Abstract
Large-scale CO2 injection for geo-sequestration in deep saline aquifers can significantly increase reservoir pressure, which, if not appropriately managed, can lead to potential environmental risk. Brine extraction from the aquifer has been proposed as a method to control the reservoir pressure and increase storage capacity. However, iterative optimization of the well controls for this scenario using high-resolution dynamic simulation models can be computationally expensive.
In this paper, we demonstrate the application of a so-called coarse–grid network model (CGNet) as a reduced-order model for efficient simulation and optimization of CO2 sequestration with brine extraction. As a proxy, CGNet is configured by aggressively coarsening the fine-scale grid and then tuning the parameters of the associated simulation graph (transmissibility, pore volumes, well indices, and relative permeability endpoints) by minimizing the mismatch of well-response data (rates, bottom-hole pressure) and saturation distribution from the fine-scale model. Calibration and optimization procedures are automated using gradient-based optimization methods that leverage automatic differentiation capabilities in the reservoir simulator in the same way backpropagation methods are used in training neural networks. Once calibrated, CGNet is employed for well-control optimization. Validation with the fine-scale model shows that CGNet closely matches the optimized net-present value (NPV). Numerical examples using the Johansen model, available as a public dataset, shows that the optimization can be accelerated up to seven times using CGNet compared with a fine-scale model. (Using a compiled language will likely result in significantly larger speedups as small models suffer from a disproportionately high computational overhead when executed in MATLAB.) This study implies that a reduced-order model such as CGNet can be a powerful data-driven tool for faster evaluation of CO2 geo-sequestration simulation, combined with proper reservoir monitoring program.
In this paper, we demonstrate the application of a so-called coarse–grid network model (CGNet) as a reduced-order model for efficient simulation and optimization of CO2 sequestration with brine extraction. As a proxy, CGNet is configured by aggressively coarsening the fine-scale grid and then tuning the parameters of the associated simulation graph (transmissibility, pore volumes, well indices, and relative permeability endpoints) by minimizing the mismatch of well-response data (rates, bottom-hole pressure) and saturation distribution from the fine-scale model. Calibration and optimization procedures are automated using gradient-based optimization methods that leverage automatic differentiation capabilities in the reservoir simulator in the same way backpropagation methods are used in training neural networks. Once calibrated, CGNet is employed for well-control optimization. Validation with the fine-scale model shows that CGNet closely matches the optimized net-present value (NPV). Numerical examples using the Johansen model, available as a public dataset, shows that the optimization can be accelerated up to seven times using CGNet compared with a fine-scale model. (Using a compiled language will likely result in significantly larger speedups as small models suffer from a disproportionately high computational overhead when executed in MATLAB.) This study implies that a reduced-order model such as CGNet can be a powerful data-driven tool for faster evaluation of CO2 geo-sequestration simulation, combined with proper reservoir monitoring program.