Daily rainfall-runoff modelling by neural networks in semi-arid zone: case of Wadi Ouahrane’s basin

Abstract

This research work will allow checking efficiency of formal neural networks for flows’ modelling of wadi Ouahrane’s basin from rainfall-runoff relation which is non-linear. Two models of neural networks were optimized through supervised learning and compared in order to achieve this goal, the first model with input rain, and the second one with rain and input ETP. These neuronal models were compared with another overall model, the GR4j model. Then, it has been optimized and compared with the three first models, a third model of neural network with rain, ETP and soil moisture (calculated by the model GR4j) input.

The neuronal models were optimized with algorithm of Levenberg Marquarld (LM), while the GR4j model was optimized with SCE-UA method. The Nash criterion (%) and the correlation coefficient of Pearson (R) allowed appreciating performances of these models.

Keywords: modeling; neural network; supervised learning; algorithm of Levenberg Marquarld; GR4J