The Application of Machine Learning to Predict CO2 Enhanced Oil Recovery:

A Case Study for the EOR33 Project

Abstract

As Trinidad and Tobago (T&T) seeks to manage its carbon emissions while boosting oil production via different EOR techniques, it will be consequential to utilize machine learning techniques to enhance the predictability of reservoir performance. Machine learning presents the opportunity to develop proxy models which can be a cost effective and can be used for making fast practical and more accurate decisions for field development.

This paper presents a machine learning approach to predicting cumulative oil production when CO₂ is injected into an oil reservoir for enhancing oil recovery. A commercial simulator CMG’s CMOST-AI was used to generate the data set required for analysis by performing a sensitivity analysis on the parameters that had uncertainty to determine their impact on the cumulative oil production when CO₂ is injected in the EOR 33 project located in South Trinidad. Supervised Learning using Linear regression, polynomial regression and Random Forest Data Analysis techniques were used analyze data generated by the commercial software so that the rate and accuracy of future predictions would be greatly enhanced.

The linear regression and polynomial regression showed and R² of 0.988 and 0.999 and a MAPE of 0.0119 and 0.0031 respectively indicating that the proxy models generated can be used for reservoirs with similar parameters for predicting oil recovery when CO₂ is injected. The Random Forest Model can be used for prediction with although it has a much lower R² of 0.91 and a MAPE of 0.0339.