Reinforcement learning based optimal dynamic policy determination for natural gas hydrate reservoir exploitation

The commercial exploitation of natural gas hydrate reservoirs requires scientific policies to reduce the exploitation cost and improve production efficiency. A multineural network composite model for NGH reservoir exploitation is designed and coupled with the secondary-developed T + H code in the PYTHON environment. Under the preconditions of realizing multiobjective optimization in terms of efficiency, safety, and economy, the model achieves the optimal dynamic production policy and quantitatively presents the advantages and disadvantages of different exploitation policies in each exploitation state. In the initial exploitation period, high-temperature and increasingly low-rate exploitation policies achieve the strongest production performance and energy efficiency. The huff and puff heat injection method can effectively reduce the permeability drop caused by secondary hydrate formation. In the middle exploitation period, the injection density of thermal energy controls the hydrate exploitation efficiency, the huff and puff method is not applicable, and more intensive heat injection results in higher exploitation efficiency and less water production. In the posterior exploitation period, the methane recovery rate does not depend on heat injection, and more energy supply can only achieve a lower energy efficiency return. The proposed methodology provides a feasible means of NGH reservoir exploitation policy optimization and analysis. © 2022 Elsevier B.V.