AI-based composition model for energy utilization efficiency optimization of gas hydrate recovery by combined method of depressurization and thermal stimulation

The combination of depressurization and thermal stimulation is one of the most promising techniques for producing gas from natural gas hydrate reservoirs. The energy utilization efficiency is the core factor that determines whether the technology can be better applied to engineering practice. In contrast to previous works, this paper proposes an AI-based composition model, which realizes the dynamic optimization of the energy utilization efficiency based on the learning and evaluation of different thermal stimulation policies. As shown by the optimization results, the proposed model achieved good performance in both the hydrate recovery target and the economic target coordination by adjusting the coefficients of the optimization control policy function. Based on a systematic assessment of the energy utilization efficiency of different thermal stimulation policies, the model increased the energy production-injection ratio by 4.5 times in Scenario I, and the CH4 recovery efficiency was increased by 1.6 times in Scenario II. The learning results of the AI model also showed that the reactor scale hydrate recovery had a size effect, such that the later the thermal energy injection was, the higher the energy efficiency that could be obtained. Moreover, through the model learning, the quantitative relationship between each thermal stimulation policy and its developmentally changing energy utilization efficiency was established. The results showed that in the initial hydrate recovery period, thermal stimulation policies with a high injection temperature and low injection rate achieved greater energy utilization efficiency. As recovery progressed, gradually lowering the injection temperature and increasing the injection rate gave the policy a higher energy utilization efficiency. The proposed methodology provides a feasible approach for dynamic thermal stimulation parameter optimization and improves the hydrate recovery economy. [Display omitted] •Combines supervised and deep reinforcement learning for gas hydrates recovery.•Thermal stimulation parameters optimization during the dynamic recovery process.•A novel model frame design reduces the training cost in high-dimensional optimization.•Agent established the mapping between recovery policies and energy efficiencies.•Minimize the injected thermal energy loss while achieving the recovery target.