Machine Learning methods for providing groundwater management insights in semi-arid alluvial aquifers

Accurate regional groundwater level (GWL) forecasting is critical for evaluating the impacts of groundwater withdrawals on water resource sustainability, especially during droughts. A lack of hydrogeological data often makes it challenging to develop water balance models that forecast the impacts of pumping groundwater throughout a catchment. Machine Learning (ML) models have considerable potential for predicting GWLs without the need for characterisation of the hydrogeology, provided there are good time series records for, among others, rainfall, streamflow, GWL, and groundwater withdrawal. In this research, ML models are developed to assess the impact of the existing water sharing allocations in the Lower Murrumbidgee Catchment (LMC) area, New South Wales, Australia. The streamflow in this catchment is controlled by water released from upstream reservoirs, which store rainfall and snow melt from the Snowy Mountains. Groundwater recharge inputs include floodwaters, river leakage, diffuse rainfall, and irrigation deep drainage. Substantial aquifer recharge only occurs after flooding. ML models are developed using Linear Regression (LR), Random Forest (RF), Gradient Boosting (GB), and Artificial Neural Network (ANN) to predict the groundwater dynamics under a range of climatic conditions, with an emphasis on understanding the impacts of droughts. All four ML methods can reproduce the GWL variations, with RF being the most robust method. A variable sensitivity analysis shows that the GWL correlates strongly with the time series associated with rainfall in upstream catchments and the annual volume of groundwater extracted. The modelling shows that for median climatic and groundwater usage conditions the GWL throughout most of the research area will gradually fall between 0.31 m/yr and 1.8 m/yr. If the region experiences a drought like the one recorded between 1939 and 1949, and groundwater is extracted at current allocations, regional groundwater levels could fall between 7.83 m and 29.94 m. This GWL drawdown prediction would trigger groundwater access restrictions under the current water sharing plan. The results of this study demonstrate that the adoption of ML methods can guide the better management of aquifers worldwide.

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2024

Author dc:creator

• Xiao, Shuang

• Groundwater management
• Groundwater modelling
• Semi-arid alluvial aquifers
• Drought
• Machine learning
• anzsrc-for: 370703 Groundwater hydrology
• anzsrc-for: 410503 Groundwater quality processes and contaminated land assessment

Statement dc:rights

• open access
• CC BY 4.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

en

Identifier

https://doi.org/10.26190/unsworks/25522

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/101818