Genetic Algorithms and Particle Filtering for Calibrating Water Demands and Locating Partially Closed Valves in Water Distribution Systems

Water distribution systems are constructed to supply water for domestic, industrial and commercial consumers. The design, operation and management of these distribution systems is usually supported by the application of hydraulic models, which are built to replicate the behavior of real systems. Conventional demand driven models simulate flows and pressures of a water distribution system requiring assumptions of known demands and known valve statuses. Due to the stochastic behavior of the water demands as well as the complexity of the piping network, these assumptions usually lead to an inadequate understanding of the full range of operational states in the water system. Installation of sensor devices in a network can provide information about some components in the system. However, calibration of the water demands and identification of valve statuses is either still not feasible or very difficult being attributable to the usual limited number of available measurement devices in most real water networks. This dissertation addresses three main issues of water distribution modelling, which include: (1) calibration of water demands under ill-posed conditions where the number of measurements is less than the number of parameter variables, (2) estimation of water demands under uncertainty in a near real-time context, and (3) calibration and localization of unknown partially/fully closed valves in a water network. The solutions for these problems, which are the main contributions of the research, are described by three journal papers included in this dissertation. The first journal paper presents a novel approach to calibration of the water demand multipliers under ill-posed (i.e. underdetermined) conditions by the multiple runs of a Genetic Algorithm model. The results from three case studies show that the average values of multiple runs of the Genetic Algorithm model can deliver very good estimates of the water demand multipliers, the flow rates and nodal heads at non-measured locations with a limited number of the measurements. In addition, the effects of the location and the number of measurement sites to the output of the demand calibration model are also analysed in the paper. The second journal paper introduces a predictor-corrector approach for the online (near real-time) estimation of demand multipliers. A conventional particle filter and an improved particle filter method, which incorporates an evolutionary scheme from Genetic Algorithms into the resampling process to prevent particle degeneracy, impoverishment and convergence problems, are investigated to implement the predictor-corrector approach. Furthermore, the paper proposes a first order approximation method to quantify the uncertainties of the model outputs caused by measurement errors. Two case studies are presented to demonstrate the effectiveness of the proposed particle filter model. The third journal paper proposes an innovative methodology for the identification of unknown partially/fully closed valves in a water distribution network. Three sequentially applied methods are proposed in the methodology, which include: a local sensitivity analysis, an application of Genetic Algorithms and an application of the Levenberg- Marquardt algorithm. In the first method, the sensitivity of the flow rates and nodal heads at measurement locations with respect to the change in the minor losses of the valves is computed. This computation is used to identify the valves that are unable to be localized by the measurement data. The second method applies a Genetic Algorithm in an extended period simulation in order to preliminarily identify the locations of the partially/fully closed valves and their setting values, i.e. the degree of opening of the valve. Finally, the application of the Levenberg-Marquardt algorithm to a steady state simulation is implemented to correct the results from the Genetic Algorithm model. This research has made significant contributions to the body of knowledge. Two novel methodologies have been developed for calibration of demands in water distribution systems. The impact of the location and number of measurement sites on the output of the demand calibration models has been evaluated in detail. In addition, a novel methodology for the calibration of unknown valve statuses has also been proposed. Results from realistic-size case studies have shown that the proposed methodologies are capable of solving real world problems, which enhances calibration approaches for water distribution system models.Thesis (Ph.D.) -- University of Adelaide, School of Civil, Environmental and Mining Engineering, 201