Predicting minimum miscible pressure in pure CO2 flooding using machine learning: Method comparison and sensitivity analysis

Abstract

CO2 injection for enhanced oil recovery (EOR) is widely recognized as an efficient technique for carbon capture, utilization, and storage (CCUS). This operation has a significant impact on various technical parameters, emphasizing the need to carefully consider and select the optimum approach. Among these factors, the minimum miscible pressure (MMP) plays a crucial role in determining the effectiveness and performance of CO2 injection. Therefore, this study aims to assess the reliability of machine learning (ML) in predicting the MMP of pure CO2 and examine the influence of different independent parameters. To achieve this, five ML methods were employed to predict the pure CO2 MMP, and the results were compared to statistical evaluations based on empirical correlations. In addition, three types of data with different functional input parameters were used in this research. Two types of data were obtained from existing literature, while the third category was collected from the thesis and PVT reports for specific Iraqi oil fields. The ML models were constructed by splitting the dataset into 20% for testing and 80% for training using Python programming. The significance of this study lies in its ability to identify the most efficient approach for forecasting MMP. The results of this work revealed that the K-nearest neighbors (KNN) model indicated the best statistical evaluation among the ML learning algorithms for two types of data (2) and (3) in predicting the MMP for pure CO2 flooding. This was evidenced by the lowest mean square error and the highest coefficient of determination. Additionally, the findings indicated that the support vector regression (SVR) method is an effective technique for smaller datasets. Moreover, the sensitivity analysis and assessment of the relative impacts of various input parameters revealed that the prediction of MMP is most sensitive to the composition of the injected gas and temperature, accounting for 46% and 28.5% of the variation, respectively. Finally, the presented ML models indicate exceptional accuracy, speed, adaptability in handling diverse conditions, and cost-effectiveness when compared to conventional approaches. These results verify the ability of ML models to provide high-quality predictions.