Study on Process Parameters to Optimize Monitoring System for Severe Accidents

Nuclear accidents have been analyzed to develop rapidly when operators failed to take correct safety action by misinterpreting information available by means of instrumentation. The problem of making right decision on safety ensuring actions becomes more challenging when instrumentation to safetycritical parameters is failed to give true plant condition due to partial or complete damage. In a nuclear power plant (NPP), most of the systems are linked due to processes of fluid flow, heat transfer etc., and their natural tendency to respond to changes during accident conditions. Variations in physical parameters during a process can be utilized to develop smart applications for plant accident monitoring and management. In this research, the statistical relationships among the process parameters have been analyzed to construct virtual sensor networks that provide the capability to monitor the safety functions from the other affected parameters. The framework employs statistical correlation and regression models developed from the accident simulation data computed at the set-points for safety systems and initial conditions. The proposed methodology has been applied to a specific loss of coolant accident (LOCA) scenario simulated from MAAP code, using correlation coefficient and artificial neural networks (ANN), for estimating post-accident monitoring (PAM) parameters at different intervals of accident progression using virtual sensor networks comprised of severe accident management guidelines (SAMG) suggested parameters. Two importance measures, accuracy improvement factor (AIF) and accuracy reduction factor (ARF) have been suggested to infer the characteristics of virtual sensor networks. AIF characterizes a sensor on the basis of its accuracy whereas the ARF characterizes a sensor on the basis of its uncertainty. Importance measures (AIF, ARF) are sensor and network specific, and are only useful while utilizing the output generated from a particular network. A measure correlation voting index (CVI) has also been suggested to identify the faulty outputs generated from a set of virtual sensor networks, which would be the problem of concern when one or more sensors from a set of virtual networks are unavailable. The estimations from virtual sensor networks are expected to improve by utilizing the importance measures and concepts to generalize the neural networks. CVI was tested for few cases for the identification of faulty sensor and estimation of output. The results found were quite consistent in those cases. CVI thus, provides a capability to select a set of related outputs, which would be used as a yardstick for comparing results in case missing or uncertain inputs are present. The technique can be extended to develop a severe accident management database to serve as an integral part of existing monitoring systems. However, relationships developed in this manner would only be of practical use, if a fair degree of similarity exists between the actual accident conditions and the one assumed while developing these relationships