Dynamic risk assessment with bayesian network and clustering analysis

While traditional probabilistic risk assessment (PRA) using event tree and fault tree is mainly concerned with static uncertainties, dynamic PRA techniques address the timeline response of plants in a systematic manner. Major technical challenges of dynamic PRA are related to a very large number of possible scenarios and many iterative simulations to accommodate the possible changes in the probability distribution of input variables. In this study, we propose a novel probabilistic mapping method using a dynamic Bayesian network with clustering analysis for discretized system space in time, which enables one to physically and logically reduce the number of required simulations as well as to quantify system evolution in a probabilistic manner. The mean shift clustering algorithm is used to cluster datasets from similar scenarios so that the proposed approach can be applied in practice at a manageable computational cost without the burden of running too many additional iterative simulations for a large variety of operational conditions of a target system. The risk effect quantification of variations in control units' configuration would lead to the verification and improvement of dynamic system safety.