Universal intelligent models for liquid density of CO2 + hydrocarbon mixtures

Abstract

Developing precise predictive methods for liquid density of CO2 + hydrocarbon mixtures is vital to model, optimize, and design of the enhanced oil recovery (EOR) and the CO2 storage processes. The CO2 injection into the oil reservoir can noticeably change the density of solution. In addition, forecasting this parameter over a wide range of conditions is extremely challenging due to the complicated behaviors of hydrocarbons. The earlier empirical models are only valid for specific hydrocarbons under very restricted conditions. Therefore, in this study, three robust machine learning based approaches were implemented to build robust and universal models based on an immense databank, comprising 17,168 experimental data extracted from 52 independent studies. The databank covered 46 varied mixtures at a broad range of conditions, including temperature, pressure and CO2 mole fraction. The statistical and the cumulative frequency analyzes showed that the radial basis function (RBF) model provides the best results among the all approaches, with an average absolute relative error of 1.61 % and a relative root mean squared error (RRMSE) of 4.29 %, respectively for the test data samples. It also provided favorable predictions for physical attitudes of the mixture density under different operating conditions, and properly described the “density crossover” phenomenon at high CO2 concentrations. The RBF-based model outperformed earlier empirical correlations in terms of precision, applicability, and generality. Overall, the machine learning-based models developed in this study are efficient and reliable methods to assist the designers and engineers involved with in CO2-EOR process.