Session: 05-07: System Performance and Safety Enhancements

Paper Number: 134769

134769 - Transient Identification of Htr-Pm Based on Graph Neural Networks 

Abstract: 

Nuclear power plants, due to the unique consequences of accidents involving radioactive leaks, face severe consequences. Therefore, safety-related design and prevention are of utmost importance in the context of nuclear power plants. High-temperature gas-cooled reactors (HTGRs) represent one of the fourth-generation advanced reactors with inherent safety. HTGRs utilize helium as a coolant in the primary loop, allowing for higher temperatures under high pressure compared to water. However, due to the density limitations of helium, a single reactor cannot achieve high power levels. HTR-PM, a type of high-temperature gas-cooled reactor, adopts a multi-module design, employing multiple reactors to drive a turbine for power generation, addressing power output challenges in this class of nuclear power plants. In contrast to traditional pressurized water reactors, HTR-PM incorporates a main pipe control design at the secondary loop's feedwater and steam pipes, resulting in a more complex overall structure. Consequently, changes in temperature and pressure at the main pipe of the secondary loop impact parameters of each reactor in the primary loop, making transient identification for HTR-PM challenging.

In recent years, the development of deep learning has led to the application of various data-driven methods in the field of transient identification in nuclear power plants. Methods based on Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) have achieved some success in transient identification. However, due to the complex physical structure of HTR-PM, it is necessary to choose appropriate deep-learning methods that consider the physical structure characteristics of HTR-PM during data feature extraction. In this regard, the use of CNN and LSTM methods for transient identification currently falls short of adequately addressing considerations of the nuclear power plant's physical structure.

Graph Neural Networks (GNNs), as a type of deep learning network structure, require constructing an appropriate topological graph based on the correlations between input features after obtaining data features. This topological graph is then embedded into the network structure to facilitate training. Transient identification for HTR-PM based on graph networks involves designing a GNN topological graph according to the physical structure of HTR-PM. This approach effectively considers the coupling of physical structures during feature extraction. The main pipe control structure design of HTR-PM aids in constructing the corresponding topological structure, enabling the neural network to perceive and learn the physical characteristics and correlations between different nodes in the topological graph, aligning well with the principles of graph neural networks. Experimental results illustrate the effectiveness of this method, enhancing the interpretability of the model by accurately locating the module stack within the primary loop responsible for a transient event.

Presenting Author: Wenji Zhang Institute of Nuclear and New Energy Technology of Tsinghua University

Presenting Author Biography: PH.D student of Institute of Nuclear and New Energy Technology of Tsinghua University

Authors:

Wenji Zhang Institute of Nuclear and New Energy Technology of Tsinghua University
Tianhao Zhang Institute of Nuclear and New Energy Technology of Tsinghua University
Jitao Li Institute of Nuclear and New Energy Technology of Tsinghua University
Duo Li Institute of Nuclear and New Energy Technology of Tsinghua University
Chao Guo Institute of Nuclear and New Energy Technology of Tsinghua University
Xiaojin Huang Institute of Nuclear and New Energy Technology of Tsinghua University