Machine Learning Based Dynamic Probabilistic Risk Assessment for Multi-Unit Small Modular Reactors

Turkmen, Gulcin Sarici; Yilmaz, Alper; Aldemir, Tunc

Permalink

https://hdl.handle.net/2142/121809 Copy

Description

Title

Machine Learning Based Dynamic Probabilistic Risk Assessment for Multi-Unit Small Modular Reactors

Author(s)

• Turkmen, Gulcin Sarici
• Yilmaz, Alper
• Aldemir, Tunc

Issue Date

2023

Keyword(s)

• Multi-unit SMR
• DPRA
• Transformer network
• RELAP

Abstract

With the recent developments in nuclear technology, the use of small modular reactors (SMRs) is becoming more widespread. SMRs are more economical and reliable designs, equipped with passive safety systems, and have more modular construction features compared to current nuclear power plants (NPPs). SMRs are also expected to increase the load-following capability of NPPs by deploying multi-units at the same site. After the Fukushima-Daiichi incident, performing detailed safety analyzes of multi-unit NPP sites has emerged as an important need, since the shared systems used by SMRs can lead to unexpected consequences in case of an accident. Dynamic probabilistic risk assessment (DPRA) has been shown to be a useful approach for assessing the safety of NPP operations. However, since NPPs are highly complex systems, it is necessary to produce large amounts of data that represent different possible situations during NPP evolution following an accident in order to carry out the DPRA comprehensively. In addition to the fact that it may take months to produce such data with NPP accident analysis codes and DPRA software developed for such a task, it also requires the use of a significant amount of computer and human resources. Recent achievements of machine learning (ML) algorithms in a wide variety of applications have given researchers incentive to use ML in nuclear safety assessments to reduce the computational effort in representing time-dependent NPP accident response data and making predictions for the consequences of accident progression. The use of DPRA datasets is illustrated to train a Transformer Network (TN) model to predict possible multi-unit light water SMR behavior under accident conditions as the accident evolves. The dataset is generated for generic light water SMRs with station blackout as the initiating event using RELAP/SCDAPSIM. The TN models are based on a multi-headed attention mechanism that renders them particularly suitable for time series data due to allowing modeling of dependencies without regard to their distance in the input or output sequences. The TN is implemented using the open-source artificial intelligence (AI) framework PyTorch. The temporal data obtained from the DPRA simulations are first pre-processed and then fed into the TN to predict peak cladding temperature and core outlet temperature. The TN model is then retrained with thousands of scenarios to increase the model accuracy by applying the Transfer Learning approach. The experimental results show that the TN can obtain good performance and possess benefits over other neural network methods.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121809

Owning Collections