Application of artificial neural networks for river flow simulation in three French catchments

For more than a decade, Artificial Neural Networks (ANNs) have been increasingly used in hydrology as flexible black-box models of non-linear type. Within this category of models, the "multi-layer feed-forward network" used in this study consists of an input layer, an output layer, and one "hidden" layer in between. The model is applied to daily data of three catchments, all located in northwest France, for river flow simulation and forecasting and its performance is compared with those of five systemtheoretic models and one conceptual model. The ANN is observed to be the best performing individual model for the catchments tested. In the subsequent application of the Neural Network Method (NNM) for combining the outputs of the individual models, in different combinations, i.e. in a "multi-model approach" for deriving consensus forecasts, the NNM (as one of three Model Output Combination Techniques (MOCTs) considered) is found to be the best performing MOCT and also better than the best individual model. The Galway Flow Modelling and Forecasting System (GFMFS), a software package developed by the present authors, is used in the study.