Application of Artificial Neural Networks to Forecast River Discharge Rates: A Case Study of the Grootfontein Aquifer

Abstract

In this paper, we attempt to provide a data driven approach for modelling discharge rates in a groundwater-fed spring emanating from the Grootfontein dolomite aquifer, within the North West province. While application of neural networks for discharge prediction in general has been investigated by many authors, in this paper, we employ several neural network architectures using different feature sets to select the most prominent feature set and neural network model in the prediction of discharge rates from the Grootfontein aquifer. Three different neural networks were tested including Long-short Term Memory, Gated Recurrent Unit and Feed Forward Neural network. The results show that by including groundwater levels, all networks significantly improved in their performance as compared to without, recording lowest MAE score of 0.046 (m3/s) and 0.219 (m3/s) respectively. The results from this paper infer that recurrent neural networks are promising tools for predicting discharge rates in the study area. Furthermore, the paper also infers that groundwater levels can be successfully applied to predict the discharge rates of groundwater-fed springs, generate discharge data or patch missing discharge data in the compartment as long as we have key groundwater levels and a weather station nearby. Due to the lack of spring discharge monitoring in many areas of South Africa, this kind of methodology could become a useful tool in water resource management and also supporting water agencies to face drought issues specifically in karst regions.