A new system for online flood forecasting : performance and implications

Flood forecasting, as well as flood risk management at the operational level, become more and more relevant. The required reliability and robustness for operational flood warning systems, which up to now present the basic problems, are accounted for by the PAI-OFF (Process Modelling and Artificial Intelligence for Online Flood Forecasting) approach. It is based on the operational advantages of artificial neural networks. The system integrates all available physical information with the aid of a training procedure, originating from a physically based hydrological model. The forecast reliability of the new approach strongly depends on the catchment models' ability to realistically portray the flood relevant processes. In this paper we present two different approaches for modelling flood peaks with WaSiM-ETH, focusing on the parameterisation strategy. In order to improve model efficiency we propose a dual parameterisation methodology. This approach allows for setting model parameters according to the dominant controls of floods of different magnitudes. Results from the study are demonstrated for a catchment in the Erzgebirge (Ore-mountains) in East Germany (1700 km 2 ). Online flood forecasting of the Zschopau River at the gauge Kriebstein is validated using PAI-OFF to predict the 2002 extreme flood event. This data did not feature in the training process of the PAI-OFF-PoNN (Polynomial Neural Network) forecast tool. The computational efficiency, together with the convincing agreement between the predicted and observed flood hydrographs, underlines the potential of the new PAI-OFF methodology in the context of operational online forecasting.