Development of a neural-network methodology for the safety justification of VVER reactors in manoeuvring modes

Abstract

The article considers the advancement of a methodology developed by JSC OKB “Gidropress” for the calculation safety justification of VVER reactors in manoeuvring modes. The main challenge of the methodology in terms of the accident analysis is the selection and justification of initial conditions, which are carried out through expert assessment. To solve the problem, it is proposed to use machine learning for automating expert assessments based on available calculation results. The article proposes methods for constructing elements of a neural network and an algorithm for its learning. The results of the work of these elements and their combinations for the solution to the given problem are analyzed. Conclusions are made about the possibility of advancing the methodology through the development and implementation of a multilayer neural network that takes into account the accident type, manoeuvring algorithm, moment of the campaign, specifics of a particular project, and other factors important for the safety justification.