Predicting ultimate hydrogen production and residual volume during cyclic underground hydrogen storage in porous media using machine learning

Underground hydrogen storage (UHS) is a promising solution for balancing renewable energy supply by storing excess energy as molecular hydrogen (H2) for later use. Large-scale UHS in geological formations requires reservoir simulations of cyclic loading scenarios to optimize storage operations. Such simulations must account for complex physical phenomena, including reservoir flow dynamics, trapping mechanisms, and reactions with minerals and bacteria. Fast H2 injection and production operations generate extensive datasets, necessitating high computational power to capture intrinsic temporal and spatial variations for reliable predictions. Machine learning (ML) offers a viable approach to addressing these computational challenges while improving operational efficiency and ultimately cost reduction. This study explores how ML models trained on simulated UHS data in porous media can predict ultimate hydrogen production and trapped hydrogen due to factors such as hysteresis, dissolution, and capillary pressure. The OPM Flow reservoir simulator was used to gen- erate cyclic field data, which was used to train time-series neural network (NN) models. The models were fine tuned through hyperparameter optimization and cross-validated before deployment. Results indicated that ML models effectively predicted hydrogen production and residual storage with high accuracy, as measured by mean squared error (MSE) and mean absolute error (MAE). A key finding was that ML-based predictions significantly reduced computational time. In one reservoir-scale case, ML reduced computation time by 6773% compared to OPM Flow simulations on a four-layer reservoir model. Even greater accuracy and efficiency were achieved using a one-layer model with a horizontal well under a complex cyclic schedule. By capturing complex physical uncertainties, ML provides a powerful complement to traditional reservoir simulation methods. This research contributes to UHS development by demonstrating the potential of ML for optimizing storage operations and enabling fast, data-driven decision-making.