Machine Learning for Underground Natural Gas Storage Optimization

Abstract

Underground natural gas storage (UNGS) plays an indispensable role in balancing seasonal and daily mismatches between natural gas supply and demand, supporting grid reliability, and increasingly facilitating the integration of variable renewable energy sources into broader energy systems. Optimizing UNGS operations requires accurate prediction of deliverability—the rate at which gas can be withdrawn from a storage site—across a wide range of geologic settings including depleted hydrocarbon reservoirs, saline aquifers, and salt caverns, each governed by distinct subsurface flow and geomechanical characteristics. Conventional deliverability estimation relies on physics-based reservoir simulation, which, while rigorous, is computationally expensive and difficult to scale across the hundreds of active storage sites that collectively constitute national and regional gas storage portfolios. Machine learning (ML) has emerged as a powerful complementary approach, capable of learning the complex, nonlinear relationships between reservoir, operational, and geologic parameters and resulting deliverability outcomes directly from extensive field datasets. This review synthesizes recent advances in ML-based prediction and optimization of underground natural gas storage, examining artificial neural network, ensemble, hybrid metaheuristic-optimized, and explainable machine learning architectures applied to deliverability prediction across depleted reservoirs, aquifers, and salt caverns. It draws on case studies from heterogeneous depleted gas reservoir optimization, geothermal-analogue petrophysical characterization, and geomechanically informed wellbore stability assessment, and further considers methodological parallels with large-scale AI-driven predictive analytics and economic impact frameworks developed for national healthcare systems. The review concludes by identifying persistent challenges in cross-formation generalizability, real-time operational integration, and model interpretability, and proposes a research agenda for advancing ML-based UNGS optimization toward routine use in supporting reliable, cost-effective natural gas storage operations.