Computationally efficient lumped floodplain modelling using data-driven methods

Modelling floodplains adequately is crucial for numerous water management applications. Many of these require a large number of simulations or long term calculations, thereby necessitating the use of models with a limited calculation time. This paper investigates two different flexible data-driven methods for computationally efficient lumped floodplain modelling that predict the inundation level in the floodplain and the flow between the river and the floodplain for a given set of river stages. The flow can be calculated using a set of optimized discharge equations. Alternatively, a hybrid approach can be followed that combines physical principles with neural networks that rely on water levels to obtain flow estimates. The derived methodology was tested on a case study using simulation results from a detailed MIKE11 model that solves the full de Saint-Venant equations. The results show that both approaches deliver accurate results, although the neural networks outperform the fixed weir equations. At the same time, the simulations are more than 10.000 faster compared to the original full hydrodynamic model. The obtained floodplain models can be used for various applications that require very short computation times.status: publishe