Scenario Analysis and PRA: Overview and Lessons Learned

ABSTRACT: The recent trend to use a best estimate plus uncertainty (BEPU) approach to nuclear reactor safety analysis instead of the traditional conservative approach can lead to the production of very large amounts of data. Therefore the need for methodologies that can facilitate the analysis of high volumes of data in terms of both the cardinality (due to the high number of uncertainties included in the analysis) and the dimensionality (due to the complexity of systems) arises. Clustering methodologies offer powerful tools that can help the user to identify scenario groups that are representative of the data and thus can reduce the effort involved in data analysis. In this paper, we consider data that are generated using the dynamic event tree approach for nuclear power reactor transients which contain a large set of state variables (e.g., temperature, pressure of specific nodes in the simulator) and information regarding the status of specific components/systems. Techniques for clustering of the raw data, using in particular Mean-Shift methodology, are discussed and evaluated. We highlight the lessons learned from the research activities at The Ohio State University and possible future research directions. In addition, pre-processing of the raw data and dimensionality reduction techniques are described and compared using several examples.