ScholarWorks@Y-Scholar Hub

초록

During geologic carbon storage (GCS), Temperature Swing Injection (TSI), a method involving the periodic changes in injection temperature at the wellhead, has been proposed to enhance storage performance. Accurate characterization of TSI requires a coupled wellbore-reservoir model, but solving the coupled partial differential equations is computationally intensive. To overcome this challenge, this study presents an integrated framework that combines physics-based simulation for quantitatively evaluating the impact of TSI on CO2 storage with deep learning-based surrogate model. A comprehensive dataset was generated by parameterizing two representative TSI strategies including Gradual Temperature Swing (GTS) and Stepwise Temperature Swing (STS). These strategies were simulated for an approximately one-year CO2 injection period under various formation properties using the coupled wellbore-reservoir simulator (T2Well/ECO2N). The simulation results were then used to develop and train deep learning-based surrogate models that accurately reproduced simulation results (R-2 > 0.995). On average, TSI increased CO2 storage by 16.3-31.2 %. Permutation feature importance (PFI) and feature sensitivity (FS) analyses identified mean temperature, swing amplitude, injection pressure, and formation permeability as the most influential parameters. Finally, optimized injection strategies identified by the genetic algorithm (GA) increased stored CO2 mass up to 22.1-32.6 % compared to pre-optimization cases. This work provides a unified framework that integrates physics-based simulation, deep-learning surrogates, and optimization, offering a computationally efficient pathway for designing complex non-isothermal injection strategies that maximize CO2 storage efficiency.