A Methodology for Global Sensitivity Analysis for Time-Dependent Code Output Applied to Reflood Simulation in TRACE
Understanding properly the impact of model parameters and their interactions on the predictions is required for an appropriate model assessment. For simulation of reflooding following a LOCA, this requirement is justified because the important parameters affecting predictions often cannot be measured and should be considered uncertain. Moreover, different process representations in the model may have strong influence at different times of the transient. These issues complicate the task of model assessment. In this work, a global sensitivity analysis (GSA) methodology tailored to analyze transient code output is developed. The methodology was based on the Morris and the Sobol’ methods. The first was utilized to screen out non-influential parameters which allowed generation of larger samples with fewer code runs for the second method, the Sobol’ method. The method results in global sensitivity indices quantifying the contribution of inputs variations to the output variation, considering interactions among them. The Functional Data Analysis (FDA) techniques were then used to post-process the output deriving new quantities of interest (QoI)s describing the overall functional variation. The method was successfully applied to a reflood simulation model using the thermal-hydraulics (TH) system code TRACE with 26 input parameters. Complementing the conventional QoIs (such as the max. temperature and time of quenching), the FDA-derived quantities gave a deeper insight on particular modes of functional variation and attributing them to the variations of the model parameters. The temperature transient was divided into two modes: the ramp phase and the descent phase. The ramp phase showed that the model was additive in terms of the parameter variations with Dispersed Film Flow Boiling-related parameters explained most of the functional output variation. Yet, the variation during the descent could only be explained through parameters interactions. This indicated the non-identifiability of the model for that specific phase of the transient with respect to the temperature response.