Comparison of physical and data-driven models to forecast groundwater level changes with the inclusion of GRACE – A case study over the state of Victoria, Australia

Abstract

Accurate prediction of groundwater levels (GWLs) is of significant importance in the sustainable use and efficient management of groundwater resources. A comparative study is conducted to evaluate the performance of machine learning and physical models in forecasting groundwater dynamics, using the state of Victoria, Australia, as a case study. GWLs are predicted using two traditional machine learning models (Random Forest (RF) and Artificial Neural Network (ANN)) and a deep learning model (Long Short-Term Memory, LSTM). The impact of utilizing GRACE-derived terrestrial water storage (TWS) anomalies as input parameters in all three models is also evaluated. The GWL estimates are compared with in-situ GWL measurements from ground networks and groundwater storage simulated from two different land surface models, World-Wide Water (W3) and Catchment Land Surface model with GRACE data assimilation (CLSM-DA). The evaluation shows that the accuracy of LSTM is significantly higher than the machine-learning models by increasing Pearson coefficient (PR) values during the prediction period by 23.89% (compared to ANN) and 41.32% (compared to RF), respectively. The inclusion of GRACE data has a significant impact on the performance of predicted GWLs, by improving PR from 0.430 to 0.643 on average. The CLSM-DA GWLs products have higher accuracy than the W3 model, which underlines the benefit of including GRACE data in land surface models. The LSTM model achieves higher accuracy in plain areas, while the performance in ocean and mountain areas is poorer due to abrupt changes in climate variables. This study supplies a feasible and accurate approach for predicting GWLs dynamics and provides a reference for model selection.