Abstract:

To objectively and comprehensively assess fire risks in high-rise residential buildings, and to address that traditional assessment methods were over-reliant on expert experience and had a high degree of subjectivity in their results. First, various machine learning algorithms, including EasyEnsemble and XGBoost, were compared based on historical fire and publicly available datasets. The optimal algorithm was selected to predict the probability of fire occurrence. Second, a fire consequence assessment indicator system was constructed from four aspects, including human, building, environment and management. Machine learning models were integrated with the structural entropy weighting method to determine indicator weights, and fire consequences were quantified based on scoring criteria. Then, the results of both approaches were coupled to quantify fire risk levels. Finally, the method's feasibility and effectiveness were validated using City J as a case study. The results indicate that the EasyEnsemble algorithm exhibits higher area under curve (AUC) values and recall rates, outperforming the Random Forest(RF), XGBoost, Support Vector Machine(SVM) and Light Gradient Boosting Machine(LightGBM) algorithms in predicting the probability of fire occurrence. The weights calculated by the XGBoost algorithm demonstrate greater discriminative power, better reflecting the relative importance of indicators. The high-rise residential fire risk assessment method proposed in this paper, which integrates machine learning with an indicator system, enables a comprehensive quantitative assessment of fire risk. It provides a reference for fire departments in formulating fire risk prevention and control strategies.

Key words: EasyEnsemble, indicator system, high-rise residential buildings, fire risk assessment, extreme gradient boosting (XGBoost), structure entropy weight method