Session: 07-10: Simulations and Predictions - II

Paper Number: 134779

134779 - Research on Accident Prediction of Nuclear Power Plants Based on Deep Learning 

Abstract: 

The safety impact of nuclear reactor accidents on nuclear power plants is mainly characterized by environmental contamination and radiation threats due to the release of radioactive material, which may trigger evacuation and sheltering on a regional scale, with significant economic losses, as well as negatively affecting the public's psyche and trust in nuclear energy technology. In order to minimize these potential impacts, nuclear power plants adopt a variety of safety measures and are regulated by international regulations. Loss of coolant accident (LOCA) refers to an accident in which there is a large breach in the reactor's first circuit, and the rate of coolant loss brought about by the breach exceeds the rate of its system replenishment, so that the temperature of the reactor core gradually rises, resulting in the heating of the fuel rod cladding, which is referred to as a meltdown. The regulator plays a role in regulating the coolant pressure in the first circuit, so the prediction of the regulator level and pressure is critical to the safety of a breach accident. In this study, based on the bidirectional long and short-term memory network (BILSTM) and the long and short-term memory network (LSTM), the rolling update mechanism (RU) is utilized to obtain the time series of the original data to form the training set data and the test set data, and to predict the water level and pressure change of the voltage stabilizer after the occurrence of the LOCA accident of the nuclear power plant, so as to provide some references for the prediction of the LOCA accident of the nuclear power plant. In order to provide a certain reference for the prediction of LOCA accidents in nuclear power plants. The data of Qinshan Nuclear Power Station is utilized as the LOCA accident simulation data as the original data, and the accident level of 0.002 is used as the training data, and the level of 0.008 is used as the test set for prediction. The data are normalized to the range of (0,1) using the MinMaxScaler function to speed up the convergence. The results show that LSTM has better prediction ability for long time series prediction, while BILSTM has better prediction results compared to LSTM model because it can deal with the correlation of before and after data information. Although the average errors of the prediction results of both models are in the order of 10^-3, the average error of BILSTM is reduced by more than 20% compared with LSTM. The existence of underfitting was also found during the study, indicating that there is room for further improvement of the LSTM and BILSTM models for abnormal working conditions.

Presenting Author: Wei Lv Harbin Engineering University, Harbin 150001, China

Presenting Author Biography: Lv Wei is mainly engaged in artificial intelligence fault diagnosis and research on system autonomous control technology based on deep reinforcement learning methods at Harbin Engineering University.

Authors:

Wei Lv Harbin Engineering University, Harbin 150001, China
Tong Li Harbin Engineering University, Harbin 150001, China
Bo Wang Fundamental Science on Nuclear Safety and Simulation Technology Laboratory, Harbin Engineering University, Harbin 150001, China
Sichao Tan Harbin Engineering University, Harbin 150001, China
Jiangkuan Li Harbin Engineering University, Harbin 150001, China
Ruifeng Tian Harbin Engineering University, Harbin 150001, China