Application of machine learning techniques in predicting groundwater levels and discharge rates in the North West aquifers
No Thumbnail Available
Date
2021
Authors
Kanyama, Yolanda
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The dolomite aquifers in the North West province have suffered from severe groundwater over abstraction causing groundwater levels to decline. This has threatened water supply security
and caused municipal and irrigation boreholes to dry up in the province. The continuous
decline in groundwater levels in the region will inevitably lead to increased costs and risks to
the sustainable development of groundwater resources.
The key principles underpinning water resource management in South Africa water law are
sustainability of groundwater resources. Accurate prediction of groundwater levels and discharge
rates is essential to the sustainable utilisation and management of groundwater resources in the
Grootfontein Aquifer. In South Africa, there is a limited body of literature in the forecasting
of both groundwater levels and discharge rates using machine learning models.
The objectives of this study were therefore (1) to present ranking by entropy determining
the primary factors governing spring discharge patterns and groundwater levels using mutual
information (2) to develop a robust model that can be used to forecast future discharge rates
and groundwater levels in the Grootfontein aquifer, (3) to predict groundwater levels under
adjusted inputs (rainfall and abstraction amounts) and (4) to test model generazability by
using the selected optimal model to predict groundwater levels in a different aquifer setting.
Results through mutual information showed that regional abstraction, discharge and precipi tation were the highest contributing factors to predict both groundwater levels and discharge
rates in the karst aquifer. An interesting observation came to light as the study highlighted
decomposed signals of rainfall produced higher entropy scores (+0.30) and better prediction for
the machine learning models as compared to actual measured rainfall. Temperature proved to
have little to no influence in the prediction of both groundwater levels and discharge rates.
The research results obtained for the predictive models showed that the long-short term memory
model was better suited to predict both groundwater levels and discharge rates in the aquifer.
Scenario testing results of groundwater levels showed a decline in groundwater levels when
abstraction amounts were doubled and a more subtle decline when rainfall peaks amounts were
reduced. Moreover, during what can be conspired as drought years (worst case scenario) the
groundwater levels in the compartment decreased by almost 2m. These results indicated the
importance the chosen variables had on groundwater levels in the region further strengthening
the entropy results. The results would also be useful to both hydrologists and water managers
in aquifer management.
For discharge prediction, two case scenarios were tested. In both scenarios tested the gated
recurrent unit and long-short term memory models were at par with one another. The results
showed that including groundwater levels from surrounding boreholes drastically improved the
prediction of discharge. In literature, the focus has been on using weather parameters and past
discharge as model inputs, however this study showed that groundwater levels inclusion had a
much better effect on the modelled results. Hence, we propose the use of groundwater levels,
precipitation and abstraction to not only infill incomplete discharge records but to generate
discharge data where there are no records at all.
Lastly, to test model generalization, the Steenkoppies aquifer was used as a case study to predict
groundwater levels in the aquifer. The best performing model for groundwater level prediction
was used with the same model pipeline as that used for the Grootfontein. The results from
this experiment proved the capabilities of the long-short term memory model in predicting
groundwater levels in a different setting. The model was able to capture the trends of the
I
observed dataset better than the other models tested. The results of the model simulations in
this experiment highlighted the ability of the long-short term memory model to capture the
broad patterns of the Steenkoppies data so we can assume, through this experiment that, the
long-short term memory model is likely to be able to predict groundwater levels from other
regions.
The results presented in this research study present a useful approach for discharge and ground water prediction. The frameworks used could be used by water managers and municipalities
to manage water resources more effectively during drier years while ensuring that groundwater
levels do not deplete. Drought warnings and water restrictions can also be issued promptly
Description
A dissertation submitted in fulfilment of the requirements for the degree Master of Science in Computer Science to the Faculty of Science, School of Computer Science and Applied Mathematics, University of the Witwatersrand, Johannesburg, 2021