Use of Neural Networks to Forecast Time Series: River Flow Modeling

This paper presents an alternative approach to time series forecasting, through use of artificial neural networks (ANNs), a relatively new concept in hydrological research. Box and Jenkins ARMAX (autoregressive moving average with exogenous inputs) models have been widely used in modeling various time series with satisfactory results. This study shows that ANNs can produce comparable, to ARMAX, and in some cases even, better forecasting results, especially for long-term prediction. By learning, through training, the underlying mapping of the time series, an ANN provides robust forecasting. The results obtained using real-life data from a catchment in Zambia suggest ANNs could be used as an efficient and effective models in forecasting hydrological variables such as river discharge, river stage, and runoff. Key-Words: Artificial Neural Networks, feedforward, ARMAX, alternative approach, training, mapping, hydrologic(al), forecasting.