Constraints of artificial neural networks for rainfall-runoff modelling: trade-offs in hydrological state representation and model evaluation

摘要

The application of Artificial Neural Networks (ANNs) in rainfall-runoffmodelling needs to be researched more extensively in order to appreciate andfulfil the potential of this modelling approach. This paper reports on theapplication of multi-layer feedforward ANNs for rainfall-runoff modellingof the Geer catchment (Belgium) using both daily and hourly data. The dailyforecast results indicate that ANNs can be considered good alternatives fortraditional rainfall-runoff modelling approaches, but the simulations basedon hourly data reveal timing errors as a result of a dominatingautoregressive component. This component is introduced in model simulationsby using previously observed runoff values as ANN model input, which is apopular method for indirectly representing the hydrological state of acatchment. Two possible solutions to this problem of lagged predictions arepresented. Firstly, several alternatives for representation of thehydrological state are tested as ANN inputs: moving averages over time ofobserved discharges and rainfall, and the output of the simple GR4J modelcomponent for soil moisture. A combination of these hydrological staterepresenters produces good results in terms of timing, but the overallgoodness of fit is not as good as the simulations with previous runoff data.Secondly, the possibility of using multiple measures of model performanceduring ANN training is mentioned.