Application of Artificial Intelligence for Estimating Severe Accidents in Nuclear Power Plants Using Offsite Information

Noh, Hyun Seok; Lee, Gee Man; Kim, Jung Soo; Jung, Woo Sik

Permalink

https://hdl.handle.net/2142/121798 Copy

Description

Title

Application of Artificial Intelligence for Estimating Severe Accidents in Nuclear Power Plants Using Offsite Information

Author(s)

• Noh, Hyun Seok
• Lee, Gee Man
• Kim, Jung Soo
• Jung, Woo Sik

Issue Date

2023

Keyword(s)

• Estimate severe accident
• Accident response
• AI learning
• MURCC
• Supervised learning

Abstract

When a nuclear power plant loses power, such as in the case of the Fukushima accident, the internal key operation parameters of nuclear power plant cannot be acquired due to the loss of power. In case of Japan, where SPEEDI is operated, it was not possible to accurately predict the spread of radioactive contamination because it was not possible to acquire source term information from internal key parameters. This paper aims to diagnose the type of accident and the size of the accident in a nuclear power plant using artificial intelligence techniques based on the measurement values of externally installed radiation measuring instruments in the event of a similar accident. There are not enough real data to learn the measurement information of major operating variables and radiation measurement data of severe accidents for AI training. Therefore, to obtain similar radiation measurement data, radiation measurement data was generated using an off-site consequences assessment code. RASCAL, MACCS, and MURCC were used in this paper. MURCC is a MACCS post-processing code developed by the Nuclear Integrated Safety and Security laboratory of Sejong University. RASCAL selects the type of radioactive material release accident and the degree of core damage in a nuclear power plant, specifies the release magnitude, and calculates the total release amount of leaked nuclides and the release amount at 15-minute intervals. The total release amount is used as an input to MACSS, and MACCS calculates the time-accumulated nuclide concentration and exposure dose at the centerline of the radiation cloud by nuclide. MURCC calculates two- dimensional nuclide concentrations at every 15-minute from the time-stepped nuclide concentrations at the radioactive plume centerline in MACCS and the 15-minute interval emissions of nuclides in RASCAL. The 15-minute nuclide concentrations at the instrument location and the initial selected release type and core damage level in RASCAL are used as training data. The algorithms used for training were compared using machine learning methods (linear regression, random forest, XGBoost) and artificial neural networks (DNN). To evaluate the learning, 900 data were divided into training and test data in a 7:3 ratio and evaluated based on the accuracy on the test data(see attachment 1). The learning results showed that XGBoost had the highest LOCA and SBO accident diagnosis in single-unit accidents with 97% accuracy, while linear regression had the lowest with 79% accuracy. Using the method presented in this paper, we were able to distinguish the type of LOCA and SBO accidents. In addition, DNN has the highest accuracy of 92% and linear regression has the lowest accuracy of 67% in classifying the degree of core damage. The method presented in this paper can classify the degree of core damage. However, in the case of multiple accidents, both accident diagnosis and core damage classification were less than 50% accurate in classifying simultaneous accidents. The reason for the low learning rate is that the combined nuclide concentration of two accidents was not well distinguished. Further research on the diagnosis of multiple-unit simultaneous accidents is needed

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121798

Sponsor(s)/Grant Number(s)

• Nuclear Safety and Security Commission Nos. 2106062-0323-SB110
• Republic of Korea Nos. 2204017-0223-SB110

Owning Collections