Chen et al. (2026) Differentiable parameter learning of reservoir operation modules

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-01-05
• Authors: Zexin Chen, Tongtiegang Zhao
• DOI: 10.1016/j.jhydrol.2026.134915

Research Groups

Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), School of Civil Engineering, Sun Yat-Sen University, Guangzhou, China

Short Summary

This paper develops three differentiable reservoir operation modules (dROMs) for WBMplus, WaterGAP2, and CWatM by utilizing Long Short-Term Memory (LSTM) to simultaneously estimate static and dynamic parameters. A case study across 88 US reservoirs demonstrates that local dROMs significantly outperform empirical ROMs, highlighting the effectiveness of differentiable parameter learning in improving reservoir simulation.

Objective

• To develop and evaluate three differentiable reservoir operation modules (dROMs) for the Water Balance Model plus (WBMplus), Water-Global Assessment and Prognosis 2.0 (WaterGAP2), and Community Water Model (CWatM) using differentiable parameter learning (dPL) with Long Short-Term Memory (LSTM) to simultaneously estimate static and dynamic parameters.

Study Configuration

• Spatial Scale: 88 reservoirs in the United States.
• Temporal Scale: Not explicitly defined for the study period, but focuses on continuous reservoir operation simulation.

Methodology and Data

• Models used: Differentiable Reservoir Operation Modules (dROMs), Water Balance Model plus (WBMplus), Water-Global Assessment and Prognosis 2.0 (WaterGAP2), Community Water Model (CWatM), Long Short-Term Memory (LSTM) network, Differentiable Parameter Learning (dPL).
• Data sources: Reservoir operational records and reservoir data for 88 reservoirs in the United States.

Main Results

• For WBMplus, WaterGAP2, and CWatM, the mean Kling-Gupta Efficiency (KGE) values for empirical ROMs were 0.44, 0.60, and 0.62, respectively.
• The mean KGE values for local dROMs were 0.54, 0.75, and 0.71 for WBMplus, WaterGAP2, and CWatM, respectively, demonstrating a significant improvement over empirical ROMs.
• The mean KGE values for regional dROMs were 0.54, 0.74, and 0.69 for WBMplus, WaterGAP2, and CWatM, respectively, being slightly lower than local dROMs.
• Reservoir characteristics influence simulation performance, with reservoir simulation becoming more challenging as the impoundment ratio increases.

Contributions

• Development of novel differentiable reservoir operation modules (dROMs) for three widely used hydrological models (WBMplus, WaterGAP2, CWatM) using differentiable parameter learning (dPL) and LSTM.
• Demonstrated the effectiveness of dPL in simultaneously estimating both static and dynamic parameters for reservoir operation modules, leading to improved simulation accuracy.
• Provided a comprehensive evaluation of empirical, local dROMs, and regional dROMs across a large dataset of 88 US reservoirs.
• Highlighted the potential of dPL to enhance reservoir simulation capabilities within hydrological models and identified the importance of collecting more reservoir data for better synergistic effects.

Funding

Not specified in the provided text.

Citation

@article{Chen2026Differentiable,
  author = {Chen, Zexin and Zhao, Tongtiegang},
  title = {Differentiable parameter learning of reservoir operation modules},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.134915},
  url = {https://doi.org/10.1016/j.jhydrol.2026.134915}
}