Estimation of Crude Oil Minimum Miscibility Pressure During CO2 Flooding: A Comparative Study of Random Forest, Support Vector Machine, and Back Propagation Neural Network

When CO2 is injected into hydrocarbon reservoirs for enhanced oil recovery (EOR), the minimum miscibility pressure (MMP), which defines as the lowest pressure of generating a miscible phase, is an important parameter to determine whether the displacement process is miscible flooding or not at reservoir conditions. Compared with the time-consuming and complicated measurement of MMP in the laboratory, an empirical estimation is a better alternative for engineering design of CO2 flooding, especially in the feasibility study stage. Machine learning based intelligent model exhibits superiority in convenient and rapid access to the precise estimations of MMP. In this paper, on the basis of experimental information from previous literature, three intelligent models, i.e., random forest (RF), support vector machine (SVM), and back propagation neural network (BPNN), are presented to estimate the CO2 MMP with various crudes. Through multivariate parametric regression (MPR) method, six main influencing factors, i.e., the mole fraction of two injected gas components (C-2-C-5 , H2S), Tcm, Tr, MWC5+ and Vol/Int, are selected as input variables. Our results show that all three intelligent models are able to exploit intrinsic dependencies between MMP and these input variables. However, different intelligent models have their own features: (1) The RF model with strong robustness and generalization capability exhibits the best performance in total database among these three models. (2) The BPNN model with artificially optimized network structures is potentially comparable with other two models although the accuracy of BPNN is vulnerable to the initialized network parameters. (3) The SVM model occupies an obvious advantage in coping with sparse samples whose MMP is over 30 MPa.