, CH4, etc. depending on the source of CO2. These impurities have different effects on CO2 MMP but there is a lack of widely accepted approaches to account for these effects on MMP calculation. In this study, a series of activities were conducted to develop a machine learning-based methodology for determining MMP for CO2 with various impurities. A database containing 234 CO2 MMP test results was built based on the reported experimental measurements in the public domain. The database was then subgrouped by three specific criteria: CO2 concentration in the injection gas, type of impurities in the injection gas, and heavier hydrocarbon content in the oil. This subgrouping was essential to capture the impact of different factors on CO2 MMP. An ensemble machine learning (ML) approach with seven ML models, including Random Forest, Adaptive Boosting, Light Gradient Boosting Machine, Extreme Gradient Boosting, Stacking, Artificial Neural Network, and Voting Regressor, was employed to calculate MMP based on the subgrouped database. The hyperparameters of these ML models were optimized by the grid search technique to minimize the relative errors between calculated and measured MMP values. The performance of each algorithm was assessed using three regression metrics: average absolute relative error (AARE), R-squared score (R2), and root mean square error (RMSE). All of these metrics exhibited satisfactory values for the optimized ML models, the average values of R², RMSE, and AARE were 0.962, 1.571, and 4.55%, respectively, for the three subgroups, indicating a high accuracy of MMP calculations using the optimized ML models. The XGBoost model emerged as the top performer across the three metrics, with an R2 of 0.979, an AARE of 2.835%, and an RMSE of 1.183 for a dataset with 190 cases. The overall high level of accuracy confirmed the reliability of these ML models in calculating MMP for CO2 with different impurities, as well as the importance of optimization in the modeling process.

Download

Share

COinS