Uncertainty and sensitivity analysis for the modeling of a start-up experiment in a VVER-1000 reactor

The transition from the application of conservative models to the use of best-estimate models raises the question about the uncertainty of the obtained results. This question becomes especially important, if the best-estimate models should be used for safety analyses in the field of nuclear engineering. Different methodologies were developed to assess the uncertainty of the calculation results of computer simulation codes. One of them is the methodology developed by Gesellschaft fuer Anlagen- und Reaktorsicherheit (GRS) which uses the statistical code package SUSA.

In the frame of the recently finished EU FP5 funded research project VALCO, that methodology was extended and successfully applied to different coupled code systems, including the uncertainty analysis for neutronics. These code systems consist of a thermal hydraulic system code and a 3D neutron kinetic core model. One of the code systems applied was ATHLET coupled with the Rossendorf kinetics code DYN3D. Two real transients at NPPs with VVER-type reactors documented within the VALCO project were selected for analyses. One was a test with the switching-off of one of two main feed water pumps at the VVER-1000 Balakovo-4 NPP. Based on the relevant physical processes in the transients, a list of possible sources of uncertainties was compiled. Besides control parameters like control rod movement and thermal hydraulic parameters like secondary side pressure, mass flow rates, pressurizer sprayer and heater performance, different neutron kinetic parameters were included into the list of possible sources of uncertainties. Sets of input data with statistical variation of the relevant parameter values were generated for a large number of runs of the coupled code.

The SUSA package was used to make a statistical analysis of the result parameters from the output data of the calculations. Time-dependent rank correlation coefficients were calculated showing the influence of the varied parameters on the output parameter under investigation. The most interesting output parameters are the physical parameters for which experimental data are available. The calculation results allowed also the determination of time-dependent tolerance intervals for given coverage and confidence. The comparison of the experimental data, the (best-estimate) reference solution and the tolerance intervals showed how the agreement between experiment and calculation could be quantified. In most of the cases the tolerance intervals include the experimental curves. A compiled list of the most important input parameters based on the rank correlation coefficients shows, which input parameters and models are responsible for the deviations. This list gives indications for further model improvements and code developments.