Towards 2D flood forecasting with the HPC-enabled shallow water solver SERGHEI-SWE

Great advancement has been achieved in the last decade in 2D shallow water solvers for flood modelling. However, their application to physically-based flood forecasting continues to be experimental and not widespread. One of the central challenges towards operational flood forecasting with 2D solvers is their computational cost, which needs to be reconciled with the required lead times for forecasts to be of use. Nonetheless, these solvers have great potential to improve flood forecasting predictions, especially when it comes to flash floods, for which the established 1D and conceptual models may be significantly less applicable.

The shallow water solver SERGHEI-SWE leverages on robust and efficient numerical techniques and is implemented for High Performance Computing (HPC), allowing its use in supercomputers and opening new opportunities in 2D flood forecasting. In this contribution, we present proof-of-concept simulations of several flood events in different catchment and river systems. We show that, with SERGHEI-SWE, it is possible to run very high resolution flood simulations for large hydrological systems with runtimes significantly lower than the event duration. This property is essential to enable operational forecasting with useful lead times.

We run simulations on three river reaches, in the Italian river Po (125 km reach between Boretto and Pontelagoscuro) and in one of its tributaries, the river Secchia (20 km reach), and a meandering reach of the Ebro river through the city of Zaragoza. We also perform flash flood simulations on a 5 km² district of Nice (France), and in a 50 km² agricultural catchment in Jaén (Spain). The focus of the exercise is on the computational performance aspect and not on the model performance. The results show that high resolution simulations can be done with runtimes in the order of 100 times faster than real time, potentially allowing a very good forecast lead time. We also explore different combinations of computational resources, model resolution and ensemble size to explore the flexibility of the modelling approach under different computational systems, which may be available for flood forecasting.