Ramon Perez Margrane
Identifiability and calibration of water network models
This thesis has been carried out with the collaboration of two research groups. One is more oriented to the application -water networks- and the other one is more oriented to the technology -control and supervision-. Experience of both groups has generated the necessity of calibration of water network models in order to be able to do good simulations, optimisations, supervisions, leak detections, etc. For this reason, the main objective of this thesis is to develop a closed methodology for the calibration process of water distribution network models.
The originality of this work comes both from the magnitude of the problem and the techniques used. On the one hand, a water network includes different elements (nodes, reservoirs, pipes, valves, and pumps), and its calibration requires the study of those elements in detail. On the other hand it has to be applied to huge systems. Special attention has been paid to the three main parts of calibration: identifiability study, macrocalibration and microcalibration. Each of those steps needs specific techniques. Some of the techniques used in this thesis are little known or unknown at all in the water industry.
The identifiability study has been developed for different case studies, ranging from simple, illustrating case to real huge networks. The simplest experiments were performed with linear and static networks. In general, networks are non-linear and the use of more than one time-step in the measurements provides better identifiability conditions. The methodology proposed allows the determination of the extended-period identifiability for general networks (non-linear). The obtained tool helps in the design of identification problems using topological information of the network.
When the model is generated, large errors are introduced. These errors are detected in a first calibration effort, macrocalibration. This process is done manually and the objective in this thesis is to give support to such work. The methodology followed by the experts has been analysed. Specific algorithms have been used in this thesis for each kind of error. In order to detect errors in huge amount of elements classification algorithms have been used. Those algorithms allow the generation of knowledge from simulation experiments and optimisation of likelihood functions.
The parameter tuning, microcalibration, is treated as an optimisation problem. The non-convexity of the problem is detected by a detailed characterisation. This non-convexity shows to be a problem for the local optimisers. The capabilities of some global optimisation algorithms have been explored. The computing cost for the global optimisers, especially when huge networks are identified, represents a major limitation. The Extended Kalman Filter has been used with promising results.
Full text
This thesis has been carried out with the collaboration of two research groups. One is more oriented to the application -water networks- and the other one is more oriented to the technology -control and supervision-. Experience of both groups has generated the necessity of calibration of water network models in order to be able to do good simulations, optimisations, supervisions, leak detections, etc. For this reason, the main objective of this thesis is to develop a closed methodology for the calibration process of water distribution network models.
The originality of this work comes both from the magnitude of the problem and the techniques used. On the one hand, a water network includes different elements (nodes, reservoirs, pipes, valves, and pumps), and its calibration requires the study of those elements in detail. On the other hand it has to be applied to huge systems. Special attention has been paid to the three main parts of calibration: identifiability study, macrocalibration and microcalibration. Each of those steps needs specific techniques. Some of the techniques used in this thesis are little known or unknown at all in the water industry.
The identifiability study has been developed for different case studies, ranging from simple, illustrating case to real huge networks. The simplest experiments were performed with linear and static networks. In general, networks are non-linear and the use of more than one time-step in the measurements provides better identifiability conditions. The methodology proposed allows the determination of the extended-period identifiability for general networks (non-linear). The obtained tool helps in the design of identification problems using topological information of the network.
When the model is generated, large errors are introduced. These errors are detected in a first calibration effort, macrocalibration. This process is done manually and the objective in this thesis is to give support to such work. The methodology followed by the experts has been analysed. Specific algorithms have been used in this thesis for each kind of error. In order to detect errors in huge amount of elements classification algorithms have been used. Those algorithms allow the generation of knowledge from simulation experiments and optimisation of likelihood functions.
The parameter tuning, microcalibration, is treated as an optimisation problem. The non-convexity of the problem is detected by a detailed characterisation. This non-convexity shows to be a problem for the local optimisers. The capabilities of some global optimisation algorithms have been explored. The computing cost for the global optimisers, especially when huge networks are identified, represents a major limitation. The Extended Kalman Filter has been used with promising results.
Full text