Seamless rainfall and discharge forecasting using a scale-dependent blending of ensemble rainfall nowcasts and NWP

Flash flood early warning requires accurate rainfall forecasts with a high spatial and temporal resolution. As the first few hours ahead are already not sufficiently well captured by the rainfall forecasts of numerical weather prediction (NWP) models, rainfall nowcasting can provide an alternative. This observation-based method, however, quickly loses skill after the first few hours of the forecast due to growth and dissipation processes that are not accounted for. In addition, providing an additional forecasting method can let users drown in the amount of available information. A promising way forward is a seamless forecasting system, which combines the aforementioned forecasting methods. By optimally combining (blending) rainfall nowcasts with NWP forecasts, we can extend the skillful lead time of short-term rainfall forecasts and provide users with more consistent, seamless forecasts.

We implemented an adaptive scale-dependent ensemble blending method in the open-source pysteps library. In this implementation, the blending of the extrapolation (ensemble) nowcast, (ensemble) NWP and noise components is performed level-by-level, which means that the blending weights vary per spatial cascade level. These scale-dependent blending weights are computed from the recent skill of the forecast components, and converge to a climatological value, which is computed from a multi-day rolling window and can be adjusted to the (operational) needs of the user. To constrain the (dis)appearance of rain in the ensemble members to regions around the rainy areas, we have developed a Lagrangian blended probability matching scheme and incremental masking strategy.

We evaluate the method using three heavy and extreme (July 2021) rainfall events in four Belgian and Dutch catchments, focusing on both the rainfall forecasts and the resulting discharge forecasts using the fully distributed wflow_sbm hydrological model. We benchmark the results of the 48-member blended forecasts against the deterministic Belgian NWP forecast, a 48-member nowcast and a simple 48-member linear blending approach. When focusing on the resulting rainfall forecasts, the introduced blending approach predominantly performs similarly or better than only nowcasting (in terms of event-averaged CRPS and CSI values) and adds value compared to NWP for the first hours of the forecast. This holds for both the radar domain and catchment scale, although the difference, particularly with the linear blending method, reduces when we focus on catchment-average cumulative rainfall sums instead of instantaneous rainfall rates. We find similar results for the resulting discharge forecasts, although the effect of the catchment size and corresponding lag times becomes influential and determines the added value of nowcasting over NWP. By properly combining observations and NWP forecasts, blending methods such as these are a crucial component of seamless hydrometeorological forecasting systems.