Can rainfall-runoff models provide accurate estimates of design flood quantiles in ungauged catchments?

Abstract

Accurate estimation of design floods is one of the most important tasks of applied hydrology, since design of hydraulic structures and flood protection measures heavily depends on these estimates. In hydrologically gauged catchments these estimates are usually obtained by applying frequency analyses over a series of annual maxima. In case of ungauged catchments, design floods are commonly estimated by applying an event-based rainfall-runoff models with design rainfall hyetographs (hereafter referred to as models). Although the design flood estimates obtained in this way are rather sensitive to every element of these models (e.g., design rainfall duration or unit hydrograph method), there is no specific guidance on the modelling decisions to obtain reliable design flood estimates. Robustness of a particular model is evaluated in gauged catchments (where observed data are available), by comparing design floods obtained with the model, to the results of the frequency analysis. The agreement between the two design flood estimates is usually quantified in terms of relative errors, which do not take into account uncertainty in the quantiles. In this paper, we propose a complementary approach to evaluation of event-based models, which implies that design flood estimates are compared to the confidence intervals of the quantiles. This approach is applied to nine models used to simulate design floods of 20-, 50- and 100-year return periods at the location of Zavlaka stream gauge on the Jadar River. The results show that taking quantile confidence intervals into consideration can provide additional insights in model performance, and, thus, should be mandatorily included in the model evaluation. Comparison across the models reveals considerable sensitivity of the design flood estimates to the models. Models that use daily design rainfall of uniform intensity systematically underestimate corresponding quantiles, while the estimates obtained with shorter rainfall durations or with rainfall of non-uniform intensities result in errors of both signs. Equifinality between the curve number and rainfall duration is also detected in the results. This study exposes challenges in evaluating event-based models, and emphasises the need for specific guidance on the application of these models for design flood estimation.