An ensemble neural network model for real-time prediction of urban floods

The real-time forecasting of urban flooding is a challenging task for the following two reasons: (1) urban flooding is often characterized by short lead times, (2) the uncertainty in precipitation forecasting is usually high. Standard physically based numerical models are often too slow for the use in real-time forecasting systems. Data driven models have small computational costs and fast computation times and may be useful to overcome this problem. The present study presents an artificial neural network based model for the prediction of maximum water levels during a flash flood event. The challenge of finding a suitable structure for the neural network was solved with a new growing algorithm. The model is successfully tested for spatially uniformly distributed synthetic rain events in two real but slightly modified urban catchments with different surface slopes. The computation time of the model in the order of seconds and the accuracy of the results are convincing, which suggest that the method may be useful for real-time forecasts.