A Novel Robust Optimization Framework Based on Surrogate Modeling for Underground Hydrogen Storage in Depleted Natural Gas Reservoirs

Abstract

Underground hydrogen storage (UHS) plays a vital role in global net-zero energy systems, enabling the storage of excess renewable energy for later use. However, physical reservoir model-based optimization for UHS system design and operation is computationally expensive due to complex geological properties and well-operational controls. This study developed a novel efficient framework for UHS robust optimization to address this challenge, integrating advanced compositional reservoir simulation, accurate surrogate modeling, and robust optimization techniques. First, a base reservoir simulation model was developed to capture compositional fluid flow, hydrogen methanation reactions, gravity segregation, hysteresis, and capillary effects. To rapidly evaluate various well controls and reservoir configurations, CNN-BiLSTM-Attention models were trained as surrogate models using a comprehensive dataset generated from reservoir simulations. The convolutional neural network (CNN) transforms three-dimensional (3D) geological fields into one-dimensional (1D) vectors, effectively capturing spatial features. The bi-directional long short-term memory (BiLSTM) network learns the temporal evolution of the input features over time by processing them in both forward and backward directions. Subsequently, the attention mechanism enhances prediction accuracy by identifying and emphasizing the most significant features at critical time steps. The well-trained surrogate models were seamlessly integrated into the robust optimization framework based on the genetic algorithm, aiming to maximize the net present value (NPV) from UHS projects. The results demonstrate that the surrogate model exhibits satisfactory performance in the context of prediction accuracy, computational efficiency, and scalability. Notably, the newly developed framework based on surrogate models provides an approximate 4878 speedup compared to an approach relying solely on reservoir simulation, while maintaining comparable accuracy. Overall, the proposed framework offers a promising solution for UHS optimization, providing valuable insights for the design and management of sustainable energy infrastructure.

Keywords: Underground hydrogen storage, Robust optimization, CNN-LSTM-Attention architecture, Surrogate model