Towards enhancing flood preparedness and early warning: a comparative study of 2D high-resolution simulations of the 2021 Ahr floods

The escalating frequency and severity of flash floods have heightened concerns, presenting a unique challenge compared to traditional fluvial flooding. Unlike large-scale fluvial events, managing flash flood risks through infrastructure approaches is less effective. Early warning systems emerge as crucial components in responding to these rapid and intense floods. For a response to be effective, an early warning system must provide stakeholders with sufficient actionable information about the imminent flood and lead time, underscoring the need for computational flood forecasting models.

Hydrodynamic models, which shine for their accuracy in predicting floods in complex topographies and urban environments affected by flash floods, face a drawback—they are computationally intensive, potentially limiting their application in early warning systems.

This contribution delves into the utilisation of two 2D flood forecasting models: the local-inertia solver RIM2D and the full shallow water equation solver SERGHEI. To minimise runtime, both solvers are implemented to run on GPUs, with a focus on maximising forecast lead time. RIM2D, less computationally intensive than SERGHEI, is expected to be well-suited for this purpose. On the other hand, to offset the higher computational cost, SERGHEI allows for multi-GPU use, specifically tailored for large-scale High-Performance Computing (HPC) systems.

The study assesses the applicability and trade-offs associated with these solvers, concentrating on the flood event in the Eifel in 2021, with a specific focus on the lower Ahr valley—from Altenahr to the Rhine. Simulations with identical conditions are conducted using both solvers, spanning resolutions from 1m to 10m. Evaluation criteria include accuracy in terms of maximum flood levels and computational performance in terms of required resources and runtime, and we explore the nature of the differences of the results produced by both solvers and their potential implications for flood forecasting and early warning.

Results indicate that at coarser resolutions, both solvers yield similar accuracy. Discrepancies emerge at higher resolutions due to the distinct mathematical formulations. Computational costs escalate rapidly with resolution for both solvers. Notably, for resolutions equal to or coarser than 5m, flood forecasts are at least 75 times faster than real-time. This efficiency makes them suitable for augmenting existing operational flood forecast systems but retaining excellent lead times, thus, enabling detailed flood impact forecasting and immediate responses. 

However, at higher resolutions, the computational demands exceed the capacity of a single scientific-grade GPU, necessitating multi-GPU implementations and some HPC capabilities for operational use. While such high(er) resolution models may seem excessive for managing specific flood events, they underscore the growing need for state-of-the-art scientific software and HPC technology in addressing larger flood domains.