Artificial neural networks in hydrological watershed modeling: surface flow contribution from the ungaged parts of a catchment

Watershed modeling is often faced with the difficulty of determining the flow contribution from the ungaged sections of the catchment. Where the main concern is making accurate streamflow forecasts at specific watershed locations, it is cost-effective and efficient to implement a simple system theoretic model. In this paper Artificial Neural Networks (ANNs) are used as system theoretic models to model the ungaged flows. Using data from the Kafue River sub-catchment in Zambia and a simple reservoir routing model, an estimate of the flow contribution from the ungaged sections is derived. Inputs: rainfall, evaporation, and previous-time-step flow are fed to a series of Feedforward-Backpropagation ANNs with target-output the current derived flow. Selected best performing ANNs are compared with Autoregressive Moving Average models with exogenous inputs (ARMAX) and they give accurate and more robust forecasts over long term than the best performing ARMAXs thereby making ANNs a viable alternative in forecasting.