Uncertainty Quantification for Reactor Safety Analysis.

Abstract

The present work developed new methodologies based on code surrogates and deterministic sampling strategies for uncertainty quantification (UQ) of nuclear power plant (NPP) transients in reactor safety analysis. These methodologies take advantage of and efficiently use the additional computational resources available to perform more simulations with system thermal-hydraulic (TH) codes obtaining additional and more reliable uncertainty information compared to conventional UQ methods used in reactor safety analysis. The methodologies were demonstrated for a Best Estimate Plus Uncertainty (BEPU) licensing calculation and the analysis of a dynamic event tree (DET) for a realistic NPP transient. The first methodology uses the Alternating Conditional Expectation (ACE) algorithm, a powerful nonparametric regression technique, to develop a dynamic code surrogate that can accurately simulate time dependent, nonlinear TH behavior of a NPP transient considering multiple safety system degradations or failures. A surrogate taking the form of a discrete time dynamic system model with four input parameters and a recursive relationship was developed to predict the subcooled water level in a reactor core during the recirculation phase of a hot leg large-break loss-of-coolant accident (HL-LBLOCA). The model uncertainty of the of the ACE surrogate was derived and the unscented transform (UT), a sampling based UQ method, was used to propagate model uncertainty in the surrogate predictions. The second methodology demonstrates the applicability of the UT as a general, sampling based UQ methodology. The UT uses a deterministic sampling algorithm to obtain estimates of the mean and variance of the output parameter of interest with significantly smaller sample sizes opposed to random sampling schemes. The primary advantage of the UT is the size of the UT sample determining the computational expense of the method scales linearly with the size of the input parameter space. Linear scaling keeps the simulation of large complex systems computationally manageable compared to geometric scaling, a common constraint in DET analysis of NPPs.