our work

Machine Learning For Groundwater Level Prediction

Groundwater in the Steenkoppies compartment is extensively used for agriculture practices that can potentially lead to groundwater storage depletion, which threatens groundwater sustainability. Groundwater models are needed to develop an understanding of the complex, interdependent relationships occurring in a groundwater system. Groundwater levels represent the response of an aquifer to changes in storage, recharge, discharge, and hydrological stresses. They are, therefore, useful to identify limits and unacceptable impacts on an aquifer to implement sustainable groundwater management decisions.

Conventionally, numerical techniques are used for groundwater modelling, and the use of machine learning techniques for groundwater modelling is relatively new in South Africa. Unlike numerical models, Artificial Neural Networks (ANNs) do not need extensive characterisation of physical properties and aquifer conditions. ANNs are data-driven and learn the behaviour of the aquifer system from measured values. The power of ANNs lies in their ability to model systems that are poorly defined or where observations may be difficult to achieve, and one does not recognise or fully understand all the existing relationships in the groundwater system.

Kirsty Gibson made use of the Neural Network Autoregression (NNAR), a sub-type of ANN in her Masters dissertation. The basic concept to model a time series using ANNs is to predict the target variable, y, assuming it has a relationship with input variables, x. Kirsty predicted y, the groundwater levels, using the x variables rainfall, spring discharge from the Maloney’s Eye spring, air temperature, and groundwater usage in the Steenkoppies Aquifer (the mutual information index was used to quantify how informative each of these variables was in predicting groundwater levels in the Steenkoppies Aquifer).

The results showed that the NNAR could be used to make groundwater level predictions with reasonable accuracy at a monthly temporal resolution for up to 30 months in the future for 18 boreholes across the Steenkoppies Aquifer (see graphs). The model could further make groundwater level predictions for scenarios of change such as dryer or wetter periods. Overall, the NNAR performed well in predicting groundwater levels in the Steenkoppies Aquifer. The transferability of the NNAR to model groundwater levels in different aquifer systems or groundwater levels at different temporal resolutions should be further explored to assess the robustness of the NNAR to predict groundwater levels. The applicability of Machine learning in the water sector is a fascinating and new topic in South Africa that Kirsty hopes to explore in more depth in the future.

Graph of groundwater level predictions
Graph of groundwater level predictions
Two examples of the NNAR ability to predict groundwater levels from two boreholes in the Steenkoppies Aquifer. Each graph represents a comparison between the actual groundwater levels (grey line) and groundwater levels predicted made by the NNAR (blue line).

8 Beach Road

muizenberg, Cape town