Downsizing parameter ensembles for simulations of rare floods

Abstract

Abstract. For extreme-flood estimation, simulation-based approaches represent an interesting alternative to purely statistical approaches, particularly if hydrograph shapes are required. Such simulation-based methods are adapted within continuous simulation frameworks that rely on statistical analyses of continuous streamflow time series derived from a hydrological model fed with long precipitation time series. These frameworks are, however, affected by high computational demands, particularly if floods with return periods > 1000 years are of interest or if modelling uncertainty due to different sources (meteorological input or hydrological model) is to be quantified. Here, we propose three methods for reducing the computational requirements for the hydrological simulations for extreme-flood estimation so that long streamflow time series can be analysed at a reduced computational cost. These methods rely on simulation of annual maxima and on analysing their simulated range to downsize the hydrological parameter ensemble to a small number suitable for continuous simulation frameworks. The methods are tested in a Swiss catchment with 10 000 years of synthetic streamflow data simulated thanks to a weather generator. Our results demonstrate the reliability of the proposed downsizing methods for robust simulations of rare floods with uncertainty. The methods are readily transferable to other situations where ensemble simulations are needed.

Clustering model responses in the frequency space for improved simulation‐based flood risk studies: The role of a cluster number

“…As a result of clustering, elements belonging to the same cluster are more similar to each other than to the elements from other clusters (Notaro et al, 2018). Clustering has been continuously gaining in popularity in hydrologic applications and has been successfully applied to describe catchment functionality (Jehn et al, 2020; Kuentz et al, 2017), to regionalization and predictions in ungauged catchments (Rao & Srinivas, 2006; Singh et al, 2016), or for reducing model computational requirements (Ehret et al, 2020; Notaro et al, 2018; Sikorska‐Senoner et al, 2020). The latter work has been the first application of the cluster analysis in the FFS.…”

“…A common way to consider uncertainty of hydrologic models is through resampling from the optimized parameter space (Sikorska & Renard, 2017) or via using multiple parameter sets derived from independent calibration runs (Westerberg et al, 2020). While the first approach requires defining a likelihood function to sufficiently sample from the posterior (Montanari & Koutsoyiannis, 2012; Sikorska et al, 2015), the latter approach requires running multiple calibration runs to minimize the risk of being trapped in a local minimum (Sikorska‐Senoner et al, 2020). The set of resulting model simulations can be used as an ensemble of hydrologic responses.…”

“…While a reduction of meteorological scenarios or initial state condition set‐ups is not straightforward due to a nonlinear input–output relationship in hydrologic models, a reduction of the parameter sample size seems to be a conceptually easier way to lower the computational needs of simulation‐based PFRA. Recently, Sikorska‐Senoner et al (2020) proposed a very promising approach for such a parameter ensemble downsizing via direct analysis of hydrologic responses in the application space (flood frequency space [FFS]). Based on this, representative parameter sets (RPSs) that best represent the full hydrologic ensemble can be selected.…”

Improving the Pareto Frontier in multi-dataset calibration of hydrological models using metaheuristics

“…One of the keys to a good GA performance is to have a crossover function that fits the problem and the search space. GA based optimization techniques have been frequently applied in hydrology (Zhu and Wu, 2013;Brunner et al, 2018;Sikorska et al, 2018;Van Tiel et al, 2018;Sikorska-Senoner et al, 2020).…”

“…In contrast to the above, the second group is based on likelihood-free optimization techniques that do not require making any assumptions on model error properties. The identifiability issue of model parameters is dealt with via using multiple parameter sets as a proxy of model uncertainty (Sikorska-Senoner et al, 2020). The likelihoodfree optimization techniques include metaheuristics methods that rely on learning strategies to find near-optimal solutions (Maier et al, 2014).…”