Improving the Predictive Power of Equation of State Models for Fluid Systems Undergoing CO2 Injection

Abstract

Abstract Phase equilibrium calculations require experimental lab data to constrain component properties in an equation of state (EOS) model. These thermodynamics-based models generally perform well when it comes to predicting conventional PVT experiments but often fall short when it comes to predicting gas injection experiments, particularly for CO2 injection. We therefore seek to develop methods that can help provide a good initial estimate of the swelling curve, in cases where laboratory data are not available. Our company PVT database compromises more than 2,200 PVT studies, which enables us to pursue three different avenues for predicting the CO2 swelling curve. The first method relies on a machine-learning algorithm, which takes fluid composition and temperature as input. In general, we find that this solution does not preserve monotonicity of the pressure-dependent properties and it extrapolates poorly outside the parameter space used for training. As an example, it fails to predict the first-contact miscible pressure defined as the maximum pressure on the swelling curve. The second option involved correlating swelling pressure, swelling factor and swelling density as a function of the amount of injected gas. We find that all three curves are well-represented by a parabolic expression and we were able to correlate the coefficients as a function methane content in the reservoir fluid only. The resulting model predicts saturation pressure, swelling factor, and density of the swollen mixtures with an absolute average deviation of 4.8%, 2.3% and 1.7%, respectively, which is an excellent starting point for tuning an EOS model for EOR screening studies until experimental data becomes available. The third strategy involved tuning a separate EOS model to each of the 34 CO2 swelling studies and then attempt to correlate the EOS component properties. We compare the values of the tuned pseudo-component properties against some standard correlations such as Pedersen, Kesler-Lee, Riazi-Daubert and others. We find that the Pedersen correlations for critical pressure, critical temperature and acentric factor provide a more accurate initial guess than the other correlations tested. However, we observed that the tuned solution depended to some extent on the initial guess. We find that for our fluid systems, the default values for the critical volume of the pseudo-components need to be reduced by 15% to better predict the viscosity using the LBC model. Despite the slightly improved property estimation, we did not manage to find a clear trend for the binary interaction coefficient between CO2 and the plus fraction. Therefore, we would recommend predicting the CO2 swelling curve with the set of parabolic correlations.