Fusion of Deep Learning Technology into Accident Diagnosis and Source Term Estimation

Kim, Sung-yeop; Park, Soo-Yong; Choi, Yun Young

Permalink

https://hdl.handle.net/2142/121824 Copy

Description

Title

Fusion of Deep Learning Technology into Accident Diagnosis and Source Term Estimation

Author(s)

• Kim, Sung-yeop
• Park, Soo-Yong
• Choi, Yun Young

Issue Date

2023

Keyword(s)

• Severe accident
• Accident diagnosis
• Source term
• Artificial intelligence (AI)
• Deep learning

Abstract

In a radiological emergency, diagnosis of accident situation and estimation of the release of radionuclides into the environment should be performed quickly and accurately. There existed continuous demand to overcome the limitation of existing methods and upgrade the source term estimation module installed in decision supporting system such as computerized technical advisory system for a radiological emergency (AtomCARE) in Korea. In this circumstance, Korea Atomic Energy Research Institute (KAERI) tried to fuse deep learning technology in order to develop faster and more accurate diagnose and estimation approach. First of all, 24 key parameters transferred from nuclear power plants (NPPs) to AtomCARE were selected as learning input by expert judgement. In the event of accident, data of 24 parameters are provided as real-time data from NPPs. A deep learning model employing Transformer encoder and fully connected layer was developed for universal purpose of accident diagnosis and source term estimation. Medium-break loss of coolant accident (MLOCA) and total loss of component cooling water (TLOCCW) scenarios were selected representing low and high pressure of reactor coolant system (RCS) in OPR1000 reactor. 9 and 3 sub-scenarios were derived by employing plant damage state event tree (PDS-ET) of Level 2 probabilistic safety assessment (Level 2 PSA). 200 to 300 learning and testing database was established for each sub- scenario by a large amount of calculation of Modular Accident Analysis Program version 5 (MAAP5). In aspect of severe accident diagnosis, developed deep learning model showed above 99% accuracy with 20,000-second NPP data and above 95% accuracy with 30,000-seconed NPP data to classify 12 sub- scenarios. Environmental release of xenon, iodine, and cesium were reasonably estimated by using 30,000- second NPP data though there exists necessity for fine tuning. 100 verification database was build for each sub-scenario and verification of the model has been carried out. In this study, current technical status of accident diagnosis and source term estimation using artificial intelligence at KAERI is summarized and introduced

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121824

Owning Collections