Recurrent neural networks in Rainfall-Runoff modeling at daily scale

This work aims to simulate potential scenarios in Rainfall-Runoff (R-R) transformation at daily scale, mainly perceived for the control and management of water resources, using feed-forward multilayer perceptrons (MLP) and, subsequently, Jordan Recurrent Neural Networks (JNN). R-R transformation is one of the most complex issue in hydrological environment due to high temporal and spatial variability; very strong and non linear interconnections among variables: a good challenge for Artificial Neural Networks (ANN). Abilities and limitations of MLP and JNN models have been investigated, especially focusing on drought periods where water resources management and control are particulary needed. The study compares the results of the two networks typologies to outputs from a conceptual linear model and then to physical context of two small Ligurian catchments. It also demonstrates the remarkable improvement obtained with the JNN approach especially when rainfall memory effect is employed as an additional input.