Combining LSTMs with a Single-Model Large Ensemble for Runoff and Water Temperature Projections in Bavaria

Both river runoff and river water temperature are experiencing highly dynamic alterations, posing serious threat to aquatic ecosystems and water resources management under climate change. Data-driven models such as Long Short-Term Memory (LSTM) networks have demonstrated remarkable skill in hydrological prediction, yet their application under non-stationary climate conditions remains challenging due to limited generalization to unseen catchments and conditions beyond the training distribution. We address these challenges by combining LSTM architectures with single-model initial condition large ensemble (SMILE) climate projections to assess non-stationary, non-linear hydrological responses considering the full range of internal climate variability and climate change, enabling robust assessment of rare and extreme events in Bavaria, Germany.

Our study builds on the ClimEx project, which provides a 50-member ensemble of climate simulations (1950–2099, RCP8.5 emission scenario) at 12 km resolution over Europe using the Canadian Regional Climate Model CRCM5.

We present two complementary application cases operating daily and at 3-hourly temporal resolution: i) For discharge prediction, we train an LSTM on observed runoff across 98 Bavarian catchments, validated against simulations from the process-based Water balance Simulation Model (WaSiM). The architecture processes dynamic meteorological forcings through stacked LSTM layers while incorporating static catchment attributes, using a composite loss function that balances performance across high and low flows. The trained model is then driven by the ClimEx ensemble to generate probabilistic discharge projections for future climate. ii) For water temperature (Tw) prediction, we developed an Entity-Aware LSTM (EA-LSTM) framework trained on observations from 44 Bavarian gauging stations, a subset of the 98 catchments constrained by Tw data availability, extended with nine French river basins to broaden the climatic gradient encountered during training. The EA-LSTM architecture explicitly separates static catchment attributes (elevation, slope, upstream river length) from dynamic meteorological forcings, using static features to parameterize the input gate rather than concatenating them at every timestep. This allows the network to learn site-specific temporal dynamics without overfitting individual locations.

To enhance model interpretability, we apply explainable AI (XAI) techniques including permutation-based feature importance analysis. Results reveal that air temperature and radiation dominate Tw predictions overall, while topographic attributes gain importance under thermal extremes, indicating the model captures physically meaningful process controls. Additionally, robustness tests with perturbed static inputs confirm smooth performance degradation rather than abrupt collapse, suggesting the EA-LSTM learns generalizable attribute-response relationships rather than memorizing site identities.

Both cases demonstrate how combining diverse training data with ensemble-based climate projections enables more robust predictions of hydrological extremes under climate change, while XAI methods provide transparency into learned representations.