Neural nets for modelling rainfall-runoff transformations

To obtain river flow data, a neural network (NN) is developed and applied to rainfall-runoff transformation. The NN has been built considering a hidden two layer net and the sigmoidal has been used as a response function. Training is conducted using a back-propagation learning rule. In the input layer, both areal and point data values may be considered. The capability to provide a suitable forecast of river runoff has been examined for the Araxisi watershed in Sardinia. Experiments have been made dividing the total extension of observed data into three ten-year periods, assuming each as a training set, learning the NN and simulating the other two decades over the same period. The obtained model efficiency confirms the capability of this approach to supplying a useful tool in the evaluation of rainfall-runoff transformations.