Forecasting Groundwater Levels: A Comparison Between Support Vector Regression and Numerical Model

This study investigates groundwater levels (GWLs) in the alluvial aquifer of the Sava River valley, located in the north-western part of Croatia. It provides the first quantitative assessment of groundwater levels using machine learning in this part of Europe. Groundwater levels from 1998 to 2017 were predicted using support vector regression (SVR). The input variables were initially monthly data on two basic elements that influence groundwater dynamics (precipitation and the Sava River levels). Later, GWLs from the previous month (GWL-1) were added as an additional predictor. Results demonstrated that the SVR model effectively predicts groundwater levels. Introducing GWL-1 reduced RMSE and MAE values by more than 47% and 46%, respectively, while increasing the R2 value by over 36%. The improvement was more pronounced farther from the Sava River, since GWLs near the river are more directly influenced by river stage fluctuations, diminishing the impact of GWL-1. Compared to the existing regional numerical model (NM), the SVR model outperformed the NM with improvements of approximately 12% to 76% across performance indicators. Our findings suggest that the SVR model provides a reliable method for predicting groundwater levels at specific observation wells, making it a valuable tool for applications such as forecasting groundwater availability for farmers during dry periods and flood risk assessment during periods of heavy rainfall.