气藏深度学习渗流代理建模关键技术与发展展望OA

The deep learning-based surrogate models for gas-reservoir seepage have been widely applied in forward simulation,inverse problem solving and production optimization of gas reservoirs because of their powerful nonlinear mapping capabilities.However,there is a lack of systematic review and comparative summary on different types of gas reservoirs.In this paper,the research status of the deep learning-based surrogate modeling for gas-reservoir seepage is systematically reviewed,the research progresses and technical bottlenecks are mainly illustrated in terms of surrogate model types,training frameworks and engineering applications,and finally the future technical research directions and development routes are proposed.The following results are obtained.First,the deep learning-based surrogate models for gas-reservoir seepage encompass a variety of fundamental networks,such as fully connected neural network,graph neural network,convolutional neural network,U-shape neural network,long short-term memory neural network,residual neural network,and generative adversarial neural network.They can be classified into three types,namely data-driven,physics-driven and data-physics-driven models.Second,the training of surrogate models includes two modes,i.e.,offline training and online training.Offline training relies on a large number of samples to guarantee model accuracy,while model updating strategy is crucial for the accuracy and speed of online training.Third,from the perspective of engineering application,surrogate models have been deployed in the scenarios like conventional and unconventional gas reservoirs,naturally fractured reservoirs,CO2 geological storage,and underground hydrogen storage.The introduction of data assimilation algorithms(e.g.EnKF and ES-MDA)and optimization algorithms(e.g.genetic algorithm and particle swarm optimization)improves the efficiency of automatic history matching and development plan optimization.Fourth,current surrogate models still have limitations in the simulation of multi-media reservoirs,unconventional gas reservoirs with complex hydraulic fracture network,and multi-phase multi-component flow systems.Future research should focus on the adaptability evaluation of surrogate models for different reservoir types,accelerating the training of surrogate models for complex gas seepage,surrogate modeling for multi-field coupled seepage,and efficient embedding of physical constraints.In conclusion,it is expected that with the improvement of the related theories and technologies,the deep learning-based surrogate modeling for gas-reservoir seepage can significantly improve the efficiency of numerical simulation of gas reservoirs and promote the intelligent construction of oil and gas fields.