Research On Leakage Detection And Location In District Metering Areas In Water Distribution Systems Based On Data-driven Methods

The leakage of urban water distribution systems has become a difficult issue to be solved in water supply industries.For leakage detection and location,water supply companies determine whether it is a leakage event according to whether there is water overflow on the ground(consumer's report or professional manual inspection).Although some online monitoring systems have alarm functions,most of them are still difficult to detect and locate leakage timely and accurately.Taking into account the National Natural Science Foundation and the actual engineering situation,this paper studied leakage detection and location in a District Metering A rea(DMA)based on data-driven methods and the methods were verified by experimental data.The main work and innovation of the thesis are as follows:1.This paper studied large leakage detection in one-inlet and multi-inlet DMAs with water demand uncertainty and studied a method based on probability distribution.The method analyzes the fluctuation characteristics of a residential DMA,and eatablishes nonlinear autoregressive(NAR)prediction model by normal flow data,then calculates the deviation between the monitoring value and the predictive value.The abnormal probability of the deviation is defined by Gaussian distribution density to determine whether it is a leakage event.The method is tested and verified for one-inlet DMA and multi-inlet DMA.The results show that the studied method has good performance for large leakage events.2.Information contained in single timestep is limited and the above method is not easy to detect small leakage.Therefore,the method of real-time leakage detection based on patterns of water demand with supervised learning is studied.The similarity between patterns of daily water demand is measured by dynamic time warping(DTW)distance.Based on DTW distance,abnormal patterns in historical dataset can be eliminated and normal patterns of daily water demand can be distinguished.The leakage patterns are simulated by adding an increment of flow into normal patterns,then a leakage identification model is modeled based on patterns of water demand with supervised learning by random forests.Compared with other methods by SCADA flow data and simulated leakage events,the results obtained demonstrate that the proposed method can detect small leakage in a fast and reliable manner with a low false alarm rate.3.In order to locate leakage zone as soon as possible,the leakage location method based on clustering algorithm and feature similarity is proposed.Using the pressure data in a DMA,pipe network nodes that pressure fluctuation characteristics are similar caused by leakage can be clustered in a leakage zone by clustering algorithm,and the leakage feature vector can be extracted from each leakage zone.When the abnormal pressure fluctuation occurs in the DMA,the similarity between the abnonnal fluctuation vector and the leakage feature vector of each leakage zone is calculated,and the most similar one is the most suspicious leakage zone.Example applications of this method have been implemented to confirm that leakage can be located in a DMA effectively.4.Based on the experimental system,this paper introduces the deman and main technologies of the system.By Java?MySQL database and other technologies,the system framework including functional module,database module,interface module,etc is introduced.The method for leakage location has been tested and verified on two cases study involving the physical experimental system and the EPANET simulation based on an real DMA.In summary,this paper focused on leakage detection and location in district metering areas in water distribution systems based on data-driven methods and introduced a prototype system.The methods have been verified by experimental data.The proposed methods are applicable to online monitoring of leakage in DMAs,which are of great value to the safety in urban water distribution systems.