Research On Intelligent Identification And Early Warning Of Leakage Signal In Water Supply Network System

Although the total amount of water resource is abundant in China,the per capita possession is low.With the increasing demand of water resource,the contradiction between supply and demand has become increasingly prominent.Water supply network system as an important water supply facilities to ensure the safety of water for residents,when it is in a long-term leakage state,it will not only cause a lot of water resources waste,affect the people’s drinking water safety,but also reduce the operation efficiency of the water department.Therefore,water supply network leakage is a problem involving resources,security and economy.In order to reduce the leakage rate of water supply network,improve the utilization rate of water resources and maintain the operational benefits of water division,this thesis proposes a data driven model based on forecast-classification to identify and early warn the leakage signal of water supply network system.Compared with previous studies,this method can avoid the dependence on long series of data and the low prediction accuracy.The detailed improvements are as follows:(1)Attention + Gate Recurrent Unit(GRU)was selected as the basic framework of the leakage signal recognition model.It effectively avoids the problems of vanishing gradients and exploding gradients in the prediction process of traditional Recurrent Neural Networks(RNN),as well as the difficulty of fitting RNN on long series of data.(2)In the model training and prediction stage,the traditional method is abandoned,and a feedback correction training method and an improved multi-step prediction method are proposed,which overcome the dependence of the traditional model on long series of data in the training process and the accumulation of errors along the time axis in the prediction process.(3)In the classification stage,multi-threshold classification method and improved Rayda criterion are used for early warning of leakage events.This method improves TPR and reduces FPR at the same time.In this study,the above methods were first constructed and verified in the L-Town case pipe network,and then the model was also constructed and applied in Qingyuan water supply pipe network system,and the following conclusions were drawn:(1)Two real leakage events in 04-02 and 06-02,2018 in the L-Town case pipeline network were successfully identified,and the response time to the leakage and leakage events was shorter than 30 min.At the same time,42 leakage and loss events were manually added to the flow data from 05-01-05-14,and the model TPR was 92.86% and FPR was 2.73%.The results show that the model has the highest recognition accuracy at night,with TPR of 98.3% and FPR of 0.86%.(2)Successfully identified and warned of 8 leakage incidents in the southern part of Qingyuan City in 2021,and only 1 false alarm.The early warning time of 7 leakage events was earlier than the actual arrival time of the water division,and the response time of 1leakage event was the same as the actual arrival time of the water division.It has played a positive role in promoting the water resources protection and the safe operation of water supply network.At the same time,the abnormal detection model of data-driven tube explosion signal identification and early warning based on forecastingclassification has been proposed in other fields and it has made a contribution to promoting the construction of smart city.