(Proceedings of symposium S7 held during the Seventh IAHS Scientific Assembly at Foz do Iguacu, Brazil, April, 2005, IAHS Publication. 303, 2006. pp. 5768).
We present PAI-OFF (Process modeling and Artificial Intelligence for Online Flood Forecasting), which combines the reliability of physically based, sophisticated modeling with the operational advantages of Artificial Neural Networks (ANN). Thus we are able to improve ANN performance in the flood forecasting context by detailed process modeling. Low computation times and robustness are the key features of ANN models and also form the basic requirements for flash flood forecasting. After presenting the theory of the new methodology, the results of a catchment related meteorological analysis for generating storm scenarios sere as the input to a coupled hydrological/hydraulic model, which is set up for a mountainous catchment in east Germany. Along these lines we operate the catchment model for all realistically possible constellations of flood formation. This results in a database consisting of corresponding input/output vectors. We complete the database for training the ANN by adding yet more flood relevant data for characterizing the hydrological and meteorological catchment situation prior to a storm event. After this preparatory step, the ANN is applied for online event, i.e. one which did not feature in the training process. The convincing agreement between the predicted and observed flood hydrograph underlines the application potential of the new PAI-OFF methodology for online flood forecasting even in smaller catchments.
Artificial neural networks, Flood forecasting, Network training
