Robust optimization of fully coupled geothermal reservoir and power plant system based on deep learning

In geothermal recovery, the reservoir production temperature decline can affect the power plant’s efficiency and electricity output. Therefore, the coupling between the reservoir and the power plant is crucial for accurate estimation of the power plant’s performance. Current studies couple the numerical simulation models of the reservoir and the power plant developed based on physics. However, such simulations are usually computationally inefficient when performing predictions, thus becoming bottleneck in both forward and inverse modeling tasks. Therefore, we aim to accelerate the forward and inverse modeling of the coupled model by replacing the numerical simulation models with deep learning-based surrogate models. In this study, we first develop surrogate models of the geothermal reservoir and power plant, and further couple them into as an integrated forward model through heat source conditions. Further, a multi-objective optimizer combining the forward model is applied to optimize the coupled system. Surrogate models of the reservoir and power plant can predict the wellhead production temperature and electricity with mean relative errors of 0.49% and 1.67% while achieving CPU speedup at 6.92×104 and 1.77×105 times compared to physics simulators, respectively. Besides, the surrogate-based optimization is 6.05×105 times faster than the simulation-based one. The results demonstrate much higher computational efficiency of our coupled model in both the forward and inverse modeling with negligible trade-off in accuracy, as compared to the current physics-based coupled simulation models. This workflow significantly accelerates the procedures of feasibility assessments of geothermal projects as well as the decision making of the geothermal reservoir and the power plant.