Safety distances estimation for hydrogen blended natural gas jet fires using neural network methods

The global trend towards hydrogen as a key low-carbon energy vector is rapidly increasing worldwide, due to its high gravimetric energy density. Following this current, there were analysed different ways to produce it, overall, the ones powered by renewable energy sources (green H2), which represents a bridge between today and a zero-emission future. All the types of H2 could be transported with several strategies and in different forms, for example in trucks through liquid or compressed gas state. However, these last two methods need sophisticated structures and elevated costs. At the same time, existing natural gas (NG) transmission and distribution pipelines represent a vast infrastructure that could carry hydrogen blended natural gas (HBNG) over long distances more economically than pure hydrogen pipelines too. Indeed, blending hydrogen into NG networks avoids massive costs for new infrastructures, but raises questions about mixing behaviour, energy transmission, compressor work and material integrity, particularly due to hydrogen embrittlement (HE) of steels under high pressure. HE could be the cause of HBNG leaks from a pressurized pipeline or vessel, and the small molecular size and low ignition energy of hydrogen can lead rapidly to turbulent jet flames. These high-momentum diffusion flames emit heat that can impinge on nearby structures and can trigger catastrophic domino events. Their behaviour is studied nowadays, because the presence of H2 modifies which could be the classic NG one. Accurate prediction of jet-fire flame geometry (length and lift-off) and radiative heat flux is therefore essential for defining safe separation distances. PHAST is a simulating software with semi-empirical models implemented that could offer optimal prevision of HBNG jet fires, Configurating these type of scenarios in simulative software like PHAST could be really useful, for evaluating the jet fire mass flowrate, the flame length and the safety distances and to study the influence of parameters such as orifice diameter, storage temperature and pressure, meteorological conditions and the volumetric concentration of the blended hydrogen. In this work, a Neural Network (NN) is developed to estimate the effects of jet fires. Firstly, jet fire scenarios of HBNG from transmission pipeline were simulated using PHAST software. A database was constructed by configuring accidental jet fire scenarios for obtaining the jet fire mass flowrate, the flame length and the safety distances (Alert zone, Intervention zone, Domino Effect zone) for different combinations of input parameters such as orifice diameter, storage temperature and pressure, meteorological conditions and volumetric concentration of the blended hydrogen. A pseudo-validation was performed to ensure that PHAST could be used for the specific ranges of parameters and compared to literature experimental cases. Before developing the NN, the database results are analysed and standardized, as pre-processing steps. A Multilayer Feedforward Neural Network with Backpropagation was trained and optimized on the database constructed and the hyperparameters were tuned to minimize the loss function. MAE, MSE and R2 metrics were used to evaluate the performance. In the end the optimized version of this NN model could successfully reproduce the original database with MSE errors in the millesimal range.