摘要
为定量分析城市火灾风险,引入随机森林算法。首先设计基于随机森林算法的递归特征消除方法优化初选的14个自变量指标,剔除无关或冗余变量;其次基于随机森林算法的平均准确度降低方法计算特征重要度,将其归一化结果作为指标的权重;最后结合指标数据,使用线性加权法评估城市火灾风险。以济南市区为例的实证研究显示:经过特征选择操作后发现只有10个指标时模型误差最小,MSE为7.43×10-4;用于计算指标权重的随机森林模型与实际火灾密度拟合的决定系数R 2>0.85,精度较高,可用于对指标的客观赋权;火灾风险评估结果不仅显示历史火灾较多的区域是高风险区,也可得到火灾发生较少区域的风险值。
In order to quantitatively analyze urban fire risk,random forest algorithm is introduced.Firstly,a recursive feature elimination method based on random forest algorithm is designed to eliminate irrelevant or redundant variables from 14 indicators selected initially;secondly,feature importance is calculated through random forest with average accuracy reduction method,whose normalized result is then taken as indicator weight;finally,combined with index data,urban fire risk is evaluated by linear weighting method.Taking Jinan urban area as an example,the case study showed that:after selecting features,random forest model only preserved a set containing 10 factors with minimum MSE=7.43×10-4;compared with fire density,the determination coefficient(R 2)of random forest model used to calculate index weight is greater than 0.85,proving that objective weight is feasible;the evaluation result not only revealed that high fire density districts were high risk areas,also can get fire risk value of those low fire density areas.
作者
吴立志
陈振南
张鹏
WU Lizhi;CHEN Zhennan;ZHANG Peng(China People’s Police University,Langfang 065000,China;School of Graduate,China People’s Police University,Langfang 065000,China;School of Intelligent Police,China People’s Police University,Langfang 065000,China)
出处
《灾害学》
CSCD
北大核心
2021年第4期54-60,共7页
Journal of Catastrophology
基金
河北省自然科学基金资助项目(D2015507046)。
关键词
城市火灾
风险评估
随机森林算法
特征选择
指标权重
urban fire
risk evaluation
random forest algorithm
feature selection
index weight
作者简介
第一作者:吴立志(1968-),男,汉族,浙江金华人,博士,教授,主要从事火灾风险评估、灭火与应急救援研究.E-mail:wulizhi119@sohu.com;通讯作者:陈振南(1992-),男,汉族,山东淄博人,硕士研究生,主要从事火灾风险评估、灭火与应急救援研究.E-mail:chenzhennan1992@163.com。
