| Ensuring the safe operation of water distribution networks(WDNs)is not only the realistic need for the current society to maintain normal function,but also the important pillar for the booming development of the national economy in the future,which is of great significance.In order to achieve this goal,it is necessary to timely and accurately estimate the hydraulic state of WDNs under normal conditions,so as to help grasp and control the dynamic operation process,which is a positive means of promotion;On the other hand,it is essential to actively detect and analyze possible abnormal events of WDNs,so as to facilitate handle anomaly in time,which is a reverse correction method.But both tasks currently face similar difficulties: poor timeliness and low accuracy.Specifically,the reason for the poor timeliness and low accuracy of hydraulic state estimation is that its core tool,the hydraulic model of WDNs,has insufficient real-time performance and high uncertainty.The reasons for the poor timeliness and low accuracy of anomaly analysis are the lack of mining of real-time monitoring data of WDNs,and the lack of efficient technical methods in key processes.In order to solve the above problems,this work systematically studies the hydraulic state estimation and abnormal event analysis of WDNs,in order to form a variety of effective means to ensure its safe operation.The nodal water demand is an important input data for the hydraulic model,and the latter is a key tool for hydraulic state estimation and anomaly analysis of WDN.Consequently,improving the accuracy and timeliness of nodal water demand prediction can enhance the performance of hydraulic state estimation and abnormal event analysis.This study proposes to replace virtual nodal water demand with monitored flow data for large users to predict water demand.Considering various factors such as weather,time and society,this study firstly introduces the efficient machine learning algorithms to efficiently and concurrently establish their own multivariable prediction models of water demand for multiple nodes in the WDN.The average value of the determination coefficient for prediction results reaches 0.847,which achieves accurate prediction of nodal water demand.Under the premise of constant updating of input data,the models can reliably predict the hourly nodal water demand in the next 24 hours,which significantly improves the timeliness of the nodal water demand prediction.Aiming at the problems of poor timeliness and low accuracy of the hydraulic state estimation of WDNs,a solution including a variety of technical measures is established.On the one hand,the constantly updated hourly nodal water demand prediction is used as the input of hydraulic model to improve the timeliness of hydraulic state estimation from the source.On the other hand,in the process of hydraulic state estimation,data preprocessing,nodes grouping and data assimilation are combined to reduce the uncertainty of hydraulic model,and Monte Carlo simulation(MCS)and First-Order Second-Moment(FOSM)uncertainty quantification methods are introduced to explore the propagation mechanism of uncertainty and its dynamic fluctuation process.The application analyses show that above three methods can play their roles in the establishment or operation of hydraulic model,reduce the uncertainty of different sources,improve the accuracy of hydraulic model and the hydraulic state estimation results.The study also shows that the uncertainty of hydraulic model is weakened in the propagation process,and the uncertainty of nodal pressure is directly affected by the fluctuation of water demand of nodes in the WDN,which changes dynamically in 24 hours a day.Aiming at the problems of poor timeliness and low accuracy of abnormal event analysis of WDNs,the key processes of the abnormal event analysis are studied,including the detection and identification of abnormal events,the location and size estimation of abnormal events,and the corresponding efficient technical methods and core frameworks are established respectively to improve the timeliness and accuracy of abnormal event analysis as a whole.The contents include:(1)The detection and identification of abnormal events: on the basis of obtaining hourly nodal water demand prediction and measurement,a hydraulic anomaly detection and identification method based on characteristic analysis of flow error curve is proposed.The method first analyzes the respective flow curve patterns of four types of abnormal events,namely burst,leakage,unauthorized water consumption and sensor failure,and then uses the feature extraction method combined with the Convolutional Neural Network(CNN)algorithm to establish recognition models.The application research results show that the method can quickly and reliably identify four types of abnormal events,and the average recognition accuracy in actual case is 85.46%.The feature extraction technique and the integration strategy based on sub-models with time windows of different lengths enable the recognition models to have better recognition performance and generalization ability,faster computing speed,and have great application prospects in engineering practice.(2)The location and size estimation of abnormal events: Firstly,a method for judging anomaly areas of WDNs based on iterative partitioning technique and identification algorithm is proposed,which includes two steps of "partitioning" and "identification"."Partitioning" refers to the sequential use of graph theory technique,clustering method,and manual fine-tuning to divide the pipeline network into two sub-regions."Identification" refers to the establishment of recognition models to determine the sub-region where abnormal events are located by combining the computational efficiency advantages of the hydraulic simulation subgroup technology and the computational accuracy advantages of the CNN model feature screening.The application research results show that the method can identify anomaly sub-region reliably and efficiently.The introduction of sub-group technology shortens the calculation time and improves the timeliness of the scheme.Feature screening reduces the complexity of the recognition models,thereby significantly improving computational efficiency.The "partitioning" and "identification" process are iteratively carried out to eventually delineate the abnormal events in the smallest possible area.Temporary pressure meters are then installed in the delineated anomaly sub-region to collect sufficient anomaly data,while using Long Short-Term Memory network(LSTM)algorithm and hydraulic simulation data to train precise localization and size estimation model.Finally,the trained model is used to analyze the anomaly data to obtain the precise location and estimated flow of abnormal events.The application research results of both the case and the actual WDN show that the established scheme can accurately locate the precise location of abnormal events,and at the same time estimate the scale of abnormal events relatively accurately,which has strong engineering applicability.The performance of the models remains robust even with high levels of uncertainty in various parameters of the hydraulic simulation and LSTM algorithm.This verifies the robustness and effectiveness of the scheme.Through the above research,this study realizes the accurate prediction of the water demand for large users(nodes)of WDNs.On this basis,the problems of poor timeliness and low accuracy in the process of hydraulic state estimation and abnormal event analysis of WDNs are solved.It forms a dual means to ensure the safe operation of WDNs,which has a strong theoretical and practical value. |
PDF Full Text Request