Uncertainty quantification of CO2 storage using Bayesian model averaging and polynomial chaos expansion

Abstract

Carbon sequestration in oil reservoirs and deep saline formations may be accomplished by many different trapping mechanisms. Use of CO2 for Enhanced Oil Recovery (CO2-EOR) leads to CO2 in three distinct phases, including CO2 dissolved in oil, CO2 dissolved in water (aqueous) and/or supercritical CO2. We evaluated the total amount of stored CO2 as well as the amount of CO2 in each phase for an active CO2-EOR site in western Texas.Three-phase relative permeability and associated hysteresis are two major sources of model uncertainty. Both are difficult to measure and are usually predicted by interpolation models. Instead of using arbitrary interpolation models, we used a model-averaging method based on Bayesian inference to estimate integrated model uncertainty for 12 alternative models. Moreover, given the uncertainty of intrinsic rock properties including permeability and porosity, uncertainty quantification (UQ) of these parameters is also necessary for forecasting CO2 storage capacity. Thus, results of this study provide uncertainty based on both model and data uncertainty.Conventional Monte Carlo methods with geocellular simulations are computationally expensive. We applied a Polynomial Chaos Expansion (PCE) methodology instead, to reduce computational requirements while minimizing the loss of accuracy. Geostatistical techniques were applied to generate stochastic realizations based on well logs and seismic survey data.For the Texas case study, we developed a systematic approach to quantify overall uncertainty, including both model uncertainty and parameter uncertainty. The approach was applied to forecast results at three important time steps, the end of the 30-year CO2-EOR injection period, the end of the 20-year post EOR CO2 injection period, and the end of the 50-year monitoring period. Results suggest that oil solubility dominates CO2 trapping and aqueous solubility has the least relative importance with respect to trapping (storage). Predictions of model averaging preserved the general pattern and captured differences among alternative models. CO2 storage of the reference model was within one standard deviation of predictions of model averaging. Estimated relative error between forecasted CO2 storage and the reference model are 0.8%, 7.4%, and 6.1% at three selected time steps. By the end of simulation, estimated CO2 storage in five selected layers in oil, supercritical, and aqueous phases are 3.4 ± 0.3 million tonnes, 2.0 ± 0.25 million tonnes, and 0.24 ± 0.04 million tonnes, respectively.