Bootstrapped artificial neural network model for uncertainty analysis in MELCOR simulation of severe accident

Abstract

This study is concerned with uncertainty analysis of MELCOR simulation of a hypothetical severe accident initiated by station blackout (SBO) in a Nordic boiling water reactor (BWR). The hydrogen mass from cladding oxidation and the vessel failure timing in the accident are selected as the figures of merit (FOMs) in this study. As a conventional approach of uncertainty analysis, 456 cases with random sampling of 31 MELCOR input parameters are executed by the code to produce the empirical cumulative distribution functions (CDFs) and the empirical 95th percentiles of the FOMs. Given the sufficient sample cases, uncertainty analyses through two nonparametric methods at various orders, i.e., the Wilks' method and the Wald & Guba's method, can then be performed to obtain the distributions of 95/95 estimates (95th percentiles estimated at a 95% confidence level) of single FOM and two FOMs. However, the conventional approach turns out to be time consuming and computationally expensive since many sample cases require iterative tuning of MELCOR input to restart and finish calculations. To overcome this issue encountered in the conventional approach of uncertainty analysis, an alternative approach is developed in the present study, in which 150 and 170 MELCOR calculation cases are used to develop bootstrapped artificial neural network (ANN) models which predict single FOM and two FOMs, respectively. The bootstrapped ANN models are then employed in uncertainty analyses through the two nonparametric methods of 95/95 estimates mentioned above. The comparative results show that the alternative approach can reproduce the distributions of 95/95 estimates for both single FOM and two FOMs with less computational costs. Moreover, while the Wilks' method or the Wald & Guba's method at a very high order (e.g., 100th order) can be used in the alternative approach to produce 95/95 estimates closer to the empirical 95th percentile, it is practically impossible to do so in the conventional approach due to unaffordable computational cost of excessive MELCOR runs. Hence, it can be concluded that the alternative approach of uncertainty analysis is not only effective (much less MELCOR cases with least fixing of unsuccessful runs), but also enabling high-order nonparametric methods for 95/95 estimates.