A statistical approach to control conservatism of robust uncertainty propagation methods; application to accidental thermal hydraulics calculations

摘要

In the safety analysis of a physical system such as a nuclear reactor, best estimate calculations associated with statistical uncertainty modelling are gradually developing. Complex physical phenomena potentially generating non-regular behaviour with respect to uncertainties tend to impose the choice of highly robust uncertainty propagation methods (such as the classical "Wilks" Monte-Carlo method) to estimate a safety quan-tile with the desired confidence interval. In turn, those methods evidence a large probability of high conservatism if the high CPU requirement of the model computation limits the number of trials. In that context, a number of alternative propagation methods (incl. structural reliability-based) have been classically designed, but inevitably require regularity hypotheses. An alternative approach is suggested: it involves the use of statistical tests as a tool to infer the conservatism of a first series of robust calculations, and to suggest physical relevance investigations and finally decide whether more calculations are worthy. The comparative advantage of this approach lies in the fact that while the robustness of the initial safety estimation is not affected, and no additional (high CPU) model trial is required, the operator has a tool to control and optimise the number of computer trials. A number of dedicated statistical tests (partly inspired by the "outlier"-control statistics) are suggested and discussed. A selection has been tested on an industrial thermal hydraulics computation as to their operational performance in conservatism control. Further research is suggested to confirm the performance of the strategy, including other physical systems or uncertainties.