Surrogate Modeling and Optimization of CO2-WAG Processes in Deepwater Low-Permeability Reservoirs

With the deepening implementation of China’s “Dual-Carbon” strategy, Water Alternating Gas (CO2-WAG) injection has emerged as a promising technological option for deepwater oil and gas development, owing to its dual advantages of enhancing oil recovery and promoting geological CO2 storage. However, influenced by depositional environments and hydrocarbon accumulation mechanisms, many deepwater reservoirs exhibit low porosity, low permeability, and strong heterogeneity. Under such conditions, conventional full-physics numerical simulation becomes computationally expensive and inefficient when applied to multi-scenario dynamic forecasting and real-time optimization. Moreover, existing CO2-WAG surrogate models often suffer from limited generalization capability under complex geological and operational conditions and usually assume fixed injection–production control schemes. To address these limitations, this study proposes a deep-learning-based surrogate model tailored for CO2-WAG development in deepwater low-permeability reservoirs. The proposed model is built upon an integrated CNN–LSTM–Attention architecture, in which geological conditions and variable well-control parameters of water and gas injectors across CO2-WAG cycles are jointly incorporated as input features. This enables the network to characterize the dynamic response of deepwater reservoirs under diverse operating conditions and to achieve high-accuracy, adaptive modeling of complex injection–production processes. Test results show that the model attains an average mean squared error (MSE) below 3 × 10−5 and an average relative error of only 2.2% for predictions of cumulative oil production, cumulative water production, cumulative water injection, and CO2 storage, with coefficients of determination (R2) for all target variables approaching 1. On this basis, the surrogate model is further coupled with an optimization algorithm, using net present value (NPV) and cumulative CO2 storage as dual objectives to perform well-control optimization for, respectively, economic benefit maximization and carbon sequestration enhancement. The results indicate that the optimal injection–production strategies obtained by the proposed framework differ from full-physics simulation results by less than 1%, while achieving a substantial improvement in computational efficiency. This study therefore establishes an intelligent optimization framework with high computational responsiveness and strong generalization capability for CO2-WAG development in deepwater low-permeability reservoirs, providing a new technological pathway for green and efficient deepwater oil and gas production.