On the propagation of errors in catchment modelling systems

Abstract

The application of catchment modelling systems is now a common approach for management of catchments. Fundamental to the application of a catchment modelling system is the calibration and validation of the many control parameters used to ensure that the predicted catchment response adequately reproduces the actual catchment response. The calibration process, in general, consists of the systematic variation of control parameter values until a set of values is obtained that results in the adequate reproduction of the recorded catchment response. While this systematic variation may be undertaken manually, there have been a number of automatic calibration techniques which are based on the minimisation of differences between the predicted and recorded catchment response. Implicit in many of these techniques is the assumption that the residuals (ie variation between the predicted and the recorded catchment response) are independent, homoscedastic and normally distributed. Presented herein are the results of an investigation into these assumptions using the Powells Creek catchment in Sydney, Australia as a test catchment. It was found that these assumptions were not achievable on this typical catchment, and therefore a range of possible corrective measures were reviewed and tested. Results obtained from testing of these possible corrective measures are presented also.