Wu et al. (2026) Predicting streamflow in heavily regulated plain river networks using physics-guided spatiotemporal deep learning models: A case study of the Four Lakes Basin

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-09-09
• Authors: Yufeng Wu, Zhihong Liu, Xianqiang Tang, Rui Li, Shouliang Huo
• DOI: 10.1016/j.ejrh.2026.103959

Research Groups

• Faculty of Infrastructure Engineering, Dalian University of Technology
• Science and Technology Promotion Centre, Ministry of Water Resources P.R.C.
• Basin Water Environmental Department, Yangtze River Scientific Research Institute
• Hubei Provincial Field Scientific Observation and Research Station for Eco-environmental Effects of the Danjiangkou Water Diversion Project
• State Key Laboratory of Environmental Criteria and Risk Assessment, Chinese Research Academy of Environmental Sciences

Short Summary

This study proposes a hybrid streamflow prediction model that integrates physical processes with data-driven methods to predict streamflow in heavily regulated plain river networks. The model effectively mitigates peak-flow misalignment and extreme errors caused by intensive cross-boundary pumping and internal regulation.

Objective

• Develop a physics-guided streamflow prediction model that integrates spatial topology extraction with temporal memory modules.
• Evaluate the ability of the hybrid model to capture complex artificial regulation events.
• Quantify the independent contributions of the core deep learning modules in reconstructing regulated streamflow evolution through ablation experiments.
• Reveal the unnatural driving mechanisms and spatial topological dependencies of regulated streamflow, and quantify prediction uncertainty and operational margins objectively.

Study Configuration

• Spatial Scale: The study area is a heavily regulated plain river network with a total drainage area of approximately 11,547.5 km² in China.
• Temporal Scale: The study period spans from 1 June 2022 to 31 December 2024.

Methodology and Data

• Models used:
  • SWAT (Soil and Water Assessment Tool)
  • Graph neural networks (GNNs) with graph attention network (GAT) as the core algorithm for spatial feature extraction.
  • Bidirectional long short-term memory (BiLSTM) networks for temporal sequence learning.
• Data sources:
  • DEM (Digital Elevation Model)
  • Land use maps
  • Soil distribution maps
  • Daily meteorological observations, including precipitation, relative humidity, solar radiation, temperature, and wind speed
  • Observed sluice-gate streamflow records

Main Results

• The hybrid model substantially improves both dynamic fitting accuracy and absolute error control across the regulated control nodes.
• The hybrid model captures sudden regulation events accurately at local time scales, including dry-season water supplementation and intensive flood-season drainage.

Contributions

• This study provides a reliable approach for streamflow simulation in complex regulated basins that lack fine-resolution operational data.
• The proposed hybrid model integrates physical processes with data-driven methods to predict streamflow in heavily regulated plain river networks.

Funding

• This research was funded by the Science and Technology Promotion Centre, Ministry of Water Resources P.R.C. (project code: [insert project code])

Citation

@article{Wu2026Predicting,
  author = {Wu, Yufeng and Chang, Qingrui and Liu, Zhihong and Tang, Xianqiang and Li, Rui and Huo, Shouliang},
  title = {Predicting streamflow in heavily regulated plain river networks using physics-guided spatiotemporal deep learning models: A case study of the Four Lakes Basin},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103959},
  url = {https://doi.org/10.1016/j.ejrh.2026.103959}
}