Enabling Clean Energy Resilience with Machine Learning-Empowered Underground Hydrogen Storage

TL;DR

The paper addresses the challenge of integrating data-driven surrogates to enable clean energy resilience via underground hydrogen storage (UHS) by reducing the computational burden of high-fidelity reservoir simulations. It proposes a roadmap for machine learning in UHS, including leveraging GCS surrogate techniques, auto-regressive time evolution, and high-resolution predictive capabilities, while outlining the data-generation strategy and associated challenges. The authors present 1000 two-dimensional UHS simulations with cyclic injection/withdrawal to illustrate the dataset and discuss how ML can predict spatial fields and scalar performance metrics, with preliminary results suggesting auto-regressive models can extrapolate over time though error accumulation remains a concern. The work aims to enable scalable, real-time risk assessment and optimization for UHS operations, promoting large-scale deployment of clean energy storage solutions.

Abstract

To address the urgent challenge of climate change, there is a critical need to transition away from fossil fuels towards sustainable energy systems, with renewable energy sources playing a pivotal role. However, the inherent variability of renewable energy, without effective storage solutions, often leads to imbalances between energy supply and demand. Underground Hydrogen Storage (UHS) emerges as a promising long-term storage solution to bridge this gap, yet its widespread implementation is impeded by the high computational costs associated with high fidelity UHS simulations. This paper introduces UHS from a data-driven perspective and outlines a roadmap for integrating machine learning into UHS, thereby facilitating the large-scale deployment of UHS.

Figures (4)

• Figure 1: Temporal evolution of the spatial distribution of H$_2$ saturation and reservoir pressure in a 2D UHS simulation. H$_2$ is injected and withdrawn from a central well in a depleted gas reservoir over ten annual cycles comprising a 6-month injection stage followed by a 6-month withdrawal stage. The initial two subfigures (top row) depict the heterogeneous porosity and permeability of the geological formation. Subsequent figures show the H$_2$ saturation and pressure distributions at various time points during UHS operations. 'Early' of a stage refers to two months after its onset, and 'end' of a stage to six months. Porosity is dimensionless; permeability is in millidarcys ($10^{-15} m^2$); H$_2$ saturation is dimensionless; pressure is in bars ($10^5$ Pa).
• Figure 2: Comparison between static and auto-regressive models. Five simulations were randomly selected from the validation set for the comparison. There are 42 time steps in each plot, where one time step is equivalent to 2 months. Figures \ref{['fig:hgasauto']}, \ref{['fig:hgasstatic']}, \ref{['fig:hgasautostatic']} concern H$_2$ saturation models, while Figures \ref{['fig:pressureauto']}, \ref{['fig:pressurestatic']}, \ref{['fig:pressureautostatic']} concern pressure models. Figures \ref{['fig:hgasautostatic']}, \ref{['fig:pressureautostatic']} plot time steps against the MAE of the auto-regressive model minus the MAE of the static model.
• Figure 3: Demonstration of extrapolation in time of auto-regressive models. Figure \ref{['fig:hgastime']} demonstrates this for the H$_2$ model, while Figure \ref{['fig:pressuretime']} demonstrates it for the pressure model.
• Figure 4: A 2D UHS simulation that captures the H$_2$ plume and pressure as time progresses. H$_2$ is injected and withdrawn in cycles of 6 months each, where the simulator further subdivides each 6 month stage into 3 parts. By start of a stage, we mean 2 months after it has started as this is the earliest when the simulator saves data. By end of a stage, we mean 6 months after it has started. The simulation is carried out in a different type of mapping than in Figure \ref{['fig:data_example']}. This figure illustrates the preferential paths that the hydrogen plume takes, making the plume go far from the well. Coarsening the image in the outer parts of the plume would delete the preferential paths of the plume. This can potentially obstruct the model learning the true behavior of the plume in these kinds of geological formations.