Flow Prediction And Abnormal Alarm Model Of DMA Based On ?-SVR And IForest

The leakage problem of water distribution network in China is very serious.It reduces the benefits of water supply enterprises and causes great waste of scarce water resources.The key to controlling leakage is the identification of leakage.Traditional user reports or regular inspection based methods are quite passive;the Minimal Night Flow?MNF?method based on District Meter Area?DMA?has a certain lag,and it is impossible to identify illegal water use during the day.Thus,this paper takes the existing DMAs of H region in G city as the research area,and discusses the prediction method for hourly water demand of DMA and its application in leakage identification.The main research contents are as follows:?1?Denoising the historical hourly flow sequence by interpolation.An interactive,convenient and fast outlier marking method for hourly flow data is proposed.The SSA algorithm with strong nonlinear relationship interpretation ability is used to iteratively interpolate the hourly flow data,and the interval quartering method is used to improve the efficiency of optimizing the interpolation parameter L_opt;the relationship between the length of the sequence to be interpolated and the interpolation result is discussed,the recommended sequence length is 3 months.?2?Constructing the hourly water demand prediction model for DMA.The model is based on?-SVR which has been world-widely proved.The optimization method of input feature,especially the historical water consumption feature,is discussed.The optimal feature set is gained through the combined method based on mutual information sorting and cross-grouping verification.The result of the grid search method shows that the linear kernel function with penalty factor C is more suitable for flow prediction.Finally,the influence of training set size on model accuracy is discussed,the recommended training set size is 3 months.?3?Constructing the abnormal alarm model for DMA.A quickly abnormal detection method combing hourly flow related information and iForest algorithm is proposed for the first time.Using iForest algorithm,a one-dimensional training set?predictive residual?based model and a two-dimensional training set?current hour and measured hourly flow?based model are constructed.The results show that each model has its own shortcomings.In order to meet the decision-making needs,it is necessary to ensure that the training set has a sufficient proportion of abnormal samples and apply the one-dimensional model and two-dimensional model simultaneously.At the same time,the iForest model can improve the robustness of the?-SVR model by correcting the historical water consumption features.