Continuous-time system identification from discrete-time measurements with application to natural gas pipeline modeling
Abstract
This work was motivated by the need to model a network of natural gas pipelines and its corresponding demand pipeline, in order to make predictions of the pressures at critical junctions in the network Development of such a model amounts to a system identification problem with limited information. In order to solve this problem, we developed a demand model that would provide estimates of the gas usage for the communities serviced by the pipeline network. The parameters of the demand model were estimated using an adaptive genetic algorithm. This new algorithm was first developed and compared with existing genetic algorithms. A discussion of the role played by crossover and mutation operators in the genetic algorithm was also presented. Based on the theory of gas dynamics and the known pipeline network topology, a resistor-capacitor network analog to the pipeline network was developed. The parameters of the resistor-capacitor model were estimated using ordinary least squares techniques. We first studied and developed a number principles and guidelines for a class of system identification problems. One of the main areas studied was the development of a generalized framework for least squares "parameter " identification of continuous-time systems from discrete-time measurements of the states of the continuous-time system. Subsequently, we extended our generalized framework to the least squares parameter identification of a class of resistor-capacitor networks. We also studied the effects on the estimated results of the integration scheme used in the process and the noise levels in the measured data. A demonstration of the benefits of the incorporation of the maximum available structural information of the system being modeled was also presented. Finally, we developed a set of guidelines for the required input signal frequencies and sampling frequencies to provide acceptable identification results for both the plant-model-match and reduced-order modeling problems. Finally, we applied these techniques to the identification of an actual natural gas pipeline network. The results provided significantly better pressure estimates than those previously reported.
This paper has been withdrawn.