Scholarlink

<p>Electrical enclosure fire scenarios represent a major hazard in nuclear facilities, underscoring the critical need to reduce its uncertainties in risk assessments. This study aims to refine and enhance peak heat release rate (HRR) distributions of electrical enclosure fires using a machine learning (ML) approach by quantifying the uncertainties of existing data analysis, thereby improving the reliability of fire probabilistic risk assessments (PRAs). Utilizing data from over 100 enclosure fire experiments, an artificial neural network (ANN) model was developed, achieving an R<sup>2</sup> of 0.85, RMSE of 21.70&#xa0;kW, and MAE of 14.69&#xa0;kW. SHapley Additive Explanations (SHAP) analysis evaluated the importance of input features, including ignition source, cabinet properties, cable properties, and ventilation conditions. The refined model provided denser peak HRR data, enriching cumulative function distributions. A Monte Carlo (MC) interface was integrated with the ML model applying 5%, 15%, and 25% uncertainties to input parameters. Sensitivity analysis, including Sobol indices, clarified the impacts of input uncertainties on model outputs. This 'MC-ML UQ Framework' was compared with current recommendations, demonstrating its contribution in the analysis of electrical enclosure fires in nuclear facilities.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Development of HRR Distributions in Electrical Enclosure Fire Scenario Through Machine Learning

  • 期刊论文

  • 发表日期：

    2025年2月11日

查看全文
  • Elvan Sahin,
  • Peter Henkes,
  • Bruno P. Serrao,
  • Mohammed A. Allaf,
  • Brian Y. Lattimer,
  • Juliana P. Duarte

摘要

Electrical enclosure fire scenarios represent a major hazard in nuclear facilities, underscoring the critical need to reduce its uncertainties in risk assessments. This study aims to refine and enhance peak heat release rate (HRR) distributions of electrical enclosure fires using a machine learning (ML) approach by quantifying the uncertainties of existing data analysis, thereby improving the reliability of fire probabilistic risk assessments (PRAs). Utilizing data from over 100 enclosure fire experiments, an artificial neural network (ANN) model was developed, achieving an R2 of 0.85, RMSE of 21.70 kW, and MAE of 14.69 kW. SHapley Additive Explanations (SHAP) analysis evaluated the importance of input features, including ignition source, cabinet properties, cable properties, and ventilation conditions. The refined model provided denser peak HRR data, enriching cumulative function distributions. A Monte Carlo (MC) interface was integrated with the ML model applying 5%, 15%, and 25% uncertainties to input parameters. Sensitivity analysis, including Sobol indices, clarified the impacts of input uncertainties on model outputs. This 'MC-ML UQ Framework' was compared with current recommendations, demonstrating its contribution in the analysis of electrical enclosure fires in nuclear facilities.