Building regional LSTM models to predict river and groundwater temperature in Germany

River and groundwater temperatures (RT and GWT) play a critical role in aquatic ecosystem functioning, drinking-water resources, and geothermal potential. Their dynamics reflect delayed responses to atmospheric temperature as well as interactions between surface water and groundwater, while for GWT, geological and geomorphological controls may be particularly important. Under ongoing climate change, understanding the dominant controls on RT and GWT at regional scales is essential for anticipating and mitigating the impacts of warming water resources.

In this study, we develop two deep learning models to predict and analyse daily RT and GWT across Germany. For river temperature, we use low-frequency (biweekly to monthly) observations from the QUADICA v2 dataset (Ebeling et al., 2024) at more than 300 locations, combined with hydroclimatic variables from CAMELS-DE and catchment descriptors as predictors. For groundwater temperature, we use daily sensor measurements from the federal state of Thuringia. After data quality control, 77 monitoring wells are retained, and model inputs include groundwater levels, atmospheric temperature, precipitation, and static well characteristics such as depth and surface elevation. We employ an LSTM architecture to account for delayed responses to atmospheric forcing, which are known to be characteristic of water temperature dynamics. All variables are transformed using a Box-Cox transformation to approximate normal distributions. Model hyperparameters are tuned using a train-validation-test split (60%, 15%, and 25%, respectively) by minimizing the root mean squared error on the validation set, and overfitting is mitigated through early stopping.

Preliminary results show that the LSTM model for river temperature achieves a median Kling–Gupta efficiency (KGE) of 0.88 for unseen periods, indicating a high predictive skill. In contrast, the groundwater temperature model yields a median KGE below zero, highlighting substantially higher complexity of the system response compared to river temperature. This reduced performance is likely attributable to a combination of data limitations and missing site-specific controls, including river proximity, land use, human abstractions, and recharge processes. These findings highlight the need for larger groundwater datasets and richer explanatory variables to better understand and predict regional-scale GWT variability.