Various studies have been conducted to optimize oil production in unconventional reservoirs using gas injection processes. Some studies determined the optimal design of hydraulic fractures through horizontal wells while others optimized the placement of horizontal wells to maximize oil recovery. In this study, an integrated optimization workflow combining Particle Swarm and Non-parametric Proxy Metamodels was adopted to optimize oil production using CO2-EOR in shale oil reservoirs.
A pseudo-component black oil reservoir model was considered to simulate CO2 flooding in shale oil reservoirs and to predict future reservoir performance over a 30-year prediction period. The cyclic CO2 flooding optimization procedure consisted of 3 cyclic operational decision factors and 5 well constraints for 5 pairs of horizontal injector and producer wells with 8 fractures each. The cyclic injection factors included injection, soaking, and production durations over the prediction period. Minimum bottom hole pressure, maximum oil production rate, and water cut were optimized for the production wells, and maximum bottom hole injection pressure and maximum gas injection rate were optimized for the injection wells. An integrated optimization approach conducted using Particle Swarm Optimization (PSO) and Proxy Metamodeling was integrated to find the optimal level for each of the 8 factors. PSO was adopted to create a search-space swarm of candidate solutions (particles) considering the range of each operational factor. These particles were then evaluated by the reservoir simulator to calculate the cumulative oil production by the end of the prediction period. To reach the optimal solution, 100 candidate solutions were created as training experiments with 4 successive iterations of approximately 20 experiments each. The optimal solution increased oil production by 322,675 surface barrels. Next, a 2nd order polynomial regression proxy model was constructed to metamodel the large reservoir simulator. The performance of the polynomial proxy model was validated via a comparison to non-parametric algorithms of Multivariate Adaptive Regression Splines. A cross-validation was considered prior to building the proxy models by sampling and subdividing the entire dataset into two subsets: 70% training for modeling and 30% testing for prediction. The testing subset and the entire dataset were compared to the three proxy models by computing the Root Mean Square Prediction Error and the Adjusted R-square for each. It was concluded that polynomial regression is the best metamodel, followed by the MARS algorithm, to produce a simplified alternative metamodel for the reservoir simulator to evaluate the cyclic CO2 flooding in shale oil reservoirs. The most influential operational factors were also identified for their effects on the CO2-EOR process performance in shale reservoirs.