Return to search

Parameter estimation using a genetic algorithm for complex catchment modelling systems.

Implementation of physically distributed catchment modelling systems reshapes the fundamental philosophy of traditional calibration approaches by supporting the concept of equifinality. Arising from the concept of equifinality, alternative behavioural parameter sets within a given catchment modelling system structure can generate similar levels of simulation performance. This concept is motivated by the existence of a variety of uncertainties associated with a complex catchment modelling system, such as an imperfect model structure, measurement errors in both the input data and the recorded flows, and unknown, or poorly defined, interactions among parameters. However, the difficulty of searching for behavioural parameter sets increases as the complexity of the catchment modelling systems increases. This study undertook an investigation on the feasibility and robustness of a real-value coding genetic algorithm (GA) for calibrating the physically distributed Storm Water Management Model (SWMM) using the Centennial Park catchment in Sydney as a case study. It was found that a real-value coding GA was a robust technique suitable to search for behavioural parameter sets and, in particular, it was found that this approach was capable of identifying the promising range of values for spatially variable parameters. Moreover, the widespread use of physically distributed catchment modelling systems has highlighted the importance of estimating the uncertainty in the parameter values and in the predictions obtained from a complex catchment modelling system as well as in catchment averaged, or lumped, systems that have been the focus of many previous studies. Bayesian inference has been shown to be a tool suitable for parameter uncertainty estimation in catchment modelling. However, the application of Bayesian inference faces difficulties in complex high-dimensional systems where there is little if any a priori knowledge about the proposal distribution of the parameters. In this study, a real-value coding GA was used to undertake uncertainty estimation on spatially variable control parameters with little a priori knowledge about the proposal distribution of parameters. After 50,000 evaluations, the marginal posterior distributions of spatially variable parameters which are associated with behavioural parameter sets were identified. The performance of a behavioural parameter set under a range of hydrological conditions was evaluated. Updating of the marginal distributions of these control parameters was implemented by adding additional calibration data. Interactions among the spatially variable control parameters were investigated also. Results based on the Pearson Correlation method indicate no clear relationship between any two control parameters. However, a methodology to detect relationships among groups of parameters was developed. Application of this methodology suggests that the simulation performance of SWMM was influenced by combinations of parameter values rather than values of the individual parameters. Finally, the predictive uncertainty associated with the existence of behavioural parameter sets was considered. A number of alternative strategies were used to evaluate the predictive performance. Consideration of the results suggests that use of a small number of parameter sets randomly selected from the large number of behavioural parameter sets was the best strategy in terms of efficiently obtaining predictive performance.

Identiferoai:union.ndltd.org:ADTP/225838
Date January 2007
CreatorsFang, Tianjun, Civil & Environmental Engineering, Faculty of Engineering, UNSW
Source SetsAustraliasian Digital Theses Program
LanguageEnglish
Detected LanguageEnglish
Rightshttp://unsworks.unsw.edu.au/copyright, http://unsworks.unsw.edu.au/copyright

Page generated in 0.0019 seconds