Modeling Sequence Timing, Technical Specification, and Code Parameter Uncertainties in Success Criteria Definition with Gaussian Process Model

Best estimate simulations of nuclear power plant (NPP) transients can be performed in support of success criteria definitions in Level 1 Probabilistic Safety Assessment (PSA). Reducing the use of conservatisms and bounding assumptions in the analysis can give a more realistic estimate of the safety margin provided by the safety systems configurations representing the success criteria. Furthermore, rigorous treatment sequence timing uncertainties in success criteria definitions is difficult within the conventional event tree/fault tree (ET/FT) methodologies used in Level 1 PSA. This paper presents a new methodology to estimate safety margin while addressing sequence timing, safety system configuration, technical specification, and thermal hydraulic code parameters uncertainties. The key aspect of the methodology is the Gaussian process model (GPM), a nonparametric regression method for multivariate regression with internal estimate of model uncertainty, is used to process data from many simulations and is a surrogate model for predicting safety parameter distributions as a function of input uncertainties. The methodology is demonstrated for the injection phase of a large-break loss-of-coolant accident (LBLOCA) and the safety margin of the Ulchin Units 3 and 4 (UCN3 and 4) success criteria are quantified. A new methodology to estimate safety margin of a NPP has been proposed and demonstrated for best estimate simulation of LBLOCA in support of Level 1 PSA success criteria definitions. The methodology simultaneously considers sequence timing, safety system configuration, technical specifications, and code model parameter uncertainties. A key aspect of the methodology is the input parameter space is partitioned into two subsets of inputs, explicit regression variables consisting of the dominant input uncertainties that are the fundamental drivers of thermal hydraulic behavior of the transient and implicit noise variables. A Gaussian process model performs regression on the explicit regression variables and output uncertainty is quantified by a measurement noise term representing the contribution of the implicit input noise variables to local variation or uncertainty of the safety parameter. This approach retains high fidelity treatment of all input uncertainties during best estimate simulation of the transient, but allows the analyst to focus regression analysis on the most important application specific parameters thereby overcoming the curse of dimensionality inherent to the analysis of complex systems