Application of Machine Learning in Identification, Classification, and Prediction of Injection Cycles in Underground Hydrogen Storage
Paper ID : 1017-UGESR (R1)
Authors
Amin Talat *, Seyyed Alireza Tabatabaei-Nejad
Department of Petroleum Engineering, Sahand University of Technology, Tabriz, Iran
Abstract
Underground hydrogen storage (UHS) in depleted gas reservoirs and saline aquifers enables seasonal balancing of renewable energy supply and demand. Repeated injection–withdrawal cycles involve complex multiphase flow, relative-permeability hysteresis, viscous fingering, gravity override, cushion-gas mixing and minor dissolution. Although full-physics simulators provide reliable predictions, they remain computationally intensive for multi-cycle optimization and uncertainty quantification. Machine learning (ML) has emerged as a powerful complementary approach for rapid surrogate modeling, injection-regime classification, thermophysical property prediction and adaptive control.This review synthesizes peer-reviewed studies (2015–2026) and organizes ML applications into five functional pillars: property prediction, regime identification, time-series forecasting, surrogate modeling, and optimization & control. Deep neural network reduced-order models achieve R² ≈ 0.99 with speedups exceeding 10⁴–10⁵×, while LSTM and ensemble classifiers demonstrate high predictive accuracy in cyclic forecasting and fingering detection. Hybrid physics-informed ML approaches exhibit superior robustness compared with purely data-driven models. Remaining challenges include limited field-scale cyclic datasets and long-term geochemical coupling.
Keywords
Keywords: Underground hydrogen storage, machine learning, cyclic hysteresis, surrogate modelling, optimisation, reservoir management
Status: Accepted (Oral Presentation)