A Novel Machine Learning-based Method for Groundwater Modelling involving Aquifer Rainfall Time Response Analysis and Clustering of Groundwater Wells

Clustering of groundwater level data is crucial for water resource management, as it increases the efficiency of models in distinctly predicting specific hydrogeological patterns in aquifer systems. Traditional methods mostly rely on spatial or time series distance metrics, neglecting the impact of external inputs (rainfall, evapotranspiration, etc.) on aquifer systems. This study introduces an innovative machine learning-based approach to model aquifer systems at the piezometer level. While our flexible methodology accommodates any model and input, we selected a random forest model for its lightweight nature and interpretability. This model-based technique enables the clustering of similar aquifers based on model parameters. By leveraging the decision trees feature importances, we derive the rainfall response time distribution of the aquifer at the piezometer level, facilitating a quantitative analysis of the local aquifer dynamics. Additionally, we demonstrate that, by selecting analogous distributions using a simple similarity measure, the predictive performance of groundwater level global forecasting models is significantly enhanced.