Thesis title: Machine Learning Modeling of High-Pressure Hydrogen Leakages
The introduction of hydrogen and its blends as fuel in gas turbines is one of the most promising solutions to reduce greenhouse gas emissions in power generation industry. On the other hand, this strategy raises several technical challenges in the design of the new generation of gas turbines, including the development of new safety protocols for fuel handling and storage. In fact, compared to natural gas, hydrogen fueling implies higher risks of ignition and explosion in case of accidental leakages from high-pressure fuel gas systems. For this reason, gas turbine manufacturers started putting even more attention than in the past on detailed quantitative risk assessments of accidental fuel gas leak scenarios within gas turbine enclosures. The goal of this analysis is to ensure that the ventilation system, with which the enclosure is equipped, is capable to dilute fuel gas accumulations, avoiding any dangerous condition. Computational Fluid Dynamics (CFD) is used to simulate the fuel gas dispersion in the enclosure coming from a specific leak location. However, a single leak analysis through CFD can be really complex and computational expensive. Moreover, a leak scenario can depend on different factors as: leak location, fuel storage pressure, leak section, interactions between flammable cloud and surrounding ventilation flow field or objects. Therefore, a large number of simulations would be needed to explore comprehensively the entire design of experiments of leak scenarios resulting from the combinations of these factors. In this context, the use of machine learning surrogate models can be a valid alternative to explore a large number of cases cutting down computational costs. This dissertation reports the milestones of the author’s doctoral studies devoted to the identification of viable strategies based on machine learning to model high-pressure hydrogen leak scenarios. After a comprehensive analysis of the state of the art practices on both CFD modeling of fuel gas dispersions and on machine learning methods, three main contributes are reported. The first one aimed to validate a reliable CFD methodology to generate a dataset of hydrogen/methane leak scenarios interacting with a ventilation flow field. The leakages were modeled as under-expanded jets, and an analytical representation of free jets was identified through selfsimilarity laws. The second contribution involves the development of a novel machine learning surrogate model based on Graph Neural Networks, trained on the previously generated dataset. The model’s inputs were derived from an analytical representation of free jet flow fields using similarity laws. In this way, the model acted as a map between the available free jet flow field and its correspondent perturbated state by the ventilation flow field. The third contribute aimed to decline the previously presented approach to model impinging hydrogen-methane gas leaks. A new dataset of impinging jets was generated. The model presented is an advanced version of the previous one, but with the same idea of mapping the free jet flow field to its perturbated state, by the presence of a solid surface this time. Collected results showed how these models can achieve high accuracy with respect to the CFD reference, lowering the computational costs of the simulations. The potential of the implemented learning strategies demonstrated to be promising for the industrial applications, net of scalability issues of the algorithms used.