Sun et al. (2026) A novel hybrid deep learning model with dynamic parameterization for accurate flood simulation

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-02-20
• Authors: bokai Sun, Shenglian Guo, Xin Xiang, Sirui Zhong, Xiaoya Wang, Jiabo Yin
• DOI: 10.1016/j.jhydrol.2026.135182

Research Groups

State Key Laboratory of Water Resources Engineering and Management, Wuhan University, China

Short Summary

This paper proposes HyDPNet, a novel hybrid deep learning model integrating a differential Xinanjiang model with dynamic parameters and an LSTM post-processor, demonstrating superior flood simulation accuracy in the Lushui River basin compared to benchmark models.

Objective

• To develop a novel hybrid deep learning model (HyDPNet) that combines physical process representation with data-driven learning to achieve accurate, physically interpretable, and process-informed flood simulation, overcoming limitations of traditional and purely data-driven models.

Study Configuration

• Spatial Scale: Lushui River basin
• Temporal Scale: Training and test periods for model evaluation (specific duration not provided in the text).

Methodology and Data

• Models used: HyDPNet (hybrid dynamic parameter network), which integrates a differential form of the Xinanjiang (XAJ) model into recurrent neural network units with parameters generated by auxiliary neural networks, and an LSTM post-processor. Benchmark models were used for comparison.
• Data sources: Large-scale hydro-meteorological datasets (specific sources not detailed in the provided text).

Main Results

• The HyDPNet model achieved Nash–Sutcliffe Efficiency (NSE) values of 0.98 in the training period and 0.97 in the test period.
• It yielded the lowest simulation errors and flood peak errors across flood events compared to benchmark models.

Contributions

• Proposed HyDPNet, a novel hybrid deep learning model for flood simulation that integrates physical mechanisms with data-driven learning.
• Introduced dynamic parameterization for the integrated physical model components within recurrent neural network units.
• Incorporated an LSTM post-processor to capture long-term dependencies and enhance simulation accuracy.
• Demonstrated superior performance in flood simulation, achieving high NSE values and lower errors compared to benchmark models.
• Provided a physically interpretable and process-informed framework for hydrological simulation, supporting disaster prevention and water management.

Funding

Not specified in the provided text.

Citation

@article{Sun2026novel,
  author = {Sun, bokai and Guo, Shenglian and Xiang, Xin and Zhong, Sirui and Wang, Xiaoya and Yin, Jiabo},
  title = {A novel hybrid deep learning model with dynamic parameterization for accurate flood simulation},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135182},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135182}
}