Leveraging reforecasts for flood estimation with long continuous simulation: a proof-of-concept study

Abstract

Flood estimation is critical for risk assessment, but traditional approaches are often constrained by the limited length of observational records. This study explores the potential of reforecasts (RFs) for flood estimation using long continuous simulation (CS) with a hydrological model at high (hourly) temporal resolution. As a proof of concept, we first processed individual RFs from the extensive archive of the European Centre for Medium-Range Weather Forecasts (ECMWF) with bias correction, stochastic downscaling and disaggregation with analogs to obtain mean areal precipitation and mean areal temperature for a set of test catchments in Switzerland. We subsequently concatenated these RFs into a time series of nearly 10 000 years and used them in long CS to derive flood return levels. Results show hydrological consistency of the concatenated RFs and demonstrate their potential in flood estimation, providing information on the magnitude and frequency of extreme events. In addition, RFs offer a relevant complementary perspective on exceptionally large floods when compared with estimates derived from long CS driven by other forcing inputs, such as a stochastic weather generator. However, structural uncertainties – particularly related to the underlying numerical weather prediction system – must be considered, along with the reliance on a single model framework, non-stationarity and internal climate variability. Further limitations arise for catchments smaller than approximately 500 km2, for which stochastic downscaling becomes increasingly inadequate. For resolving the relevant convective events in such catchments, dynamical downscaling would be more appropriate; however, this was not feasible with the currently available data.