Sun et al. (2026) Streamflow prediction in the Danube River Basin using a multi-source graph-integrated GCN-LSTM model

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-02-24
• Authors: Mingze Sun, Yifan Sun, Xin Yu, Yuyue Ye
• DOI: 10.1016/j.ejrh.2026.103275

Research Groups

• Key Lab of Spatial Data Mining and Information Sharing of Ministry of Education, Fuzhou University, Fujian, China
• National Engineering Research Center of Geospatial Information Technology, Fuzhou University, Fujian, China

Short Summary

This study developed a multi-source graph-integrated GCN-LSTM model for regional streamflow forecasting in the Danube River Basin, demonstrating its superior performance over conventional baselines and highlighting the critical role of daily temporal resolution for capturing extreme events, while revealing limitations in spatial generalization to ungauged sites.

Objective

• How does temporal resolution affect streamflow prediction accuracy?
• To what extent does station density influence model stability and robustness?
• Can the GCN-LSTM model maintain predictive performance and generalization when applied to ungauged locations?

Study Configuration

• Spatial Scale: Danube River Basin, utilizing data from 97 gauging stations.
• Temporal Scale: Monthly (January 1960 to December 2020) and daily (2000–2014) resolutions for input data and predictions. Prediction horizon is one-step-ahead.

Methodology and Data

• Models used: Graph Convolutional Network–Long Short-Term Memory (GCN–LSTM) framework. Benchmarked against Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Bidirectional LSTM (biLSTM).
• Data sources:
  • Streamflow observations: Global Runoff Data Centre (GRDC).
  • Meteorological and land-surface variables (static and dynamic predictors): ERA5-Land reanalysis dataset (European Centre for Medium-Range Weather Forecasts - ECMWF), provided at approximately 9 km resolution and regridded to 0.1° × 0.1° latitude–longitude grid.

Main Results

• The GCN-LSTM model consistently outperformed all baseline models, achieving the highest R² (0.9525) and lowest Mean Absolute Error (MAE) (114.4235 m³/s) and Root Mean Square Error (RMSE) (211.3520 m³/s) in streamflow prediction.
• Daily input data significantly improved the GCN-LSTM model's ability to reproduce streamflow dynamics, particularly in capturing the timing and magnitude of flood peaks, with a mean Kling–Gupta Efficiency (KGE) of 0.9649 for daily inputs compared to 0.4426 for monthly inputs.
• The model demonstrated high robustness to station density, maintaining stable performance (R² consistently above 0.93) even when the number of gauging stations was reduced from 97 to 17, with only a modest degradation observed when fewer than approximately 30 stations were used.
• Cross-validation experiments showed a substantial degradation in performance at ungauged sites (KGE decreased from 0.96 at training stations to 0.24 at unseen stations), with accuracy loss increasing with distance from training stations, and a tendency to underestimate peak streamflow.

Contributions

• Developed and validated a novel multi-source graph-integrated GCN-LSTM framework that effectively leverages spatial dependencies among gauging stations and temporal dynamics for enhanced regional streamflow forecasting.
• Quantified the critical importance of high temporal resolution data (daily vs. monthly) for accurately capturing short-term streamflow fluctuations and extreme events like flood peaks in large river basins.
• Demonstrated the robustness of the GCN-LSTM model to varying station densities, suggesting its applicability in regions with sparse or heterogeneous gauging networks.
• Provided new insights into the challenges of spatial generalization in hydrological modeling, highlighting that predictive skill at ungauged locations is distance-dependent and limited by the representativeness of training conditions in hydrologically diverse regions.

Funding

• National Key Research and Development Program of China (2025YFE0102700)
• National Natural Science Foundation of China (grant 42171426, 42374041 and 42304099)

Citation

@article{Sun2026Streamflow,
  author = {Sun, Mingze and Sun, Yifan and Yu, Xin and Ye, Yuyue},
  title = {Streamflow prediction in the Danube River Basin using a multi-source graph-integrated GCN-LSTM model},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103275},
  url = {https://doi.org/10.1016/j.ejrh.2026.103275}
}