ABSTRACT

Water main breaks are a common recurring problem in water distribution networks, resulting in cascading effects in the whole system and the interconnected infrastructures such as transportation. Having integrated the physical features of pipes such as diameter and environmental factors like precipitation, we propose predictive models based on spatiotemporal data and machine learning methods. In this study, the dataset is the main breaks recorded from 2015 to 2020 in the city of Tampa, Florida. First, a spatial clustering is conducted to identify vulnerable areas to breaks. A time series analysis is also carried out for the temporal data. The result of these analyses informed the machine learning algorithms as independent variables. We then compared the predictive models based on information-based and rank-based criteria. Obtained results indicated that Boosted Regression Tree (BRT) model was superior to the others. Finally, we present predicted normalized failure rates for the water distribution network to inform rehabilitation and fortification decisions at the municipality level.

摘要

供水总管断裂是配水管网中常见的反复出现的问题，会导致整个系统和交通等互联基础设施的级联效应。在整合了管道的物理特征（如直径）和环境因素（如降水）后，我们提出了基于时空数据和机器学习方法的预测模型。在这项研究中，数据集是佛罗里达州坦帕市 2015 年至 2020 年记录的主要中断。首先，进行空间聚类以识别易受破坏的区域。还对时间数据进行了时间序列分析。这些分析的结果告知机器学习算法作为自变量。然后，我们比较了基于信息和基于等级的标准的预测模型。获得的结果表明，Boosted Regression Tree (BRT) 模型优于其他模型。最后，我们提出了供水网络的预测归一化故障率，以告知市政一级的修复和设防决策。