Using Machine Learning to Assess Spill Fire Data for Use in Fire PRA

A fuel spill can be a risk-significant fire scenario in nuclear power plants (NPPs). Fuels can leak from main components, such as pumps, hydraulic valves, and diesel generators. A spill fire has a complex, dynamic changing burning area due to the fuel spread and fuel burning characteristics. To quantify spill-fire risk in NPPs, it is essential to analyze the spread of the liquid fuel and the fuel-burning behavior. However, the current spill fire models have large uncertainties and do not account for some variables in the dynamic process of the spill fire, such as different surface angles, ignition delay time, substrate thermal conductivity, and some fuel physical properties. In the current study, past spill fire experimental data are collected to analyze the spreading and burning behaviors for different spill sizes and fuels. The spill is classified as continuous/fixed quantity in unconfined area. Identifying the dominant parameters is required for creating such a fire model able to determine the spill size and maximum heat release rates. Random forest machine learning (ML) models are used to further evaluate the most important parameters that impact the fire scenario conditions and provide a computationally efficient model to support fire PRA. The results of ML models quantify the importance of the fuel leakage rate or quantity, the slope of spill surface, substrate material, and fuel properties, which are the main key factors in predicting the peak heat release rate.