A framework to characterize flood events of defined return period ranges using functional boxplots

To assess the safety of dams, design floods are typically used as a basis. Of particular interest are events with a return period of 1’000 years and even rarer events derived from that with help of simple return period conversion factors given by design codes. However, both the peaks and even more the flood volumes of such rare events are subject to large uncertainties due to limited length and spatial coverage of gauge records. Bivariate approaches can help reduce the uncertainty related to the flood volumes. Nevertheless, both univariate and bivariate approaches require long-term observations on which the return periods of flood events can be calculated.

In this study, we make use of very long simulated hydrographs in hourly resolution for Swiss catchments (scale range: ~300–18’000 km²). The hydrographs span about 300’000 years each and stem from a hydro-meteorological modelling chain starting with a stochastic multi-site weather generator. With these hydrographs, we develop a framework to characterize design floods through a realistic hydrograph using functional data analysis as well as hydrographs that envelope 50%, say, of the most central observations (corresponding to the 25% and 75% quantiles in a univariate setting).

In a first step, we assigned the simulated annual maximum flood events to return period classes of 100, 1000 and 10000 years. We then built clusters of similar events within each class using functional clustering. Here we explore some of the possibilities of the approach and in particular show how sensitive the functional clustering is to the choice 1) of event characterization (peak only, flood peak and volume, flood volume given a minimum flood peak), and 2) to the separation of event and baseflow of the selected events in the bivariate case and 3) to the different functional latent mixture models that are applied within the functional clustering.