Artificial intelligence methods application for reactor dynamics predicting in the tasks of maneuverable modes safety assessment

Abstract

The work aimed to the development of a methodology for calculating safety assessment of VVER reactor plants in maneuvering modes. This methodology was developed at OKB “GIDROPRESS” to solve the problem of conducting safety analyzes of high-power VVER reactor plants in a flexible operation mode. Performing such work directly requires significant computing resources and multi-parameter expert assessments. Therefore, the main direction of development was the use of a numerical method using neural network models. In particular, the possibility of efficiency increasing of calculations show in terms of choosing the moment in time when the occurrence of the initial event leads to the most conservative results.

In this work, we study the possibilities of further development of the method by constructing neural networks with deep learning aimed at predicting the development of non-stationary processes, taking into account a large number and complex relationships of available parameters. The capabilities of convolution and recursive architectures for constructing neural networks analyzed to estimate the reactor plant dynamics, taking into account maneuvering after an accident occurs. The analysis examines the interpretability of the results in terms of accounting for xenon transients, water exchange operations, and control movement. For software implementation of the method, the VELETMA/GP program is used.

Based on the results of the work, conclusions drawn about the practical significance of the methods used for solving the tasks set for the calculation substantiation of designs of reactor plants with VVER in maneuvering modes. The work uses both the experience of computational justification and the results of validating full-scale tests of maneuvering modes on modern high-power VVER reactors.