Reliability of calibration of a monthly rainfall-runoff model: the semiarid case / Validité de l'étalonnage d'un modèle pluie-débit à pas de temps mensuel: cas d'un bassin semi aride

ABSTRACT

A monthly conceptual rainfall—runoff model that enjoys fairly widespread use in South Africa was calibrated for each of 50 calibration samples of lengths 3–20 years, drawn from a synthetic 101-year semiarid streamflow time series generated with the Stanford Watershed Model. The ability of Pitman's model to reconstruct the original 101 years of monthly streamflow for each of the 50 calibrations was then examined against a set of statistics of monthly and annual streamflows. The variabilities of key model parameters associated with different lengths of calibration period were also investigated. The results show that it is well worthwhile to increase the calibration period to about 15 years in order to reduce errors in reconstructed flow statistics. Merely increasing the length of calibration period from 6 to 10 years may decrease the error in most regenerated flow statistics by 30–50%. A fair amount of variability in “optimum” parameters, however, seems unavoidable, even at longer calibration periods, though this may also partly be due to imperfect model calibration. The effects of this variability are greatly attenuated in the reconstructed flow statistics due to parameter interdependence.