Stochastic explosion risk analysis of hydrogen production facilities

Highlights

• A new stochastic Bayesian Regularization Artificial Neural Network (BRANN) method for explosion risk analysis.

• The BRANN method enable to minimize the uncertainty and the stimulation intensity for the risk study.

• The BRANN method assist in estimating the exceedance frequency of maximum overpressure.

• The exceedance frequency is used for design and explosion risk management purpose.

Abstract

Explosion risk analysis (ERA) is an effective method to investigate potential accidents in hydrogen production facilities. The ERA suffers from significant hydrogen dispersion-explosion scenario-related parametric uncertainty. To better understand the uncertainty in ERA results, thousands of Computational Fluid Dynamics (CFD) scenarios need to be computed. Such a large number of CFD simulations are computationally expensive. This study presents a stochastic procedure by integrating a Bayesian Regularization Artificial Neural Network (BRANN) methodology with ERA to effectively manage the uncertainty as well as reducing the stimulation intensity in hydrogen explosion risk study. This BRANN method randomly generates thousands of non-simulation data presenting the relevant hydrogen dispersion and explosion physics. The generated data is used to develop scenario-based probability models, which are then used to estimate the exceedance frequency of maximum overpressure. The performance of the proposed approach is verified by analyzing the parametric sensitivity on the exceedance frequency curve and comparing the results against the traditional ERA approach.

Introduction

The global demand for energy keeps growing, and analyst forecast has been given that the energy demand will increase by 30%–40% until 2050 [1]. Hydrogen is one of the most favorable energy alleviating the high demand in the coming years due to its advantages, such as being environment-friendly, high combustion energy, and being renewable. Also, there have been many hydrogen-related applications all over the world. One of the most promising applications is fuel cell vehicles, which may significantly contribute to reducing environmental pollution compared to the current fossil fuel vehicles.

In spite of many advantages, one needs to beware of the potentially high risks of the hydrogen and its applications [2]. For example, the hydrogen is prone to be ignited due to its wide flammability range, which may cause the potential fire and explosion threat to public safety. So far, many hydrogen-related fire and explosion accidents have been documented [[3], [4], [5]]. To prevent or mitigate the hydrogen-related risks, the Quantitative Risk Analysis (QRA) researches have been conducted [2,[6], [7], [8], [9], [10], [11], [12], [13], [14], [15]]. Most of the works focused on the hydrogen refueling stations with the relatively open surroundings, however, few research works were for the hydrogen production plants where most of the produced and consumed hydrogen in the industry are generated [[8]]. Furthermore, the hydrogen production plants are equipped with many facilities, and the relative obstructions would significantly affect the QRA results by influencing the consequence modeling of the QRA methodology compared to the hydrogen refueling stations [16]. Therefore, concerns should be given to the relevant risks of the hydrogen production plants as well [8,17].

As one of the most significant parts of the QRA, Explosion risk analysis (ERA) could provide accidental design loads that contribute to the explosion risk-based mitigation design. Generally, the ERA framework consists of two parts, namely the consequence modeling and probability modeling. The consequence modeling provides the hydrogen dispersion-explosion scenario-related consequence results, and by integrating the associated results, the probability modeling is then applied to generate the exceedance frequency curve. For those located in relatively open areas such as hydrogen refueling stations, the consequence modeling generally employs the simplified hydrogen dispersion and explosion models, such as Muti-Energy and Baker-Strehlow-Tang models [2,[6], [7], [8], [9], [10], [11], [12], [13], [14], [15]]. Regarding the hydrogen production plants in the obstructed environment, the CFD tool is a suitable alternative for the hydrogen dispersion-explosion consequence modeling due to its numerical models, which are possible to take account of the obstacle-inducing turbulence effect on the consequence of hydrogen explosion accidents [16,18].

The CFD-based ERA approaches have widely used for the onshore and offshore chemical plants [[18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30]]. However, the traditional approaches still have the dispersion-explosion scenario-related parametric uncertainty, which would affect the robustness of the ERA process. Stochastic approach is the widely used alternative to manage the scenario-related uncertainty. In particular, several stochastic approaches have been recently proposed within the scope of renewable energy applications [31]. Some of the proposed approaches immediately employ the Monte Carlo Simulation (MCS) to generate numerous scenarios [[32], [33], [34], [35], [36]]. However, for our objective, a large number of CFD simulations should be conducted if immediately using the MCS, which is much computationally expensive. Alternatively, the data-driven approaches have been proposed for scenario generation of renewable resources such as wind and solar as well as demands [37]. Furthermore, for the CFD-based ERA process, some data-driven approaches, such as response surface methodology and frozen cloud approach, have been incorporated into to reduce the computational intensity [[25], [26], [27]]. However, the authors' previous work has already verified the poor performance regarding the robustness and accuracy of both the above approaches and accordingly proposed the advanced non-intrusive approach, namely Bayesian Regularization Artificial Neural Network (BRANN), to study the hydrocarbon dispersion in process facilities [28]. The BRANN, which is one type of artificial neural networks, has a better generalization capacity, especially under minimal simulation data, mainly attributed to integrating the Bayer's theory and generalization term. Also, the authors further verified the accuracy of the integration of the BRANN approach and the CFD-based ERA process, where the ERA results based on the whole CFD simulations were regarded as the benchmark [29,30]. However, due to the computational cost-related restriction of the whole CFD simulations, the scenario-related parametric uncertainty has not yet been comprehensively analyzed mainly for the relevant chemical plants during the detailed design phase.

This study proposes a stochastic ERA procedure of hydrogen production facilities. The BRANN-based non-intrusive approach is integrated with the CFD-based ERA to manage the scenario-related uncertainty effect comprehensively. (1) FLACS is applied to model the hydrogen dispersion-explosion-related physics in the obstructed environment. (2) Based on the simulations, the BRANN-based models for dispersion and explosion studies are developed. (3) The hydrogen dispersion-explosion scenario-related parametric probability models are assumed. By systematically integrating the above three steps, the converged exceedance frequency curve is derived based on the MCS. Sensitivity analysis of the number of the hydrogen dispersion-explosion scenarios on the exceedance frequency, as well as comparison with the traditional approach, is conducted. The parametric uncertainty effect reduction and computational efficiency of this proposed procedure are verified by a case study of hydrogen production plants.