It has been a challenge in dynamic probabilistic risk assessment (PRA) world that a large number of scenarios from one initiating event with time-related scenario evolutions give complexness on an understanding of the transient/accident scenarios. The understanding of risk which enhances the safety of the entire system requires not only the full understandings of scenario evolutions but also the key characteristics of the events: Both success events and failed events. Since the time evolution is now in consideration of the plant risk assessment, a lot of difficulties such as organizing such large amounts of information and interpreting its physical meaning should be properly resolved. Clustering analysis, one of the unsupervised machine learning (ML) techniques, has been discussed in years to group scenarios with similar characteristics and to identify key patterns of each group so that an analyst can understand entire scenario behaviors by groups. Here we propose a novel methodology of identifying key patterns of scenarios in an accident case of a nuclear power plant system with dynamic reliability analysis. In clustering analysis four items need to be considered: 1Clustering algorithm, 2distance matrix, 3variables in clustering algorithm, and 4cluster validity evaluation. In this paper, partition around medoids (PAM) clustering algorithm with global alignment (GA) kernel distance is utilized. GA kernel, which is considered suitable for clustering time series data, is to assess the similarity between time series data by casting the dynamic time warping (DTW) distances and similarities as positive definite kernels. In order to find variables which will be embedded in the clustering algorithm, multilevel flow model (MFM) methodology is leveraged. For a case study, dynamic PRA tool, MOSAIQUE (Module for SAmpling Input and QUantifying Estimator) coupled with a RELAP-5 generates 2,500 scenarios of SBLOCA. Advanced power reactor 1400 MWe (APR- 1400) is used as a reference plant model. The proposed classification and identification approach has grouped the 8000 scenarios with only 77 clusters and the result can show key patterns shown in core damaged and safe cases which static PRA may not present.