Study of a transient identification system using a neural network for a PWR plant

This paper presents the procedure and results of a system for identifying PWR plant abnormal events, which uses neural network techniques. The neural network recognizes the abnormal event from the patterns of the transient changes of analog data from plant parameters when they deport from their normal state. For the identification of abnormal events in this study, events that cause a reactor to scram during power operation were selected as the design base events. The test data were prepared by simulating the transients on a compact PWR simulator. The simulation data were analyzed to determine how the plant parameters respond after the occurrence of a transient. A method of converting the pattern of the transient changes into characteristic parameters by fitting the data to pre-determined functions was developed. These characteristic parameters were used as the input data to the neural network. The neural network learning procedure used a generalized delta rule, namely a back-propagation algorithm. The neural network can identify the type of an abnormal event from a limited set of events by using these characteristic parameters obtained from the pattern of the changes in the analog data. From the results of this application of a neural network, it was concluded that it would be possible to use the method to identify abnormal events in a nuclear power plant