Incorporating phase behavior constraints in the multi-objective optimization of a warm vaporized solvent injection process

The warm vaporized solvent injection process has been proposed as a more environmentally friendly alternative to steam-based technologies for bitumen recovery. The process typically involves injecting heated solvent vapor into a horizontal injector; the solvent condenses and dissolves into bitumen, while the diluted oleic phase would flow towards a horizontal producer. Despite the promising results reported from several pilot projects near Fort McKay, Alberta, successful commercial-scale extraction is costly and would require a detailed optimization of the pertinent design variables. The main challenge is that this is a multi-objective optimization (MOO) problem, which aims to balance the trade-offs between conflicting performance objectives while honoring the various operational constraints. In this study, a systematic workflow is formulated to optimize these multiple conflicting performance objectives considering phase behavior constraints. A 2D synthetic model based on typical Athabasca oil sands properties is constructed to simulate the warm vaporized solvent process. The addition of non-condensable gas (methane) into the solvent (propane) is examined. The resultant changes in thermodynamic properties and equilibrium phase behavior are considered in determining the practical limits of the decision variables (e.g., bottom-hole injection pressure and temperature). The objective functions, including oil recovery factor, solvent retained-to-oil ratio, and energy consumption, are defined, and a factorial experimental design is employed to identify a subset of decision variables that exhibit minimal redundancy internally and create the dataset for proxy model development. To reduce the computational costs associated with reservoir simulations, proxy models, e.g., the artificial neural network (ANN), is developed and applied. Finally, a Pareto-based MOO scheme is implemented to estimate the optimal decision variables. Despite the higher front-end loading requirement of the ANN proxy modeling, the MOO with proxy modeling still requires significantly less execution/running time as compared to a MOO with traditional flow simulation (e.g., a 97% reduction in CPU time). This reduced running time is important for alleviating the computational load when evaluating the objective functions during the optimization process. More importantly, this optimization scheme is capable of identifying a set of optimal decision variables. This work presents a practical workflow fo