Advancing water resource management: Insights and implications from global machine learning models in groundwater prediction

Abstract

Accurate and reliable predictions of groundwater levels are essential for sustainable water resource management. Faced with the impacts of climate change and the increasing stress on groundwater resources, there is a growing necessity to balance domestic, agricultural, and industrial utilization. To address these challenges, innovative and progressive approaches in predicting groundwater levels are necessary, such as the application of machine learning (ML) methods. The advancement of these ML-based prediction models is a crucial component of the BMBF project KIMoDIs. Within this research initiative, an AI-based monitoring, data management, and information system for coupled prediction and early warning of low groundwater levels and groundwater salinisation is being developed. Currently, the state-of-the-art in hydrogeology involves individual models for each groundwater monitoring well (local models). While local models can achieve high predictive accuracy, their application to a multitude of measurement points is impractical. Conversely, global models allow training and prediction for multiple measurement wells simultaneously. This model class has the potential to learn and capture dynamics beyond a single well and their dependency on dynamic input variables (e.g., meteorological parameters) as well as static variables (e.g., specific hydro(geo)logical or morphometric site properties). Particularly with extensive training datasets, global model approaches can provide predictions at measurement points sharing similar site properties to those used in training (generalization). Additionally, they offer advantages in terms of computational requirements as well as model management, as only one model needs to be trained and applied over a large area. The objective of this study is to demonstrate the predictive capabilities of modern ML methods in the context of groundwater level prediction. Further, it provides insights and recommendations regarding the extent to which global models, with a wealth of spatial and temporal information, can contribute to improve prediction accuracy. Global ML models are used for short-term prediction of groundwater levels on a regional scale: Two model architectures (Temporal Fusion Transformer and Neural Hierarchical Interpolation for Time Series Forecasting) are applied to over 5000 groundwater monitoring points in Germany in order to predict groundwater levels for up to 12 weeks. Meteorological data and historical groundwater level data dating back to 1990 (dynamic features) as well as hydrogeological, soil and morphometric properties (static features) are used as input data. Additionally, feature importance is assessed, and eliminating various inputs enabled to identify suitable features for groundwater level prediction.