Transient Identification of HTR-PM Abnormal Feedwater Flow Based on Graph Neural Networks

Abstract The safety requirements for nuclear power plants are very strict, as an accident may lead to radioactive leakage, causing unbearable losses. Therefore, safety related research in nuclear power plants is very important. High Temperature Reactor-Pebblebed Modules (HTR-PM) are a fourth generation advanced reactor with inherent safety features. HTR-PM adopts modular reactors, using multiple reactors to drive one turbine for power generation. To achieve a multi-module design, HTR-PM has a fluid network for steam and feedwater. When the feedwater flow rate of a modular reactor changes, the physical quantities in the fluid network are also affected, resulting in a very complex transient of the HTR-PM at this time. In recent years, many deep learning methods have been applied to the fault diagnosis of nuclear power plants. These methods cannot consider the physical structural characteristics of nuclear power plants when extracting data features. Therefore, graph neural networks can effectively solve this problem. The graph neural network constructs a topology graph based on the fluid network structure of HTR-PM and combines it with the topology graph when the features propagate with the network layer, fully considering the physical structural characteristics of the nuclear power plant. After experiments, the network structure of LSTM-GCN has achieved excellent results in transient identification of water supply anomalies.