Data-driven Approaches To Pressure Transfer Modeling In An Urban Water Supply System And Daily Water Demand Forecast

Identification of the pressure transfer relationship among nodes in a water supply system forms the basis of pressure control and scheduling optimization. Accurate water demand forecast helps to make plans for the production of treated water. Both topics are crucial for water supply management and decision making. In this thesis, a water pressure transfer model for the pipe network in a city is constructed from the real-time pressure measurements. Based on the historical data of water use, models for the prediction of urban water demand are proposed.An approach to modeling the water pressure transfer in an urban water supply network is first presented for the purpose of pressure control. The network is divided into different sub-networks based on the Pearson correlation analysis of the nodal pressure measurements. The Pearson correlation analysis is performed to find out the set of nodes, whose water pressures are highly correlated, and thus a corresponding sub-network is formulated. As a case study, 47 sub-networks are recognized for a region with an area of 250 km2 and 77 nodes in total. For each sub-network, a linear model is constructed to quantify the pressure transfer. The output of the model is the pressure estimate for the node of our interest which is called the center node. The rest nodes in the sub-network are called the correlated nodes of the center node, and the pressure measurements at the correlated nodes constitute the input to the model. The average relative error of the model is found less than 3%. A pressure regulating method based on the model is proposed and tested numerically.In addition, a time series analysis based model and artificial neural networks are built for daily water demand forecast in a city. While the time series analysis based model extracts water demand patterns of different periodicities from historical data of water use and calibrates the residuals aside from the periodic patterns with the observed water demand, artificial neural networks directly approximate the relationship between the daily water demand and the related factors. Two different network architectures, traditional multilayer perceptron neural network and dynamic artificial neural network, are applied to develop the water demand forecast model. Network weights that connect neurons are updated regularly to adapt to the time-varying feature of water demand. At last, the prediction errors of different methods for water demand forecast are compared and it is concluded that the multilayer perceptron neural network outperform the rest, with an average relative error of 1.47%.