ATHLET BASED TRAINING OF NEURAL NETWORKS FOR THE ANALYSIS OF NUCLEAR POWER PLANT (NPP) SAFETY

Katkovsky Е.А., Katkovsky S.E, “Energoautomatika” Ltd, Moscow, Russia, Nikonov S.P., NRC “Kurchatov Institute”, Moscow, Russia, I. Pasichnyk, K. Velkov, T. Voggenberger, GRS mbH, Garching, Germany

22nd Symposium of AER on VVER Reactor Physics and Reactor Safety (2012, Průhonice, Czech Republic)
NEUTRON KINETICS AND REACTOR DYNAMICS METHODS

Abstract

A diagnostics of initiating events, states of reactor systems and relevant safety
variables is important for safe operation of an NPP. On the one side one needs to correctly
interpret signals from measuring instrumentation. On the other side a fast classification of the
transients significantly improves the reliability of NPP operation.
This paper presents an original approach for detection of NPP equipment failures and
for classification of complex transient states of NPP. It is based on the use of artificial neural
networks. As a learning system to train the recognition of safety relevant processes the bestestimate system code ATHLET (GRS) is used.
The sequence and the principles of network-learning, as well as the justification for
selecting the type of neural network are presented. The paper contains two applications of the
approach: a) the integral and local estimates of recognition accuracy of latent and false
rejections in measuring channels and recommendations on the application of the implemented
methodology in automatic process control systems of NPP units; b) a robust classification of
the transients in German NPPs using the implementation of the ALADDIN tool developed in
the Halden Reactor Project.in the framework of ATLAS (GRS) environment.

Presentation / full paper