최철원  서동균

Vol. 40,  No. 5, pp. 93-99, Oct.  2025
10.14341/JKOSOS.2025.40.5.093

Risk Assessment power plant safety Machine Learning hazardous energy Logistic Regression Random Forest

PDF

Thanks to governmental policy efforts and heightened public safety awareness, the mortality rate due to accidents has decreased by two-thirds over the past 20 years. However, the accident rate remains unacceptably high considering the economic level of our nation. In response, the government established the "Roadmap for Major Accident Reduction" at the end of 2022, aiming to lower accident rates to compare with advanced nations by 2026. A key component in this roadmap is reforming the risk assessment process. Current Job Safety Analysis (JSA) practices at our power plants align well with standard procedures, including preparation, identification of hazards, risk estimation and evaluation, and development of countermeasures. Nonetheless, two key challenges persist: (1) the inevitability of qualitative estimations in existing power plant risk assessments, and (2) over-representation in initial risk assessments. This study addresses these challenges by exploring the applicability of Hazardous Energy concepts when estimating accident severity via historical data, while also proposing a methodology for real-time risk assessment. In this study, we conducted machine learning on 407 power plant accident data points, examining the relationships with and without Hazardous Energy attributes. We applied two machine learning techniques—Random Forest and Logistic Regression algorithms—to each relationship model. The results indicate that adding the Hazardous Energy attribute to the existing two attributes enhances accuracy in both algorithms. Notably, the Logistic Regression algorithm demonstrated an additional capability to predict one more fatality compared to previous models.