Evaluation Tests of Event Identification Method Using Neural Network at Kashiwazaki Kariwa Nuclear Power Station Unit No.4

Abstract

An event identification method using a neural network has been evaluated in an on-line environment during the plant startup test at Unit No.4 Plant in the Kashiwazaki Kariwa Nuclear Power Station. In the method, the neural network identifies the event from the change pattern of analog data, such as reactor pressure signals, and then the result is confirmed or similar events are discriminated using digital data, such as valve open signals. Before the test the neural network is trained for the events causing a reactor scram by using analysis results.

For the test the method is incorporated into a prototype of the alarm handling system which is connected to the plant facilities. Five kinds of analog data are acquired and eight sampled data from each, namely a total of 40 data, are input to the neural network after normalization. The results show that the load rejection, the turbine trip and the main steam isolation valve closure events are correctly identified from 9 kinds of subject events, regardless of the difference between the trained analysis results and the recognized plant data. This is owing to the data sampling and normalization methods as well as the robustness of the neural network.

KEYWORDS:

• event identification
• neural network
• on-line evaluation
• nuclear power plants
• BWR type reactors
• alarm handling
• operation support