Using finite mixture models in thermal-hydraulics system code uncertainty analysis

Abstract

Nuclear Power Plant safety analysis is mainly based on the use of best estimate (BE) codes that predict the plant behavior under normal or accidental conditions. As the BE codes introduce uncertainties due to uncertainty in input parameters and modeling, it is necessary to perform uncertainty assessment (UA), and eventually sensitivity analysis (SA), of the results obtained. These analyses are part of the appropriate treatment of uncertainties imposed by current regulation based on the adoption of the best estimate plus uncertainty (BEPU) approach. The most popular approach for uncertainty assessment, based on Wilks' method, obtains a tolerance/confidence interval, but it does not completely characterize the output variable behavior, which is required for an extended UA and SA. However, the development of standard UA and SA impose high computational cost due to the large number of simulations needed. In order to obtain more information about the output variable and, at the same time, to keep computational cost as low as possible, there has been a recent shift toward developing metamodels (model of model), or surrogate models, that approximate or emulate complex computer codes. In this way, there exist different techniques to reconstruct the probability distribution using the information provided by a sample of values as, for example, the finite mixture models. In this paper, the Expectation Maximization and the k-means algorithms are used to obtain a finite mixture model that reconstructs the output variable probability distribution from data obtained with RELAP-5 simulations. Both methodologies have been applied to a separated effects experiment, and to an integral effects simulation.