HESS Hydrology and Earth System Sciences HESS Hydrol. Earth Syst. Sci. 1607-7938 Copernicus Publications Göttingen, Germany 10.5194/hess-9-111-2005 Constraints of artificial neural networks for rainfall-runoff modelling: trade-offs in hydrological state representation and model evaluation de Vos N. J. 1 Rientjes T. H. M. 1 2 Water Resources Section, Delft University of Technology, Delft, The Netherlands Department of Water Resources, International Institute for Geo-Information Science and Earth Observation (ITC), Enschede, The Netherlands 05 07 2005 9 1/2 111 126 Copyright: © 2005 N. J. de Vos 2005 This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 2.5 Generic License. To view a copy of this licence, visit https://creativecommons.org/licenses/by-nc-sa/2.5/ This article is available from https://hess.copernicus.org/articles/9/111/2005/hess-9-111-2005.html The full text article is available as a PDF file from https://hess.copernicus.org/articles/9/111/2005/hess-9-111-2005.pdf

The application of Artificial Neural Networks (ANNs) in rainfall-runoff modelling needs to be researched more extensively in order to appreciate and fulfil the potential of this modelling approach. This paper reports on the application of multi-layer feedforward ANNs for rainfall-runoff modelling of the Geer catchment (Belgium) using both daily and hourly data. The daily forecast results indicate that ANNs can be considered good alternatives for traditional rainfall-runoff modelling approaches, but the simulations based on hourly data reveal timing errors as a result of a dominating autoregressive component. This component is introduced in model simulations by using previously observed runoff values as ANN model input, which is a popular method for indirectly representing the hydrological state of a catchment. Two possible solutions to this problem of lagged predictions are presented. Firstly, several alternatives for representation of the hydrological state are tested as ANN inputs: moving averages over time of observed discharges and rainfall, and the output of the simple GR4J model component for soil moisture. A combination of these hydrological state representers produces good results in terms of timing, but the overall goodness of fit is not as good as the simulations with previous runoff data. Secondly, the possibility of using multiple measures of model performance during ANN training is mentioned.