Coupling high-resolution precipitation forecasts and discharge predictions to evaluate the impact of spatial uncertainty in numerical weather prediction model outputs

Abstract

River hydrograph forecasts are highly sensitive to the space–time variability of the meteorological inputs, particularly in the case of watersheds characterised by a complex topography and whose hydrological processes are simulated by means of distributed rainfall–runoff models. An accurate representation of the space–time structure of the event that might occur is, therefore, essential when atmospheric and hydrological models are coupled in order to achieve successful streamflow predictions for medium-sized catchments. Even though the scale compatibility between atmospheric and hydrological models no longer seems to represent a serious problem for a direct one-way coupling, the quality and the reliability of deterministic quantitative precipitation forecasts (QPFs) are often unsatisfactory in driving hydrological models. This is because uncertainties in QPFs are, nowadays, still considerable at the scales of interest for hydrological purposes. In this work, different configurations of the non-hydrostatic meteorological model Lokal Modell (LM) have been tested for four rain events, with the aim of improving the description of the phenomena related to the precipitation. Then, LM QPFs have been coupled with the distributed rainfall–runoff model TOPKAPI, in order to assess the results in terms of discharge forecast over the Reno river basin, a medium-sized catchment in northern Italy. The coupling of atmospheric and hydrological models offers a complementary tool to evaluate the meteorological model performance. In addition, an empirical approach is proposed in order to take into account the spatial uncertainty affecting the precipitation forecast. The methodology is based on an ensemble of future rainfall scenarios, which is built by shifting in eight different directions the precipitation patterns forecasted by LM. An ensemble of discharge forecasts is then generated by feeding the hydrological model with these rain time series, thus, enabling a probabilistic approach for flood management. This empirical methodology has been tested for the same four events. Moreover, a statistical analysis has been performed on the discharge predictions driven by the operational version of LM for the autumn seasons in the years 2003–2005, with the aim of assessing systematic model deficiencies. In particular, this investigation aims to define, in terms of the number of LM grid points, the spatial shift more suitable for taking into account the spatial uncertainty of the precipitation forecast, and to establish which shift directions would improve the forecast over the area of interest. The results show that the very-high-resolution (2.8-km) configuration of LM, where an explicit description of the deep precipitating convection is adopted, improves the rainfall forecast considerably, both in terms of the total amount and the timing. Furthermore, the south, west and south-west shift-adjusted scenarios turn out to be useful in improving the discharge forecast for the Reno river basin when the main error source which affects the LM forecast is the spatial localisation of the precipitation field.