Prediction of Optimal Production Time during Underground CH4 Storage with Cushion CO2 Using Reservoir Simulations and Artificial Neural Networks

Injection of carbon dioxide (CO2) in aquifers for underground natural gas storage (UNGS) can improve the operating efficiency of the storage facility during gas injection and production and later serve as a solution for permanent CO2 storage. However, mixing of the gases can lead to undesired CO2 production during seasonal withdrawal periods, which requires costly treatment of the produced gas. This work quantifies the optimal production time (until the well stream exceeds 1% mole fraction of CO2) in idealized, CO2-filled reservoirs subjected to an injection and production cycle of methane (CH4) in a single well. A set of 1200 compositional reservoir simulations with systematic variation of reservoir temperature, porosity, permeability, height, and injection time shows that reservoir height and permeability have the most significant impact on the production time. The generated data enable the development of artificial neural networks (ANN) that describe relations between the varied input parameters and the optimal production time with excellent accuracy. Reducing the amount of data sets in the ANN training from 1050 (87.5%) to 600 (50%) and augmenting the ANN output with other parameters, like maximal reservoir pressure, average CO2 mole fraction, and well-block pressure at the end of production, only marginally reduces the accuracy of the data-driven models. In all cases, the developed ANNs exhibit an RMSE less than 0.02 and an R2 score above 0.99. Hence, we conclude that trained and validated ANNs are useful tools to determine relations between important parameters in UNGS operations where CO2 is used as cushion gas, with the aim at reaching higher CH4 production time and larger amounts of delivered gas with minimal CO2 production.将二氧化碳（CO₂）注入含水层用于地下天然气储存（underground natural gas storage, UNGS），可提升注采作业阶段储库的运行效率，后续还可作为CO₂永久封存的可行方案。然而，气体混合会导致季节性采气周期中出现非预期的CO₂产出，这需要为采出气体承担高额的处理成本。本研究针对充满CO₂的理想化储层，在单井注入与采出甲烷（CH₄）的循环工况下，量化了最优采气时长（直至井筒气流中CO₂摩尔分数超过1%）。通过系统改变储层温度、孔隙度、渗透率、储层厚度与注入时长，本研究开展了1200组组分油藏模拟（compositional reservoir simulations），结果表明储层厚度与渗透率对采气时长的影响最为显著。所生成的数据集可用于构建人工神经网络（artificial neural networks, ANN），能够以极高精度描述可变输入参数与最优采气时长之间的关联关系。将ANN训练所需的数据集规模从1050组（占总数据集的87.5%）缩减至600组（占比50%），并将最大储层压力、平均CO₂摩尔分数以及采气结束时刻的井块压力等额外参数纳入ANN输出，仅会小幅降低数据驱动模型的精度。在所有实验场景中，所构建的ANN的均方根误差（RMSE）均小于0.02，决定系数（R² score）均高于0.99。综上，经训练与验证的ANN可作为有效工具，用于厘清以CO₂作为垫层气（cushion gas）的地下天然气储存作业中各关键参数间的关联关系，以期实现更长的甲烷采气时长、更高的输气量，同时最大限度减少CO₂产出。