Issue 7

Depicting hydrodynamic processes in flooding with artifical neural networks under consideration of the rainfall-runoff processes in the intermediate watershed

by Ronny Peters (2008)

Summary: Due to shortcomings of existing flood-routing models for forecasting, an alternative method combining the speed and robustness of artifical neural networks with the reliability of hydrodynamic-numerical modelling is developed in this work. One-dimensional, hydrodynamic models incorporate excact knowledge of river bed and foreshore geometry and consider the physical processes of wave propagation. Such a deterministic model is used as a basis for comprehensive scenario calculations to generate a data basis covering the wide range of theoretically possible flood events. This data basis is used for training of artificial neural networks that can thus generate forecasts even for extreme flood events. This work examines and evaluates the performance of both supervised and unsupervised training networks, namely multilayer feedforward networks and selforganizing maps. Furthermore, the method was enhanced to include features of rainfall-runoff processes in the ungauged intermediate watershed for the consideration of lateral flows along the modelled flow paths. For this, the data basis was generated by a rainfall-runoff model. The transfer of input data to characteristic features for the mapping of target variables, namely discharge and water level at the target gauge, is of major importance. The deterministic models are not only used for the generation of a reliable data basis for the network training, but also allow sensitivity analysis of modelling results due to changes in input variables for both the rainfall-runoff processes as well as the hydrodynamic processes involved. This analysis helps in uncovering the most relevant features, as the introduction of a single state feature that consolidates the total meteorological history of the event for the characterization of the watershed condition is key to the successful integration of the rainfall-runoff processes into the forecast network. The methodology developed here was successfully tested using the Freiberger Mulde watershed.