Abstract: An underground gas storage well is the most important gas storage mode at compressed natural gas (CNG) filling stations, and its safety is the key to the normal operation of CNG filling stations. In this paper, a probabilistic dynamic safety assessment method based on Bayesian network was proposed to dynamically evaluate the safety of the gas storage wells at CNG filling stations. In this method, a complete causal logic on the leakage failure of gas storage wells at CNG filling stations is established by means of Bow-Tie model, and the corresponding safety assessment model based on Bayesian network is developed according to the mapping relationship between Bow-Tie model and Bayesian network. Then, the prior probabilities of top event and accident consequence, the posterior probabilities in the case of top event, the failure risk when anti-corrosive paint is completely off and the probabilities of accident consequence when the reliability of alarm response is changed are calculated respectively. Furthermore, sensitivity of root nodes is analyzed based on the characteristics of Bayesian network. The dynamic safety assessment results show that the key to improve the safety of gas storage wells at CNG filling stations is to reduce the failure probabilities of high-sensitivity basic event, predict the failure risk periodically based on new evidences and improve the reliability of alarm response to minimize the accident loss.