Li et al. (2025) Dynamic Graph Transformer with Spatio-Temporal Attention for Streamflow Forecasting

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

Identification

• Journal: Hydrology
• Year: 2025
• Date: 2025-12-08
• Authors: Bo Li, Qingping Li, Xinzhi Zhou, Mingjiang Deng, Hongbo Ling
• DOI: 10.3390/hydrology12120322

Research Groups

[Information not available in the provided text.]

Short Summary

This study introduces DynaSTG-Former, a novel deep learning architecture designed to enhance multi-step-ahead streamflow forecasting by adaptively integrating diverse spatio-temporal dependencies. The model demonstrated exceptional performance in the Delaware River Basin, significantly outperforming baseline models and providing a robust tool for water management.

Objective

• To develop a novel deep learning architecture, DynaSTG-Former, that effectively captures spatio-temporal evolution characteristics and dynamic interdependencies among streamflow gauges for accurate multi-step-ahead streamflow forecasting.

Study Configuration

• Spatial Scale: Basin-scale (Delaware River Basin)
• Temporal Scale: Multi-step-ahead forecasting (12 hours, 36 hours, and 72 hours)

Methodology and Data

• Models used: DynaSTG-Former (novel deep learning architecture), LSTM (baseline), GRU (baseline), Transformer (baseline)
• Data sources: Streamflow gauge data (implied, for the Delaware River Basin)

Main Results

• DynaSTG-Former achieved basin-scale Kling–Gupta Efficiency (KGE) values of 0.961 for 12-hour forecasts, 0.956 for 36-hour forecasts, and 0.855 for 72-hour forecasts.
• The model significantly outperformed baseline models including LSTM, GRU, and Transformer.
• Ablation studies confirmed that the dynamic graph module was a core contributor to performance, with the Pearson correlation graph playing a dominant role in error reduction.
• The model demonstrated enhanced accuracy, stability, and strong robustness for streamflow forecasts at the basin scale.

Contributions

• Introduction of DynaSTG-Former, a novel deep learning architecture that employs a multi-channel dynamic graph constructor to adaptively integrate physical topology, statistical correlation, and trend similarity for streamflow forecasting.
• Design of a dual-stream temporal predictor within DynaSTG-Former to collaboratively model long-range dependencies and local transient features.
• Empirical demonstration of superior performance and robustness of DynaSTG-Former in multi-step-ahead streamflow forecasting compared to conventional data-driven approaches.

Funding

[Information not available in the provided text.]

Citation

@article{Li2025Dynamic,
  author = {Li, Bo and Li, Qingping and Zhou, Xinzhi and Deng, Mingjiang and Ling, Hongbo},
  title = {Dynamic Graph Transformer with Spatio-Temporal Attention for Streamflow Forecasting},
  journal = {Hydrology},
  year = {2025},
  doi = {10.3390/hydrology12120322},
  url = {https://doi.org/10.3390/hydrology12120322}
}