%0 Journal Article %T 机器学习在给排水管网中的应用
The Application of Machine Learning in Water Supply and Drainage Networks %A 朱远建 %A 黄扬 %A 梁一为 %A 岑敬明 %A 李文丽 %A 陈祈安 %A 朱建全 %A 侯辉 %J Software Engineering and Applications %P 479-493 %@ 2325-2278 %D 2026 %I Hans Publishing %R 10.12677/sea.2026.153045 %X 随着城市化进程的持续加速，给排水管网作为城市生命线工程的核心组成部分，其安全、高效、稳定运行直接关系到居民日常生活质量、城市生态环境可持续性以及城市整体韧性。传统水力建模方法在处理大规模、强非线性的管网运行问题时，普遍存在参数标定耗时久、计算成本高昂、对突发故障响应滞后等突出不足，难以满足现代城市智慧管网的运维需求。近年来，以深度学习、图神经网络为代表的机器学习技术快速发展，凭借其强大的数据挖掘、特征学习和复杂模式拟合能力，为给排水管网的漏损检测、爆管识别、水质预警、需水量预测以及城市内涝预报等关键运维任务提供了全新的解决思路与技术路径。本文系统梳理近五年机器学习在给排水管网领域的研究进展，从方法层面全面回顾传统机器学习、深度学习、图神经网络与物理信息神经网络的发展脉络及核心优势，从应用层面重点探讨其在供水管网漏损识别与定位、污染事件检测、需水量预测，以及排水管网管道缺陷识别、内涝预测等核心场景中的最新研究成果与实践应用。最后，深入分析现有研究方法面临的数据稀缺、模型可解释性不足与跨场景泛化能力薄弱等现实挑战，并对物理信息图神经网络与数字孪生融合等未来发展前景进行展望，为推动机器学习技术在城市给排水管网领域的深度应用与产业化落地提供参考。
With the accelerating urbanization process, water supply and drainage networks—as critical components of urban infrastructure—play a pivotal role in ensuring safe, efficient, and stable operations that directly impact residents' quality of life, urban ecological sustainability, and overall urban resilience. Traditional hydraulic modeling methods exhibit significant limitations when addressing large-scale, highly nonlinear network operations, including time-consuming parameter calibration, high computational costs, and delayed responses to sudden failures, making them inadequate for modern smart urban pipeline management requirements. In recent years, rapid advancements in machine learning technologies—particularly deep learning and graph neural networks—have leveraged their robust data mining, feature learning, and complex pattern recognition capabilities to provide innovative solutions for critical operations tasks such as leakage detection, pipe rupture identification, water quality monitoring, water demand forecasting, and urban flooding prediction. This paper systematically reviews five years of research progress in machine learning applications for water supply and drainage networks, comprehensively examining the developmental trajectories and core strengths of traditional machine learning, deep learning, graph neural networks, and physical information neural networks at the methodological level, while focusing on cutting-edge research achievements and practical implementations in key scenarios—including leakage identification and localization in water supply networks, pollution event detection, water demand forecasting, pipeline defect detection, and flood prediction. Finally, this study conducts an in-depth analysis of the practical challenges faced by existing research methodologies—such as data scarcity, insufficient model interpretability, and weak cross-scenario generalization %K 机器学习， %K 图神经网络， %K 供水排水管网， %K 漏损检测， %K 物理信息神经网络
Machine Learning %K Graph Neural Networks %K Water Supply and Drainage Pipeline Networks %K Leakage Detection %K Physical Information Neural Networks %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=145451