Zhang et al. (2026) A physics-driven hybrid transformer model for hydrologic simulation under nonstationary environmental conditions

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-02-14
• Authors: Hongwei Zhang, Zexing Tao, Kaiwen Wang, Gang Zhao, Jiewei Chen, Duanyang Xu, Ronggao Liu, Longhao Wang, Lei Wang, Quansheng Ge
• DOI: 10.1016/j.jhydrol.2026.135133

Research Groups

• Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China
• University of Chinese Academy of Sciences, Beijing 100049, China

Short Summary

This study proposes a Physics-Driven Hybrid Transformer Hydrologic Model (PD-HTHM) that integrates deep learning with the conceptual SIMHYD model to generate time-varying parameters, improving hydrologic simulation under nonstationary environmental conditions. It demonstrates that dynamically adjusting a few key parameters significantly enhances model robustness and predictive accuracy compared to static parameterizations.

Objective

• To develop a hydrologic model capable of robustly simulating nonstationary hydrological processes by dynamically adjusting model parameters using a physics-driven deep learning approach.

Study Configuration

• Spatial Scale: 71 catchments in the River Severn Basin (UK).
• Temporal Scale: Long-term simulations to capture nonstationary conditions (specific duration not provided).

Methodology and Data

• Models used: Physics-Driven Hybrid Transformer Hydrologic Model (PD-HTHM), which integrates an LSTM–Transformer hybrid encoder with the conceptual SIMHYD model.
• Data sources: Not explicitly listed, but the model addresses nonstationary environmental conditions driven by climate change and human activities, implying the use of relevant climate and land-use data.

Main Results

• All eight designed dynamic parameterization strategies outperformed the static SIMHYD model on common evaluation metrics.
• The configuration jointly adjusting the soil moisture storage capacity (SMSC) and the baseflow recession coefficient (K) achieved the best overall performance.
• Optimal parameter combinations varied among catchments, indicating pronounced spatial heterogeneity.
• Dynamically adjusting a few key parameters associated with dominant hydrological processes is more effective than simply increasing the number of variable parameters, improving model robustness and predictive accuracy under nonstationary conditions.

Contributions

• Proposes a novel Physics-Driven Hybrid Transformer Hydrologic Model (PD-HTHM) that integrates deep learning (LSTM-Transformer) with a conceptual hydrological model (SIMHYD) for dynamic parameterization.
• Introduces a physics-driven approach to generate time-varying parameters, addressing the limitations of static parameters in conventional models under nonstationary conditions.
• Systematically evaluates process-oriented dynamic parameterization strategies, providing insights into which parameters are most effective for dynamic adjustment.
• Demonstrates the potential of integrating deep learning with process-based modeling to improve the representation of nonstationary hydrological dynamics.

Funding

• No funding information was provided in the text.

Citation

@article{Zhang2026physicsdriven,
  author = {Zhang, Hongwei and Tao, Zexing and Wang, Kaiwen and Zhao, Gang and Chen, Jiewei and Xu, Duanyang and Liu, Ronggao and Wang, Longhao and Wang, Lei and Ge, Quansheng},
  title = {A physics-driven hybrid transformer model for hydrologic simulation under nonstationary environmental conditions},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135133},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135133}
}