Combining machine learning and multi-objective optimization algorithms to optimize key parameters for underground hydrogen storage

The intermittency of renewable energy sources often leads to surplus energy curtailment, emphasizing the need for efficient large-scale energy storage. Hydrogen, with its high energy efficiency and clean combustion, is an attractive energy carrier. However, its low density and stringent phase transition conditions limit large-scale storage applications on the surface. However, its low density and stringent phase transition conditions limit large-scale storage applications on the surface. Underground hydrogen storage (UHS) has been proposed as a solution for large-scale storage and utilization of surplus renewable energy. The hydrogen injection rate is a critical operational parameter, governing hydrogen storage and production efficiency. Balancing dynamic changes in key indicators (hydrogen production rate, dissolution rate, and storage mass) is essential. This study prioritized hydrogen production rate and dissolution rate (or storage mass) as primary objectives, employing multi-objective optimization to determine cycle-specific optimal injection rates. Advanced machine learning algorithms were used to develop and compare surrogate models across varying parameters and neural network architectures, identifying the most accurate predictive framework. This methodology significantly enhanced computational efficiency for both hydrogen storage modeling and optimization. The study established Pareto front for multiple objectives and provided corresponding injection rate schemes. Results demonstrated that the Long Short-Term Memory (LSTM) model achieved superior predictive performance, and dividing the Pareto front into three regions (low hydrogen loss mode or high storage mode, balanced mode, and high production mode) to meet different needs. These findings offer theoretical guidance for practical UHS applications. • A general framework for optimizing underground hydrogen storage is developed. • LSTM shows excellent performance in predicting key outcomes for hydrogen storage. • Surrogate model couples NSGA-II to obtain Pareto front for objective variables. • Three modes are established to address diverse operational requirements.