Uncertainty Quantification and Sensitivity Analysis of a Machine Learning-Based Spill Fire Model for Nuclear Power Plants

Sahin, Elvan; Henkes, Peter; Lattimer, Brian Y.; Duarte, Juliana P.

Permalink

https://hdl.handle.net/2142/121829 Copy

Description

Title

Uncertainty Quantification and Sensitivity Analysis of a Machine Learning-Based Spill Fire Model for Nuclear Power Plants

Author(s)

• Sahin, Elvan
• Henkes, Peter
• Lattimer, Brian Y.
• Duarte, Juliana P.

Issue Date

2023

Keyword(s)

• Spill fire
• Machine learning
• Uncertainty quantification
• Sensitivity analysis

Abstract

Fuel spill fire is a significant risk scenario in nuclear power plants (NPPs) that requires accurate and reliable models to assess potential consequences. However, current models suffer from large uncertainties, making it challenging to predict the spread and burning behavior of liquid fuel in a spill fire accurately. To address these challenges, this study proposes a machine learning-based spill fire model that incorporates uncertainty quantification and sensitivity analysis to enhance the accuracy and reliability of risk assessment. To develop the spill fire model, past experimental data on spill fire for different spill sizes and fuels are collected. The spill is considered a continuous/fixed quantity in an unconfined area, and the dominant parameters are identified using sensitivity analysis and Sobol indices. The study also employs the DAKOTA software to perform uncertainty quantification on the model and evaluate the uncertainty propagation of the dominant parameters. Moreover, the study uses the Monte Carlo method with a sampling size of 2048 to generate data from DAKOTA-MLP simulations. The data is used to train a machine learning model which is integrated with DAKOTA to obtain statistics of the spill fire peak heat release rate (HRR). The algorithm can also provide uncertainties associated with the lack of data or data uncertainty and, therefore, indicating regions where more data is needed. Additionally, sensitivity analysis is carried out for four input parameters and one output parameter for each spill fire scenario. The results show that the fuel leakage rate or quantity, the slope of the spill surface, ignition delay time, substrate material, and fuel properties are the most critical parameters that impact the peak HRR of the spill fire.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121829

Sponsor(s)/Grant Number(s)

U.S. Department of Energy, Office of Nuclear Energy Award No. DE-NE0008981

Owning Collections