Catchment scale hydrodynamic flash flood simulation for early warning: insights from the 2021 Ahr flood event.

The 2021 Ahr Valley flood during storm Bernd exemplifies the severity of flash floods and the challenges in flood risk and emergency management. This event underscores the growing threat of flash floods, even in regions where they are not traditionally considered common. The sudden, localized nature of flash floods makes early warning systems (EWS) critical. Failures in EWS played a relevant role during the Ahr floods and have likely played roles in other catastrophic events, such as the 2023 Libya floods under storm Daniel and the October 2024 floods in Valencia. Effective warnings require better, actionable information from flash flood models, a challenge due to the rapid onset of such events, the high resolution needed, and significant computational demands.

This study examines the application of the fully dynamic 2D shallow water solver SERGHEI, specifically designed for multi-GPU systems in large-scale High-Performance Computing environments. The focus is on simulating the rainfall-runoff process and subsequent flooding during the 2021 Ahr floods.

In earlier work, we explored flood propagation dynamics in the lower Ahr valley using SERGHEI at a very high resolution of 1m. We showed that simulations could be performed quickly enough for early warning, but with two key limitations. Firstly, since the domain only includes the lower valley, an inflow hydrograph is required at the upstream boundary to force the model. When performing forecasts for early warning, such inflow hydrograph would need to be generated by some hydrological model for the catchment down to the point of inflow, thus requiring a modelling chain. Second, the domain of interest for flood impact modelling, and thus the location for the hydrograph generation, is a priori unknown.

To address these limitations, we scale up the simulation by simultaneously modelling runoff generation and flood propagation over the entire catchment (900 km2). We perform SERGHEI simulations informing the model with a 1m resolution DTM, and openly available land cover, land use and soil data to parametrise hydraulic roughness and infiltration processes. The model is forced using radar precipitation measurements (1km spatial resolution 5 minutes temporal resolution). The target simulation resolution is 1m, leading to a computational grid of ~900 million cells, requiring 128 A100 GPUs in the JUWELS supercomputer, running roughly 5x faster-than-real-time. To perform sensitivity analysis to the infiltration and roughness parameters, we perform simulations at 5m resolution, for which the 36 million cell domain only required 16 GPUs to perform computations ~45x faster than real time. We also explore other resolutions to understand the effects of resolution on the quality of the forecast, computational resources and attainable lead time.

The results show the tradeoffs among modelling approaches for this event and demonstrate the feasibility and advantages of this approach for early warning in flash flood events. They underscore the maturity of the technology and provide strong arguments for using it to augment existing operational flood forecasts, while still achieving excellent lead times and far better detailed flood impact forecasting.