This paper addresses the challenge of optimizing subsurface hydrogen storage in porous media, a crucial component for advancing energy transition. The multifaceted nature of this challenge stems from the complex physics governing the process, coupled with operational limitations, and subsurface geological uncertainties. We apply a stochastic gradient-based optimization method with novel deep-learning acceleration components, tailored to maximize the efficiency of hydrogen storage by tuning well locations while honoring operational constraints.
The key objective of optimization is to maximize the amount of recoverable hydrogen while maintaining operational constraints. We adopt a robust optimization approach that maximizes the mean objective function over a set of realizations representing subsurface uncertainty. The objective function, defined as the hydrogen deliverability index, is calculated using a compositional reservoir simulator with high-resolution grids to minimize numerical dispersion. Our approach leverages a deep-learning-accelerated-gradient (DLAG) method alongside these simulations. This method is applied to the Brugge field case study, which is divided into two distinct optimization scenarios. In the first case, we evaluate the effectiveness of the optimization method with only one subsurface realization, optimizing the placement of eight storage wells and comparing outcomes with and without the application of DLAG. In the second case, we extend the analysis to include five different subsurface realizations and impose specific location constraints on each of the storage wells to optimize their placement. In the first case, the application of the DLAG method showed a clear advantage over the non-DLAG approach, resulting in faster convergence. The optimization of hydrogen storage well locations in the Brugge field model yielded notable improvements in storage efficiency, demonstrating the practicality and effectiveness of our approach.
