基于多源数据融合的供水DMA泄漏诊断机理与方法研究

Water supply pipelines face serious leakage problem, which restricts the safety of urban water supply. Leak diagnosis using data driven model in district metering area (DMA) has been an emerging technology in recent years. However, the conventional predictive-diagnosis method suffers with high false alarm rate due to limited understanding of mechanism and the structure of model. ..To solve ultra-short time water demand prediction problem for DMAs, this project takes data fusion to improve the information utilization as principle. A minute level water prediction algorithm based on deep learning technique is developed using a gated recurrent units network (GRUN) model as the realization frame, by exploring the multi-source data fusion mechanism of the pipe network. To solve the problem of abnormal flow detection under the condition of non-uniform residuals, an outlier detection algorithm based on the similarity of accumulated residuals is developed by analyzing the inhomogeneous distribution characteristics of the predicted and measured residuals. By studying the differential collaborative response of pipe flow under different conditions, the discriminating mechanism of inducement for abnormal water demand is then revealed. Finally, the leak diagnosis problem can be solved by building a DMA leak diagnosis method based on multi-source data fusion. ..The purpose of this project is to form a new theory and new method for DMA leak diagnosis through multidisciplinary intersection. It will also provide a new solution for leakage control, which is of great significance for the safety of water supply and sustainable management of water resources.

配水管线漏损严重，制约城市供水安全。利用数据驱动模型进行供水DMA泄漏诊断是近几年出现的新兴技术。然而，受机理认识和模型结构所限，常用的预测-诊断二阶段法存在误报率高等问题。.针对该问题，本项目以数据融合提高信息利用度为学术思想，以门控循环神经网络为实现手段，通过探究管网多源数据融合机制，构建基于深度学习的分钟级水量预测算法，解决DMA超短时水量高精度预测难题；通过厘清水量预测值与实测值残差的非均匀分布特征，开发基于累积残差相似度的异常数据检测算法，解决残差非均变条件下水量异常的检测难题；通过阐明不同工况下管网流量的差异性协同响应特征，揭示水量异常诱因的甄别机理，构建基于多源数据融合的供水DMA泄漏诊断方法，解决泄漏诊断误报率高的问题。.本项目旨在通过多学科交叉，形成DMA泄漏诊断的新理论与新方法，为供水管网漏损控制提供新思路。对于促进供水安全和水资源可持续管理，具有重要意义。

配水管线漏损严重，制约城市供水安全。受机理认识和模型结构所限，传统预测-诊断二阶段法存在误报率高等问题。本项目以数据融合提高信息利用度为学术思想，综合利用深度学习、统计过程控制、时间序列分解、时空分析等技术手段，将多种数据驱动方法与供水管网计量区系统相融合，主要开展了供水DMA水量变化规律与数据特征提取、超短时区域水量高精度预测、残差非均变条件下异常泄漏检测、基于趋势迁移的多泄漏甄别等方面的研究，形成了基于多源数据融合的供水DMA泄漏诊断方法体系。项目重要结果包括：（1）基于统计分析视角，解析了供水管网计量区水量数据的波动关联性，揭示了水量数据的多源波动特征，以数据融合提高信息利用度为学术思想，构建了高效的数据融合机制。（2）以门控循环神经网络为实现手段，开发了基于门控循环神经网络的超短时流量预测方法，解决了计量区5分钟流量的高精度预测难题。（3）阐明了计量区流量预测残差随时间变化的分布特征，提出了基于单区持续性时变阈值的泄漏识别方法，克服了传统单阈值方法低准报率、高误报率的问题。（4）阐释了不同工况下管网流量的差异性协同响应特征，开发了DMA流量序列时空分解方法，揭示了水量异常诱因的甄别机理，提出了基于多区迁移估计的多工况泄漏甄别方法，克服了计量区多工况泄漏与季节趋势的混叠难题。综上，形成了包含预测-检测-甄别三阶段的供水 DMA 泄漏诊断方法。在我国大力推广分区计量管理的背景下，本研究的成果能够为我国分区计量管理泄漏控制工作提供有效的新方法、新思路、新途径，为快速发现泄漏、降低管网漏损、促进水资源可持续管理提供科学有效支撑。