Machine learning approaches are increasingly used to predict groundwater levels, with local models for single monitoring wells currently being state of the art. Global models enable training and forecasting at multiple monitoring wells simultaneously, incorporating dynamic (e.g., meteorological) and static (e.g., hydro(geo)logical) features. These models can generalize predictions to wells with similar site characteristics and offer computational scaling benefits by requiring only one model for a larger area.
This study presents two global machine-learning models for short-term groundwater level prediction (up to 12 weeks): the Temporal Fusion Transformer (TFT) and Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS). The TFT combines recurrent neural networks with an attention mechanism and can determine the significance of individual input variables (feature importance). The N-HiTS model uses a fork architecture with multiple stacks to model different data frequencies, enhancing prediction accuracy.
We used a dataset of approximately 5300 monitoring wells across Germany, with groundwater levels from 1990 to 2016 (around 4.5 million values). Input features included groundwater levels, meteorological parameters, and site-specific environmental features such as hydro(geo)logical, soil, and spatial characteristics.
The TFT model showed a median NSE of 0.34, while the N-HiTS model performed better with a median NSE of 0.5 for the 12-week forecast. Around 25% of the test sites achieved an NSE over 0.68. Key features for forecast quality included historical groundwater levels, precipitation, the standard deviation of groundwater levels, and major hydrogeological districts. The topographical wetness index was the most important static feature, though its impact on model performance was minimal.