Predicting initial spread rate of continuous spill fires using machine learning

Abstract

Continuous spill fires pose significant hazards in industrial settings, with their rapid initial spread being a critical factor in fire escalation. To effectively predict the development of fires and achieve early control, this study applies machine learning models for prediction, specifically by collecting a large amount of previous experimental data on sustained spill fires, the spreading and combustion behaviors of fuels at different spill rates are thoroughly analyzed. In order to evaluate the key parameters affecting the development of spill fires, three advanced ML models, namely, Random Forest, Gradient Boosting, and Support Vector Machine Regression prediction model, were employed, and a Random Forest prediction model with a coefficient of determination (R²) of 0.91 and a mean square error of 0.15 was successfully constructed for accurately predicting the spread rate of the fuel in the initial spreading stage. The results of the study showed that fuel discharge rate was the most important factor influencing the initial spread rate, and among the eight influencing factors fuel discharge rate accounted for 56.3 % of the overall importance, next in sequence are slope, which accounts for 14.3 %, followed by substrate width at 11.1 %, substrate thermal conductivity making up 8 %, fuel properties representing 6 %, and fuel heat of combustion contributing 3 %, and lastly longitudinal wind speed, it was found that the open space and the tunnel space did not have a significant effect on the initial spread rate of spill fires.