Research On Water Demand Prediction And Leak Detection Of Water Supply Network Based On Artificial Neural Network

To promote the construction of water-saving cities,optimize the operation of water supply systems,and reduce the leakage rate,this paper takes a university as an example to study the model of water consumption prediction based on neural networks,and proposes a water supply network leakage detection model based on water meter data and DMA.Firstly,for a relatively independent water supply division,the relationship between water usage in student dormitories,public buildings,and family areas and factors such as time,holidays,temperature,and weather was analyzed,and their importance was ranked.Based on the BP neural network,this paper learns from measured water demand data,using time,holidays,temperature,and weather as input parameters,hourly water consumption as output parameters,and mean square error as evaluation indicators.A water demand prediction model is established.This model predicts the water demand of three regions,with an average relative error of less than 11%.It is found that the error first increases and then decreases with the increase of training data size.The performance of short-term prediction is good,the performance of mid-term prediction is poor,and the performance of long-term prediction will be slightly reduced.Therefore,in order to avoid overfitting or under fitting,the number of neurons should be moderate.At the same time,scale of training data should be sufficient,and introducing the real water consumption of 1h as a reference can significantly reduce errors.The model can predict hourly water consumption in the next three months,providing guidance for the operation of water supply systems.Then,based on EPANET,this article sets various typical leakage conditions on DMA.A leakage dataset is constructed.With time,holidays,temperature,weather,and water consumption as input parameters,and the leakage status of each partition as output parameters,BP neural network is used to learn the leakage data and compare the deviation of water consumption and detects leakage areas.Using accuracy as an evaluation indicator,this paper achieved a short-term detection accuracy of 100%.It is found that the increase in the number of DMA aggravates under fitting phenomenon and weakens overfitting phenomenon.When the number of neurons in hidden layers reaches more than 6,accuracy rate can be stable at more than 90%.Accuracy first decreases and then increases with the expansion of training data scale,until it stabilizes at over 90%.The more DMA there are,the less training data is required to achieve an accuracy of 90%.Increasing the number of DMA is beneficial for improving accuracy.The longer the leakage time of training data,the higher the accuracy.If the flow coefficient is too large or too small,the accuracy will be severely reduced.When the number of DMA reaches 3,the accuracy decreases monotonically with the increase of leakage points.Therefore,when using ANN for leak detection,within the scope of 128 nodes,the number of DMA should reach more than 3,the training data size should reach 200 d,and the number of hidden layer neurons should reach more than 6 and increase with the number of partitions.The leak duration of the training data should reach 12 hours,and the flow coefficient should be consistent with the actual leak condition as much as possible.The number of leak points should be 1,so that the accuracy of leak detection can be stable at over 90%.This article is based on BP neural network.By learning the water consumption characteristics obtained from water meter,the leakage area can be accurately located without other sensors in water supply network.The method saves construction investment and reducs the difficulty of data acquisition.With the popularization of real-time water meters,real-time detection of leakage areas can be achieved.