Résumé
This paper investigates the way to adapt neural network design to non-Gaussian environmental data. The process of model selection is specifically investigated in order to make the model more robust. It appears that the standard method of cross-validation gains to be applied on an ensemble model, rather than a unique model, in order to apply additional regularization. Specific architectures of neural networks based on multilayer perceptron were used simultaneously with various methods of complexity selection. Prediction results on floods and droughts show that the sensitivity to the initial value of parameters could be greatly reduced. A case-study is chosen on a exceptionally complex hydrosystem, subjected to critical water resource conflicts.