Property Loss And Casualty Prediction For Building Fires Based On Ensemble Method

With the continuous acceleration of urbanization process and the rapid growth of population,the scale and density of urban buildings are increasing in order to meet the needs of life and work in today’s society.The occurrence of building fires is inevitable,which has brought immeasurable property loss to the country and society,and even caused serious casualties.Accurate prediction of property loss and casualty in building fires will be beneficial to disaster relief organization and post-disaster recovery.Therefore,property loss and casualty prediction for building fires are both important problems to be solved in fire research.Machine learning is an essential core technology in the era of big data.In recent years,some studies have applied machine learning methods to the construction of disaster loss prediction models.However,the traditional machine learning cannot meet the actual needs,so it should be improved continuously to meet the developmental needs.To improve the accuracy of building fire loss prediction,we present an interpretable Boosting tree ensemble method(IBTEM)for property loss prediction that uses R2-based weighted voting strategy to create an ensemble of boosting trees and interpretable Shapley additive explanations to analyze the model internal mechanism,improving the prediction performance and interpretability of the model.For casualty prediction,we present an imbalanced classification ensemble method based on XGBoost with random oversampling and adaptive threshold(ROS-AT-XGBoost)that combines XGBoost with random oversampling method to create an ensemble learning model and uses an adaptive threshold adjustment strategy to determine the optimal equilibrium point,improving the classification performance and application value of the model in practical problems.In order to verify the effectiveness of the proposed method,this paper combines American building fire data provided by the United States Fire Administration(USFA)with the weather data collected by the National Oceanic and Atmospheric Administration(NOAA)from 2012 to 2016 in the experiments.The proposed IBTEM and ROS-AT-XGBoost are compared with other popular machine learning methods as well as state-of-the-art methods proposed in recent literatures,e.g.,decision tree,neural networks,random forest.The experimental results show that the proposed method achieves better prediction performance.In addition,a series of experiments on improvement including winsorization,logarithmic transformation,recursive feature elimination(RFE),Pearson correlation analysis,sample quantile of order P and multisource data fusion are conducted to verify the effectiveness of the introduced strategies.In conclusion,the method proposed in this paper has achieved favorable prediction performance,which can be effectively applied to property loss and casualty prediction for building fires and,thus,assists relevant departments in making disaster relief decisions of dispatching assistance and mobilizing resources in a timelier manner.Meanwhile,this method can make up for the deficiency of the traditional prediction methods in accuracy,so as to predict loss more accurately in advance,which is practically significant for the rational planning of rescue,reduction of disaster loss and improvement in recovery efficiency.