Data driven simulations for accurately predicting thermodynamic properties of H2 during geological storage

Abstract

The emission of greenhouse gases is a major contributor to global warming, making it crucial to find alternatives to fossil fuels. Renewable energy sources such as solar and wind are desirable options, but their fluctuating production highlights the need for underground storage. Conversion of energy into hydrogen (H2) and its underground storage has gained attention as a solution. Accurate simulation of H2 geological storage involves modeling highly complex H2 behavior under various conditions using precise and sophisticated fluid models and predicting H2′s thermodynamic properties such as density, viscosity, and thermal conductivity. While traditional analytical modeling methods face challenges from H2′s highly complex thermodynamics, this study endeavors to develop and test eight machine learning-based fluid models for such prediction. These models include multi-layer perceptron artificial neural network (MLP-ANN), convolutional neural network (CNN), K-nearest neighbors (KNN), support vector machine (SVM), decision tree (DT), ensemble learning, random forest (RF), and adaptive boosting (AdaBoost). The results of the study demonstrate that machine learning approaches have great potential in determining thermodynamic properties of H2. Among the developed models, DT and KNN are proven to be the best options for integration within reservoir simulators for the most efficient and accurate outcomes. It is estimated that the score values of DT and KNN models are more than 0.994 for all properties. Furthermore, different types of error parameters are determined to be near zero. These results show the potential of suggested models in the prediction of hydrogen properties. In addition, sensitivity analysis showed that temperature is the least effective factor in the determination of H2 thermodynamic properties. As a novel point in this study, the adoption of machine learning-based methods successfully addresses the challenges in prediction of thermodynamic properties of H2 by enabling accurate and comprehensive data-driven modeling of H2′s behavior, and also highlights the promising performance and potential of DT and KNN models for simulations in which experimental data are particularly difficult and expensive to be obtained.