FORECASTING FLOOD HYDROGRAPHS at TIBER RIVER BASIN in ITALY by ARTIFICIAL NEURAL NETWORKS G. TAYFUR Prof. Dr.,Dept. Civil Engineering, Izmir Institute of Technology, Gülbahçe Campus, Urla, 35340 İzmir, Turkey (gokmentayfur@iyte.edu.tr ) T. MORAMARCO Res., IRPI-CNR, Via Madonna Alta 126, 06128 Perugia, Italy (t.moramarco@irpi.cnr.it) ABSTRACT Tracing the flood waves in natural rivers in order to mitigate flood damages is a challenging problem for engineers. To this end, engineers have strived to develop hydrologic and hydraulic methods to predict flood hydrographs. Hydraulic methods are based on solving flood wave equations that require complex numerical tech- niques while the hydrologic methods route flood wave using only the conservation of mass principle. Alternatively, this study developed artificial neural network (ANN) to forecast flood waves in natural channels. The developed ANN is based on three layer feed forward network using back propagation training algorithm, and sigmoid activation function. The model is applied to 4-h, 8-h, and 12-h lead time flood forecasting at three different stations located on Tiber River in Upper Tiber River Basin in central Italy. The network uses flow stages at upstream and down- stream stations in the input vector to predict flow discharge at the downstream sta- tion. For each case, six different hydrographs were used to train the network and four other hydrographs were used for testing the model. The model showed satisfac- tory performance in forecasting the hydrographs. Its performance was also com- pared to that of Muskingum model which is limited to floods whose wave travel time is equal or shorter than the lead time. ANN produced better performance with minimum errors. Keywords: Forecasting, Hydrograph, Flood, Basin, Modeling INTRODUCTION Forecasting floods is a major task to protect human life and as well as surround- ings around the natural rivers. Engineers have developed mainly hydrologic and hydraulic methods for this purpose. Hydrologic methods are based on conservation of mass principle. They basically assume linearity and therefore they are limited in