Kenechukwu Okoli, Luigia Brandimarte, Francesco Liao, Anna Botto, Giuliano Di Baldassarre
Tuesday 30 june 2015
14:05 - 14:20h at North America (level 0)
Themes: (T) Extreme events, natural variability and climate change, (ST) Hydrological extremes: floods and droughts
Parallel session: 6I. Extreme events - Flood Drought
This study investigated the performance of “model averaging” on the estimation of the design floods in view of uncertainty in model choice (i.e., probability distribution), as well as inaccuracy of hydrological data (i.e., times series of annual maximum flows). Two model averaging techniques were tested: i) simple model averaging, whereby a number of probability distribution models are used to infer the data and the design flood is computed as a simple average, and (ii) weighted model averaging, whereby the estimates provided by the diverse probability models are combined by giving higher weights to the distribution models that fit better the hydrological data. Model averaging outcomes were also compared to the results of model selection, whereby a single best probability model was selected by means of the Aikaike Information Criterion (AIC). In particular, numerical experiments were carried out by generating synthetic time series of annual maximum flows using the Wakeby probability model as parent distribution. For this study, comparisons were made in terms of relative errors and referring to the 1-in-100 year flood, i.e the quantile corresponding to a return period of 100 years. Weighted model averaging and simple model averaging showed a similar level of performance in improving the estimate of design floods. Interestingly, for shorter sample size drawn from a highly skewed population (not rare in hydrology), the experiments showed that model averaging (that might lead to over-fitting) improves design flood estimates even for short sample size in comparison to blindly selecting the parsimonious Gumbel distribution (also known as EV1).