The use of narx neural networks to predict the timing and location of dike breaches during river floods

Generally, two dimensional hydraulic models are used to simulate the consequences of river flood events. However, the computational times of these models are commonly in the order of hours to days making them inappropriate to be used as an early flood forecasting system. Therefore, we studied whether neural networks, having computational times of less than a second, can be used to predict the locations and timing of potential dike breaches. The Dutch bifurcating system of the Rhine river delta was used as a case study. Training data was created with a 1D-2D coupled hydraulic model, in which the main channel and floodplains were schematized by 1D profiles and the hinterland was discretized on a 2D. The dike breach locations were assumed to fail if the simulated water level reached the dike crest levels. In total, 80 potential flood events were simulated by varying the upstream peak discharge and shape of the discharge waves. It was found that only two out of the 28 dike breach locations failed during one or more of the 80 simulated flood events. The discharge that left the river system through the dike breach, flowing into the hinterland, (i.e. the outflow hydrograph) was used to train the neural networks. NARX neural networks were developed since this type of neural network is capable of  predicting a time-varying output based on an input time series (Shen & Chang, 2013). For both dike breach locations, a NARX network was set up. It was found that the NARX networks always accurately predicted if the specific dike location failed. Furthermore, the timing of the dike breach and the outflow hydrographs were predicted accurately. Finally, the cumulative outflow flood volume was accurately predicted, only deviating 1.5%, which is highly relevant for the  prediction of inundation extents in the hinterland