堆叠集成算法在城市地下排水管网风险评估中的应用

Utilizing stacked integration algorithm for risk assessment of urban underground drainage networks

近年来排水管网事故频发,因此迫切需要一种客观、高效的风险评估方法。基于此,提出基于堆叠模型的城市地下排水管网风险评估方法,显著提高建模的精度。研究选择管长、坡度等14个特征变量进行建模,将堆叠集成算法与独立机器学习算法的预测性能进行比较,探究排水管网风险的影响因素。结果显示:(1)研究区内4级风险和3级风险排水管段数占比分别为21.23%和21.38%,需要尽快在现场进行处置。(2)堆叠集成算法预测精确度为93.7%,高于随机森林(91%)、决策树(89%)和支持向量机(78%)的预测精确度,比独立机器学习算法的评估性能更好。(3)除了管长、管径等常规属性对排水管网风险有重要影响,沉积、障碍物等因素对排水管网的风险也不容忽视。结果有助于排水管网健康评估,对各类机器学习算法评估管网风险建模具有启示意义。

In response to a recent surge in drainage network accidents,there is a pressing need for an objective and efficient risk assessment method.To address this,a new method for assessing the risk of urban underground drainage networks,based on stacked models,has been proposed.Stacked ensemble models,as heterogeneous ensemble algorithms,use the results of three primary learners from the first layer as input variables for training meta-learners of the second layer.This approach helps mitigate disruptions in results caused by data deficiencies in independent machine learning models and significantly enhances modeling accuracy.Fourteen feature variables,including pipeline length,gradient,burial depth,and sedimentation,representing various aspects of pipe networks such as the network's attributes,surrounding environment,and structural defects,are selected to build models aimed at exploring the influencing factors on urban underground drainage network risks.The performance of the stacked ensemble model versus independent machine learning models is evaluated using determination coefficient R-squared,root mean square error,and standard deviation.Additionally,the performance between stacked ensemble algorithms and independent machine learning algorithms is assessed by comparing the operating characteristic curves and areas under the curves of the subjects.The research findings reveal that:(1) Within the study area,the proportion of drainage pipes categorized as Level 4 risk and Level 3 risk is 21.23% and 21.38% respectively,indicating a relatively high proportion of high-risk pipes among the total pipes.This underscores the need for prompt intervention to eliminate pipeline risk factors.(2) The predictive accuracy of the stacked ensemble algorithm is 93.7%,which is significantly higher than that of random forests(91%),decision trees(89%),and support vector machines(78%).This demonstrates the superior evaluation performance of the stacked ensemble algorithm in drainage network risk assessment,compared with independent machine learning algorithms.(3) Besides conventional attributes such as pipeline length and diameter,sedimentation,obstacles,and other factors also play crucial roles in drainage network risk.