Development and application of objective uncertainty measures for nuclear power plant transient analysis

For the development, design and licensing of a nuclear power plant (NPP), a sound safety analysis is necessary to study the diverse physical phenomena involved in the system behaviour under a wide range of operational and transient conditions. Such studies are based on detailed computer simulations, backed by appropriate experimental information where available. With the progresses achieved in computer technology and the greater availability of experimental and plant data, the use of so-called best estimate codes for safety evaluations has been gaining increasing acceptance in the nuclear community (i.e. among regulators, policy makers and utilities). These codes predict the physical response of a NPP to a postulated (or real) transient event by using a physically more realistic, and hence less conservative, description of the system. The application of best estimate safety analysis has opened new prospects but also raised new problems that need to be addressed. Thus, it has become more crucial to assess as to how reliable code predictions are, especially when, for example, they need to be compared against safety limits that must not be crossed. It hence becomes necessary to identify and quantify the various possible sources of uncertainty that could affect the reliability of the results. Currently, such uncertainty evaluations are generally based on experts' opinion, although, when data are available, analytical methods based on parametric statistics can also be employed. In the present doctoral research, a novel methodology based on a non-parametric statistical approach has been developed for objective quantification of one of the most demanding contributors to best-estimate code uncertainties, viz. that related to the physical models which the code uses. The basis is an evaluation of the accuracy of a given physical model achieved by comparing its predictions with experimental data from an appropriate set of separate-effect tests. The discrepancies, i.e. the differences between measurements and predictions, can be considered stochastically distributed for several reasons, and thus a statistical approach can be employed. The first step which has been taken is the development of a procedure for investigating the dependence of a given physical model's accuracy on the experimental conditions. Each separate-effect test effectively provides a random sample of discrepancies between measurements and predictions, corresponding to a location in the state space defined by a certain number of independent system variables. As a consequence, the samples of "errors", achieved from analysis of the entire database, are associated to various individual points over the state space. By applying a novel multi-dimensional clustering technique, based on the non-parametric statistical Kruskal-Wallis test, it has been possible to achieve a partitioning of the state space into regions differing in terms of the quality of the physical model's predictions. The second ste