- Meneses Gelves, Jairo Andrés|||0009-0002-1790-3997
- Vacca, Pascale|||0000-0001-5863-1030
- Planas Cuchi, Eulàlia|||0000-0002-7053-3959
- Pastor Ferrer, Elsa|||0000-0002-2985-3635
As hydrogen deployment grows, hydrogen refuelling stations (HRSs) are becoming a fundamental infrastructure in urban and industrial environments. This expansion makes land-use planning (LUP) an essential component of hydrogen safety assessment, since decisions on facility siting, separation distances, and exposure of nearby populations depend on the risk assessment methodology to support planning and permitting. In this context, quantitative risk analysis (QRA) provides a conventional framework for estimating accident frequencies and consequences to help LUP decisions.
Conventional QRA typically relies on fault tree (FTA) and event tree (ETA) analyses. Within this structure, fault trees represent causes leading to a top event, such as hydrogen release, while event trees depict possible post-release outcomes, such as a jet fire, flash fire or vapour cloud explosion (VCE). However, the frequency of accident-initiating events involving hydrogen, which is key for FTA and ETA frameworks, is limited by the scarcity of hydrogen-specific failure data. Many studies derive these frequencies from conventional fuels or generic databases, but hydrogen's unique properties (such as low ignition energy and rapid dispersion) need more tailored risk analysis approaches.
In this context, Bayesian Networks (BNs) offer a probabilistic framework for representing causal dependencies among events and updating probabilities when new evidence is introduced. Unlike static FTA/ETA approaches, a BN enables posterior inference across parent and child nodes and incorporates operational observations, expert insights, and accident evidence. This feature allows the analysis of how prior failure probabilities change after conditioning on observed evidence, the examination of how the model's probabilistic assumptions relate to recorded accident information, and the diagnosis of the system by identifying the causes most strongly associated with the observed outcomes.
The present study develops a Bayesian network-based analysis for a HRS. This study follows the approach by Xiao et al. (2024) in which fault tree and event tree are mapped into a dynamic Bayesian network for an HRS (see Figure 1). Based on information gathered in the Hydrogen Incident and Accident Database (HIAD), which compiles recorded hydrogen-related incidents and accidents from different applications and operational contexts and includes information on accident circumstances, initiating events, involved equipment, and observed consequences, evidence was set into the BN to analyse how the probabilities of parent and child nodes change using 12 HRS accidents recorded up to 2025 (Joint Research Centre, 2025). In this way, the accident records were used to examine the behaviour of the original network structure under evidence-based updating.
The results show that incorporating HIAD-based evidence modifies the probability distributions of both parent and child nodes. The posterior probabilities of the final scenarios vary depending on the accident evidence introduced. The analysis also shows that accident records can be used to compare prior model assumptions with posterior results obtained after evidence propagation. From a methodological perspective, this allows the FTA/ETA-based accident representation to be linked with an updateable probabilistic structure for hydrogen-specific risk analysis in HRS applications. This approach is intended to support the treatment of hydrogen-specific uncertainty in QRA.