Application of a Deep Learning Technique to the Development of a Fast Accident Scenario Identifier

Abstract

To obtain more accurate results of probabilistic safety assessment (PSA), it is necessary to reflect more complete dynamics of nuclear power plants. In analyzing these more realistic PSA models, numerous thermal-hydraulic code runs should be performed that typically take from a few minutes to several hours. This paper proposes a fast running model using deep learning techniques to obtain plausible accident scenarios while reducing the resources required to conduct PSA. The developed model is built from a conditional autoencoder, and an analysis of its performance is carried out under both trained and untrained ranges. Taking about one second per scenario, the developed model shows about 0.4% and 1.6% error in the trained and untrained ranges, respectively. As a feasibility study, the aggressive cooldown operation under a small break loss-of-coolant accident in the APR1400 plant was considered. The proposed method can reduce uncertainty in PSA and contribute a key technique to dynamic PSA.

Keywords

• Deep learning
• dynamic PSA
• fast running
• model uncertainty
• parameter uncertainty
• probabilistic safety assessment (PSA)