Parameter Optimization of Conceptual Hydrological Models

The form of modelling used in this research for the simulation of the rainfall/runoff regime of catchment areas by mathematical models is of particular importance to civil engineers in the building of dams, river bridges and other works affected by high and low flows in rivers and streams. The parametric conceptual models can also be used in the management of water resources and as a basis for the assessment of long term risks associated with water storage and transmission of supplies. The objectives of this research are to examine the problems arising from the conceptual modelling of catchment areas with large data sets, and the effective determination of model parameters using gradient and non-gradient optimization techniques in the field of hydrology. A simple model package was developed from the application and modification of ideas current at the time which allowed a good fit to observed hydrographs to be achieved with the input of rainfall data and data for an evaporation loss function. Nine parameters were available for optimization in this model. The practical demand for the assessment of land use and its variations on catchment water yield led to the development of a more complex model with thirty five parameters based on the latest vegetation process studies. One of the first modifications was to the criterion for convergence where it was changed from the rate of change of parameter values to that of the model coefficient of determination or efficiency of fit. The least squares objective function was investigated, and retained for model explained variance. However, for parameters involved in the simulation of base flows it was found to be more effective to use a proportional function, whilst for intense storm events an eighth power function exaggerated the information available in the data for determination of surface runoff parameters. The models employ an input data 'overlay' technique which allowed the use of large data sets running over many years. The simulation results from land use changes with large data sets from the highlands of Scotland, a clay catchment in Buckinghamshire and montane rain forest in Kenya are compared and contrasted for both models. The results for these catchments using gradient and non-gradient optimization algorithms are also examined, including the use of a genetic algorithm, and recommendations made for the values of algorithm parameters. Hybridized algorithms are developed and tested. A combination of the Rosenbrock and Nelder and Mead Simplex techniques was found to be an efficient hybrid; particularly with the land use model