Gradient-boosted spatiotemporal neural network for simulating underground hydrogen storage in aquifers

Abstract

Underground hydrogen storage (UHS) in aquifers has emerged as a viable solution to address the seasonal mismatch between supply and demand in renewable energy. Numerical simulation of the UHS serves as a crucial foundation for optimizing storage operations and conducting system risk assessments. However, numerical simulation methods employed for these purposes often demand substantial data, making data collection challenging and computationally expensive, especially in scenarios involving the coupling of multiple physical fields. Deep learning serves as an effective tool in resolving this challenge. Here, we proposed a spatiotemporal neural network architecture with gradient enhancement, denoted as gradient-boosted spatiotemporal neural network (GSTNN) and its variant GSTNN-s. The GSTNN combines a convolutional neural network (CNN), a long short-term memory network (LSTM), and an autoencoder architecture. To incorporate physical constraints into the network, the spatiotemporal gradient operators from the gas-water seepage and gas convection-diffusion equations are introduced as regularization terms, imposing physics-informed constraints on the training process in both temporal and spatial dimensions. In predicting the multiphase flow of UHS in both homogeneous and heterogeneous formations, GSTNN outperforms CNN and CNN-LSTM in terms of the accuracy of pressure, saturation and H₂ concentration fields. In terms of predicting UHS in formations with different permeabilities and porosities, GSTNN-s demonstrates improved performance as well. The proposed GSTNN architecture is promising in improving the efficiency of UHS numerical simulation, and has great potential to be applied for optimizing UHS operations in the future.