Robust on-line diagnosis tool for the early accident detection in nuclear power plants

Any loss of coolant accident mitigation strategy is necessarily bound by the promptness of the break detection as well as the accuracy of its diagnosis. The availability of on-line monitoring tools is then crucial for enhancing safety of nuclear facilities. The requirements of robustness and short latency implied by the necessity for fast and effective actions are undermined by the challenges associated with break prediction during transients.
This study presents a novel approach to tackle the challenges associated with the on-line diagnostics of loss of coolant accidents and the limitations of the current state of the art. Based on the combination of a set of artificial neural network architectures through the use of Bayesian statistics, it allows to robustly absorb different sources of uncertainty without requiring their explicit characterization in input. It provides the quantification of the output confidence bounds but also enhances of the model response accuracy. The implemented methodology allows to relax the need for model selection as well as to limit the demand for user-defined analysis parameters. A numerical case-study entailing a 220 MWe heavy-water reactor is analysed in order to test the efficiency of the developed computational tool.