Abstract. Model benchmarking is needed in order to establish how newly developed forecasting approaches perform against current state-of-the-art systems. In many cases, resources for re-forecasting long periods of time are limited and therefore, a period of parallel-operations is evaluated. For this study, the forecasting chain presented in the companion paper by Antonetti et al. (2018) has been set up for the Verzasca basins in the southern Swiss Alps. In this region, an operationally running system is available from previous studies on probabilistic flash flood (FF) forecasts. This current system relies on the calibrated semi-distributed hydrological model PREVAH. The new model RGM-PRO includes the concept of dominant runoff processes and requires a priori estimation of parameters but no direct discharge observations for calibration. This is a significant benefit to FF prediction in ungauged catchments.

Both FF forecasting chains are forced by information from numerical weather prediction COSMO-1 and COSMO-E. Real-time rainfall is provided by the CombiPrecip product, which combines rain gauge and weather radar data. As RGM-PRO is an event-based model, initial conditions are not computed internally. Such initial conditions are obtained from operationally available gridded simulations of the PREVAH model. The current PREVAH-HRU setup uses rainfall data as obtained by interpolating real-time data of the station network maintained by MeteoSwiss. Initial conditions are tracked internally day-by-day.

The PREVAH-HRU runs forced by COSMO-1 and COSMO-E during real-time operations in the period May to August 2016 have been compiled. Corresponding model runs using RGM-PRO have been computed a posteriori. Both sets of forecasts are evaluated against discharge observations using deterministic and probabilistic verification metrics.

Results showed that the novel approach was able to compete with the operational benchmark prediction system and was consistently superior for high-flow situations. The new forecasting chains were able to react faster on precipitation in comparison with the benchmark forecasts. Confirming previous studies for all forecasting chains, a clear preference for using a meteorological ensemble as forcing data was found. In a synthesis of the two companion papers, more skill was found in the Verzasca basin than in the Emme catchment, suggesting a better forecast performance in strongly topography driven basins with shallow soils and weak dependence on initial conditions.

The findings of the two studies suggest that the novel forecasting chains can compete with the traditional ones in operational setup without the need of long-term discharge measurements and extensive calibration. With the new runoff generation module, extension of FF prediction to ungauged catchments is possible, provided that spatially distributed information on dominant runoff processes is available.