On the optimisation of numerical weather prediction model configuration for improved flood forecasting

Hydrological forecasts are essential for the timely and accurate prediction of flooding events, which are among the most impactful natural hazards for both infrastructure and human life in Europe and many other regions worldwide. Most existing flood warning systems are supported by hydrological models. Their accuracy depends not only on the representativeness and proper calibration (when required) of the model itself, but also on the quality of its inputs. While static inputs, particularly soil parameters, are highly uncertain, weather forecasts are arguably the most influential drivers.

In this study, we recreate the entire operational modelling framework used in Switzerland. Weather forecasts are provided by ICON (MeteoSwiss) and are used as input for WaSiM (FOEN), which produces streamflow predictions and issues warnings when necessary. We focus on several case studies, including selected catchments (e.g., Thur) and historical events that exceeded national flood warning levels (e.g., 30 May–2 June 2024).

This setup allows us to experiment with different configurations of the numerical weather prediction (NWP) model and to assess their downstream impacts on hydrological forecasts. We test different lead times to evaluate how early flood peaks can be detected, varying ensemble sizes to determine how many members are required to capture “extreme” flooding scenarios, and different spatial resolutions (500m – 2km) to assess the impact of resolving small-scale processes (e.g., convection).

Model performance is evaluated using classical hydrological metrics (NSE, KGE, RMSE, etc.), as well as more operationally relevant metrics for warning systems, such as whether thresholds are exceeded, how early exceedances occur, and their duration. Finally, we test different products for initializing model runs, either interpolated station-based products or NWP analysis products and assess the influence of the hydrological model itself through a sensitivity analysis of its parameters.

The results of this study will shed light on how NWP model configurations affect flood forecasting and, in turn, improve flood early warning design and decision-making.