LSTM-based proxy model combined with wellbore-reservoir coupling simulations for predicting multi-dimensional state parameters in depleted gas reservoirs

Abstract

Although most CO₂ geologic storage projects focus on deep saline aquifers, depleted gas reservoirs are a more economical option as a potential site. However, due to the extremely low initial pressure, the injection of high-pressure supercritical CO₂ into the reservoir can result in dramatic changes in CO₂ properties, which may affect the well head pressure and bottom hole pressure. Aside from the injection rate, the injection of supercritical CO₂ at different temperature and pressure has varying degrees of impact on reservoir pressure. In order to design the optimal injection condition, a deep learning proxy model, combining wellbore-reservoir coupling numerous simulations, is proposed to quickly interrogate status response of wellbores and reservoirs. Based on 567 simulation cases of supercritical CO₂ injection into a deep depleted gas reservoir, the model uses the T2Well/ECO2N software to capture the time evolution of the pressure, temperature, and rate fields of wellbore and reservoirs, and is trained to get the optimal LSTM-based proxy network. Compared with simulation results, the proxy model predicts in less than 0.1s while ensuring an overall coefficient of determination (R²) of up to 99.9%. The maximum prediction errors of pressure, temperature, and rates at all times are also not more than 0.04, 0.02, and 0.08 for a single case, respectively. The assessment findings of the ultimate reservoir pressure based on the model show that injecting supercritical CO₂ under low initial pressure and high temperature is beneficial to the long-term safety of CO₂ sequestration engineering in depleted gas reservoirs.