Application of Artificial Intelligence to the Source Term Database: Clustering of Accident Scenarios and Prediction of Offsite Consequences

Jin, Kyungho; Cho, Jaehyun; Kim, Sung-yeop; Vechgama, Wasin

Permalink

https://hdl.handle.net/2142/121827 Copy

Description

Title

Application of Artificial Intelligence to the Source Term Database: Clustering of Accident Scenarios and Prediction of Offsite Consequences

Author(s)

• Jin, Kyungho
• Cho, Jaehyun
• Kim, Sung-yeop
• Vechgama, Wasin

Issue Date

2023

Keyword(s)

• Source term
• Probabilistic safety assessment
• Artificial intelligence (AI)
• Clustering
• Prediction

Abstract

The level 2 probabilistic safety assessment (PSA) is an important risk analysis method of nuclear power plants (NPPs) for estimating large early release frequency (LERF) and the amount of radioactive material release from postulated accidents as the safety goal requirement of the Republic of Korea (RoK). Previously, only dominant accident scenarios among large accident scenarios were selected as representatives for source term estimation using logical tree analysis or expert judgement. Consequently, it is difficult that these source term results from representative scenarios would cover the risk insight of all scenarios in level 2 PSA. To increase the significant imbalance between the number of scenarios in the risk assessment, the Korea Atomic Energy Research Institute (KAERI) proposed an approach called exhaustive simulation to mechanically construct an enormous database of source terms for a large number of severe accident scenarios. This approach allows us to address some legacy issues not previously considered in the existing level 2 PSA. To effectively utilize this database and support the PSAs, KAERI is currently developing a framework based on artificial intelligence (AI) technology. In this paper, the authors introduce two studies of AI applications using this database. The first study quantitatively categorizes severe accident scenarios to select the best representative one using a clustering method after extracting a feature through an autoencoder for the performance of the clustering. It turned out that this approach can reduce the uncertainties in selecting representative scenarios compared to the qualitative logical tree or expert judgement. The second study develops an accident consequence prediction model using source term database with a convolutional neural network to provide appropriate information for quick and better decision-making during an actual accident. The prediction model successfully provides expected accident consequences when various severe accident conditions are given. These studies were performed based on the database of about 650 severe accident scenarios, including the results of source term and accident consequence analysis

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121827

Owning Collections