Surrogate-Assisted Evolutionary Optimization of CO₂-ESGR and Storage

CO2-enhanced shale gas recovery (CO2-ESGR) could efficiently recover gas with synchronous carbon sequestration, which is safer and more reliable than that in conventional reservoirs due to adsorption. This study developed an optimization scheme for the CO2-ESGR by integrating stochastic algorithms with an artificial neural network (ANN) surrogate model derived from compositional modeling. First, a shale gas reservoir sector model endowed with characteristics from the Sichuan Basin and hydraulically fractured horizontal wells was built. Then, sensitivity analysis was conducted for five critical parameters, namely, gas injection rate, bottom-hole pressure, fracture spacing, half-length, and conductivity, which are predictors for training surrogate models. Aided by the Latin hypercube sampling (LHS) method, 206 parameter combinations were designed, and CH4 recovery and CO2 storage were obtained by compositional simulation. These data sets with the objectives of CH4 recovery and CO2 storage were then partitioned to train and test the ANN models. A correlation coefficient of 0.99 was achieved for both training and testing, indicating high prediction accuracy and stability. Inputting new parameter sets demonstrated that the ANN surrogate model is comparable with a compositional simulator in terms of prediction accuracy with about 6500 times of speedup. Based on the obtained surrogate model, the five predictors were simultaneously optimized using genetic algorithm and particle swarm optimization, achieving a higher recovery factor than sensitivity analysis and LHS with just around1/220 of the computational time of a single compositional simulation. This study serves as a useful reference for the proactive design and further implementation of CO2-ESGR and concurrent storage.