Quantifying uncertainties in the estimation of safety parameters by using bootstrapped artificial neural networks

For licensing purposes, safety cases of Nuclear Power Plants (NPPs) must be presented at the Regulatory Authority with the necessary confidence on the models used to describe the plant safety behavior. In principle, this requires the repetition of a large number of model runs to account for the uncertainties inherent in the model description of the true plant behavior. The present paper propounds the use of bootstrapped Artificial Neural Networks (ANNs) for performing the numerous model output calculations needed for estimating safety margins with appropriate confidence intervals. Account is given both to the uncertainties inherent in the plant model and to those introduced by the ANN regression models used for performing the repeated safety parameter evaluations. The proposed framework of analysis is first illustrated with reference to a simple analytical model and then to the estimation of the safety margin on the maximum fuel cladding temperature reached during a complete group distribution header blockage scenario in a RBMK-1500 nuclear reactor. The results are compared with those obtained by a traditional parametric approach