Author(s): Napolitano; G. ; See; L. ; Savi; F.
Linked Author(s):
Keywords: Real-time forecasting; Flooding; Artificial Neural Networks
Abstract: An Artificial Neural Network (ANN) model, trained with backpropagation and Bayesian regularization, has been developed to forecast hourly water levels at Ripetta gauging station in Rome for a lead time of 12 hours. Rome is located at the lower part of the Tiber river basin and it is considered a flood-risk prone area for extreme events, characterized by a return period of about 200 years. The model relates the forecasted water levels in Rome to the water stage observed in three stream gauging stations located upstream. Rainfall in the middle and lower Tiber basin can also be considered as inputs to improve the model forecasts. The effects on the model results of using the net rainfall estimated by means of a hydrologic abstraction model instead of observed rainfalls are analysed and discussed. The ANN model proves to be reliable on the basis of absolute and relative performance measures as well as a visual inspection of the water level hydrographs.
Year: 2009