Assisted upscaling of miscible CO¬2-enhanced oil recovery floods using an artificial-neural-network-based optimisation algorithm

Abstract

The fine-scale compositional simulations required to accurately model miscible CO2 flooding are unrealistic and highly computationally expensive, and upscaling procedures are required to approximate the behaviour of these fine-scale grids on more realistic coarse-scale models. These procedures include the pseudoisation of relative permeabilities which ensures the matching of large-scale effects such as the volumetric fluxes of the phases and the use of transport coefficients to better represent small-scale interactions such as the time-dependent flux of the components within the hydrocarbon phases (molecular diffusion). Most times, a mismatch between the phase fluxes of the integrated fine-scale and that of the coarse-scale is observed. The accuracy of the upscaling results can be improved by tuning some of the derived upscaled coarse-scale quantities such as the alpha factors, absolute permeability or the relative permeabilities endpoints. This process is generally computationally expensive and can be treated as a reservoir history matching problem. In this study, we present a framework to represent the dynamics of small-scale molecular diffusion, and macro-scale heterogeneity-induced channelling associated with miscible CO2 displacements on upscaled coarser grid reservoir models. The approach applied was based on the pseudoisation of relative permeability and transport coefficients and was applied in two Society of Petroleum Engineers (SPE) benchmark reservoir models. Our results showed that the use of efficiently tuned transport coefficients led to better results so that the derived pseudo-relative permeability functions can be neglected. To reduce the associated computational expense, we proposed a novel methodology for upscaling miscible floods where a neural-network-based genetic-algorithm assisted upscaling procedure was applied. The optimisation algorithm was applied to reduce the error between the predictions of the upscaled models and a data-driven approximation model significantly reduced the computational expense associated with the assisted tuning procedure. In summary, the framework presented in this study adequately represented the small- and large-scale behaviour associated with the miscible displacements on upscaled coarse-scale reservoir models.