Assessing the Robustness of Long Short-Term Memory Networks for Streamflow Simulation under Climate Change

The quantification of present and future water resources at the regional scale is critical for designing climate change (CC) adaptation strategies and assessing flood/drought risks. Process-oriented Hydrological Models have traditionally been used for hydrological projections under climate change. Recurrent neural networks of the Long Short-Term Memory (LSTM) type have recently emerged as a powerful substitute, showing strong performance in temporal induction (TI) task [1], (when historical data of predicted basins are used for model training). 10-year daily streamflow predictions over French catchments demonstrated excellent accuracy [2], highlighting the potential of LSTM for predictions over future periods. However, their ability to operate under non-stationary CC conditions remains largely unexplored. 

The Explore 2 project [3] provides a multi-scenario and multi-model ensemble of future streamflow projections over France (3 greenhouse gas emission scenario - 17 climate projections - 9 hydrological models). The selection of hydrological models was based on the state of the art in 2021, just before the emergence of LSTM in hydrological applications. Deep learning based models were therefore not included in the ensemble. 

In this study, we harness the Explore2’s spatially consistent datasets to evaluate LSTM performance against the process-oriented hydrological models. A subset of catchments with low anthropogenic influence and sufficiently long streamflow records is selected. These catchments are used for model training. Model evaluation is then performed under the RCP 8.5 emission scenario until 2100. A focus on streamflow elasticity curves [4] is proposed to assess the hydrological response to a change in climatic forcing. Catchments are clustered using their hydrological signatures to assess the ability of models to preserve hydrological dynamics.

This work aims to clarify the role of LSTM in hydrological CC studies and opens new opportunities for reliable streamflow projections. Beyond this application, predictions outside the country could be investigated to extend the projections outside the national boundaries. 

[1] Kratzert, F.; Klotz, D.; Brenner, C.; Schulz, K.; Herrnegger, M. Rainfall–Runoff Modelling Using Long Short-Term Memory (LSTM) Networks. Hydrology and Earth System Sciences 2018, 22 (11), 6005–6022. https://doi.org/10.5194/hess-22-6005-2018. 

[2] Assessing Temporal and Spatial Generalization of Lstms for Streamflow Modeling in French Watersheds with and Without European Training Data. https://doi.org/10.2139/ssrn.5286855. 

[3] Sauquet, E. et al. A Large Transient Multi-Scenario Multi-Model Ensemble of Future Streamflow and Groundwater Projections in France. EGUsphere 2025, 1–41. https://doi.org/10.5194/egusphere-2025-1788. 

[4] Anderson, Bailey J., Manuela I. Brunner, Louise J. Slater, et Simon J. Dadson. « Elasticity Curves Describe Streamflow Sensitivity to Precipitation across the Entire Flow Distribution ». Hydrology and Earth System Sciences 28, n^o 7 (2024): 1567‑83. https://doi.org/10.5194/hess-28-1567-2024.