Monthly runoff forecasting by means of artificial neural networks (ANNs)

Document Type: Research Paper


1 Department of Range and Watershed Management, Faculty of Natural Resources, University of Guilan, Somehsara, Iran

2 Department of Water Resources Engineering, Lund University, Box 118, SE-22 100, Lund, Sweden


Over the last decade or so, artificial neural networks (ANNs) have become one of the most promising tools for
modelling hydrological processes such as rainfall runoff processes. However, the employment of a single model does
not seem to be an appropriate approach for modelling such a complex, nonlinear, and discontinuous process that
varies in space and time. For this reason, this study aims at decomposing the process into different clusters based on
self-organizing map (SOM) ANN approach, and thereafter modelling different clusters into outputs using separate
feed-forward multilayer perceptron (MLP) and supervised self-organizing map (SSOM) ANN models. Specifically,
three different SOM models have been employed in order to cluster the input patterns into two, three, and four
clusters respectively so that each cluster in each model corresponds to certain physics of the process under
investigation and thereafter modelling of the input patterns in each cluster into corresponding outputs using feedforward
MLP and SSOM ANN models. The employed models were developed on two different watersheds, Iranian
and Canadian. It was found that although the idea of decomposition based on SOM is highly persuasive, our results
indicate that there is a need for more principled procedure in order to decompose the process. Moreover, according to
the modelling results the SSOM can be considered as an alternative approach to the feed-forward MLP.