Author(s): Carcano Elena Carla; Bartolini Paolo; Muselli Marco; Croke Barry
Linked Author(s):
Keywords: Dvanced numerical modelling; Neural networks; Data management
Abstract: An analysis of potential scenarios of daily streamflows starting from rainfall, runoff and temperature series comparing Artificial Neural Networks (ANNs) to a conceptual model approach (Ihacres) is herein conducted. Particular attention is reserved to the reconstruction of hydrograph drought periods where water sources design and control are mostly needed. Concerning ANNs two connectionist models are introduced: a feedforward Multilayer Perceptron (MLP) is put aside to an hybrid version of Jordan Recurrent Neural Network in order to reproduce real world data coming from two small catchments (192 and 69 km2 of size) with torrential regime. Different networks configurations are explored in order to determine a good combination of input features (mostly runoff and rainfall data) able to catch the variability of Rainfall-Runoff process. Temporal dependence is introduce with traditional Tapped Delayed Line (TDL) technique and with memory effects. Results suggest that when poor measures are available total empirical models are able to stand the comparison to conceptual ones and, in general, to perform slightly better.
Year: 2007