| dc.description.abstract | The application of optimisation to Water Distribution Network (WDN) Modelling involves the use of computer-based techniques to many different problems, such as leakage detection and localisation. The success in the application of any model-based methodology for finding leaks highly depends on the availability of a well-calibrated model. Both leak detection and localisation, as well as model calibration are procedures that constitute the field of inverse problems in WDN modelling. The procedures are interlinked and dependent as when a leak is found and the model is updated its quality improves, while when a model is calibrated its ability to detect and localise leaks also improves. This is because both inverse problems are solved with the aim to mimic the behaviour of the real system as closely as possible using field measurements. In this research, both inverse problems are formulated as constrained optimisation problems. Evolutionary Optimisation techniques, of which Genetic Algorithms are the best-known examples, are search methods that are increasingly applied in WDN modelling with the aim to improve the quality of a solution for a given problem. This, ultimately, aids practitioners in these facets of management and operation of WDNs. Evolutionary Optimisation employs processes that mimic the biological process of natural selection and “survival of the fittest” in an artificial framework. Based on this philosophy a population of individual solutions to the problem is manipulated and, over time, “evolves” towards optimal solutions. However, such algorithms are characterised by large numbers of function evaluations. This, coupled with the computational complexity associated with the hydraulic simulation of WDNs incurs significant computational burden, can limit the applicability and scalability of this technology across the Water Industry. In addition, the inverse problem is often “ill-posed”. In practice, the ill-posed condition is typically manifested by the non-uniqueness of the problem solution and it is usually a consequence of inadequate quantity and/or quality of field observations. Accordingly, this thesis presents a methodology for applying Genetic Algorithms to solve leakage related inverse problems in WDN Modelling. A number of new procedures are presented for improving the performance of such algorithms when applied to the complex inverse problems of leak detection and localisation, as well as model calibration. A novel reformulation of the inverse problem is developed as part of a decision support framework that minimizes the impact of the inherent computational complexity and dimensionality of these problems. A search space reduction technique is proposed, i.e., a reduction in the number of possible solution combinations to the inverse problem, to improve its condition considering the accuracy of the available measurements. Eventually, this corresponds to a targeted starting point for initiating the search process and therefore more robust stochastic optimisations. The ultimate purpose is to increase the reliability of the WDN hydraulic model in localising leaks in real District Metered Areas, i.e., to reduce the number false positives. In addition, to speed up the leak search process (both computationally and physically) and, improve the overall model accuracy. A calibrated model of the WDN is not always available for supporting work at distribution mains level. Consequently, two separate problem-specific methods are proposed to meet the abovementioned purpose: (a) a Leak Inspection Method used for the detection and localisation of leaks and; (b) a Calibration Method for producing an accurate average day model that is fit for the purpose of leak detection and localisation. Both methods integrate a three-step Search Space Reduction stage, which is implemented before solving the inverse problem. The aim is to minimize the number of decision variables and the range of possible values, while trying to preserve the optimum solution, i.e., reduce the inverse problem dimensionality. The search space reduction technique is established to generate a reduced set of highly sensitive decision variables. Eventually this is done to provide a viable, scalable technique for accelerating evolutionary optimisation applications in inverse problems being worthwhile on both academic and practical grounds. The novel methodologies presented here for leak detection and localisation, as well as for model calibration are verified successfully on four case studies. The case studies include two real WDN examples with artificially generated data, which investigate the limits of each method separately. The other two case studies implement both methods on real District Metered Areas in the United Kingdom, firstly to calibrate the hydraulic network model and, then, to detect and localise a single leak event that has actually happened. The research results suggest that leaks and unknown closed or open throttle valves that cause a hydraulic impact larger than the sensor data error can be detected and localised with the proposed framework which solves the inverse problem after search space reduction. Moreover, the quality of solutions can dramatically improve for given runtime of the algorithm, as 99.99% of infeasible solution combinations are removed, compared to the case where no search space reduction is performed. The outcomes of the real cases show that the presented search space reduction technique can reduce the search area for finding the leak to within 10% of the WDN (by length). The framework can also contribute to more timely detection and localisation of leakage hotspots, thus reducing economic and environmental impacts. The optimisation model for predicting leakage hotspots can be effective despite the recognized challenges of model calibration and the physical measurement limitations from the pressure and flow field tests. | en_GB |
