Physics-informed data-driven Bayesian network for the risk analysis of hydrogen refueling stations

The safety of hydrogen refueling stations (HRSs) is receiving increasing attention with the growth use of hydrogen energy. Existing risk assessment methods of HRS are primarily based on expert knowledge, which is affected by potential subjectivity. This paper aims to present a new hybrid risk assessment method incorporating HRS accident data and physical knowledge into a Bayesian network (BN) model to analyze the key risk influencing factors (RIFs). The HRS accident data in HIAD 2.1 from 1980 to 2023 is used in this paper, and 30 RIFs are identified based on the accident report information and physical knowledge. To address the issue of the insufficient accident data for BN modeling, the accident data is expanded by Conditional Tabular Generative Adversarial Networks (CTGAN). Bayesian search, Peter-Clark algorithm and Greedy Thick Thinning methods are adopted for structure learning. The expectation maximization algorithm is employed for parameter learning in the BN model. Additionally, K-fold cross validation is used when testing the performance of different BN models.