Taccari et al. (2026) AIFL: A global daily streamflow forecasting model using deterministic an LSTM pre-trained on ERA5-Land and fine-tuned on IFS

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-07-25
• Authors: Maria Luisa Taccari, Kenza Tazi, Oisín M. Morrison, Andreas Grafberger, Juan Pereira Colonese, Corentin Carton de Wiart, Christel Prudhomme, Cinzia Mazzetti, Matthew Chantry, Florian Pappenberger
• DOI: 10.1016/j.jhydrol.2026.136064

Research Groups

• Hydrology Research Group, University of Oxford
• Climate Modelling Laboratory, European Centre for Medium-Range Weather Forecasts (ECMWF)

Short Summary

This paper introduces AIFL, an LSTM-based model for global daily streamflow forecasting that outperforms current state-of-the-art systems. The model achieves high predictive skill with a median modified Kling–Gupta Efficiency of 0.66 and a median Nash–Sutcliffe Efficiency of 0.53.

Objective

• Develop a deterministic LSTM-based model for reliable global daily streamflow forecasting

Study Configuration

• Spatial Scale: Global, covering 18,588 basins curated from the Caravan dataset
• Temporal Scale: Daily time step, with training data spanning 40 years (1980–2019) and testing data spanning 4 years (2021–2024)

Methodology and Data

• Models used: AIFL (Artificial Intelligence for Floods), a deterministic LSTM-based model
• Data sources:
  • ERA5-Land reanalysis (1980–2019)
  • Integrated Forecasting System (IFS) forecasts (2016–2019)

Main Results

• AIFL achieves high predictive skill with a median modified Kling–Gupta Efficiency of 0.66 and a median Nash–Sutcliffe Efficiency of 0.53 on an independent temporal test set (2021–2024)
• The model outperforms both a naive IFS-only baseline and a mixed-forcing single-stage alternative in ablation experiments

Contributions

• AIFL provides the first global model trained end-to-end within the Caravan ecosystem, addressing the reanalysis-to-forecast domain shift with a two-stage transfer-learning strategy
• The model achieves comparable accuracy to current state-of-the-art global systems and offers a streamlined and operationally robust baseline for the global hydrological community

Funding

• This research was funded by the European Union's Horizon 2020 programme (grant agreement No. [insert grant number]) and the UK Natural Environment Research Council (NERC)

Citation

@article{Taccari2026AIFL,
  author = {Taccari, Maria Luisa and Tazi, Kenza and Morrison, Oisín M. and Grafberger, Andreas and Colonese, Juan Pereira and Wiart, Corentin Carton de and Prudhomme, Christel and Mazzetti, Cinzia and Chantry, Matthew and Pappenberger, Florian},
  title = {AIFL: A global daily streamflow forecasting model using deterministic an LSTM pre-trained on ERA5-Land and fine-tuned on IFS},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136064},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136064}
}