Персона: Старков, Сергей Олегович
Email Address
Birth Date
Результаты поиска
Application of spiking neural networks for modeling the process of high-temperature hydrogen production in systems with gas-cooled reactors
© 2019 Obninsk Institute for Nuclear Power Engineering, National Research Nuclear University 'MEPhI'. All rights reserved.The article considers a simulated possible scenario for the joint production of hydrogen and electrical energy using a high-temperature gas-cooled reactor. The considered model is based on a neural network system, which is used as a technological tool for generating control signals. The multi-layer direct-acting neural network is composed of spiking neural elements, the architecture of which is based on interacting reverberation loops. The electro-optical commuting system considered in this article is the base for building a switching communication system between neurons. The use of optical communication and liquid crystal modulators simplifies the mass distribution of a signal to many neurons from different populations and the change of its parameters. This property is necessary to ensure the neural controller high performance. The approximating properties of a neural network are used to control a group of dual electrolytic cells. Each electrolyzer has a set of variables controlling the temperature, chemical composition and current density through the cell. The spiking network, exerting a control action on pairs of electrolytic cells, completely controls the process of low-temperature electrolysis in a copper sulfate solution. The amounts of hydrogen produced at the cathodes of the grouped electrolyzers will be proportional to the amount of gas produced by the high-temperature electrolysis systems, in which nuclear reactors are the sources of thermal and electrical energy. Information coding is carried out by sequences of spiking pulses from groups of 4 neurons. This method of representing the control sequence elements minimizes a false change in the parameters of low-temperature electrolysis. Learning of the neural network system is carried out by a scattered search algorithm. The evaluation of the simulation efficiency has shown the feasibility of constructing hybrid models with a neural network control system, which do not require the use of expensive materials.