Never Train a Deep Learning Model on a Single Well? Revisiting Training Strategies for Groundwater Level Prediction

Abstract. Deep learning (DL) models are increasingly used for hydrological forecasting, with a growing shift from site-specific to globally trained architectures. This study tests whether the widely held assumption that global models consistently outperform local ones also applies to groundwater systems, which differ substantially from surface water due to slow response dynamics, data scarcity, and strong site heterogeneity. Using a benchmark dataset of nearly 3000 monitoring wells across Germany, we systematically compare global Long Short-Term Memory (LSTM) models with locally trained single-well models in terms of overall performance, training data characteristics, prediction of extremes, and spatial generalization. For groundwater level prediction, we find that global models provide no systematic accuracy advantage over local models. Local models more often capture site-specific behavior, while global models yield more robust but less specialized predictions across diverse wells. Performance gains arise primarily from dynamically coherent training data, whereas random data reduction has little effect, indicating that similarity matters more than quantity in this setting. Both model types struggle with extreme groundwater conditions, and global models generalize reliably only to wells with comparable dynamics. These findings qualify the assumption of global model superiority and highlight the need to align modeling strategies with groundwater-specific constraints and application goals.