Xiang et al. (2025) An explainable deep learning model based on hydrological principles for flood simulation and forecasting

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-12-19
• Authors: Xin Xiang, Shenglian Guo, Chenglong Li, Yun Wang
• DOI: 10.5194/hess-29-7217-2025

Research Groups

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

Short Summary

This study develops an explainable deep learning (EDL) model for flood simulation by integrating the Xinanjiang (XAJ) model's runoff generation and flow routing principles into a recurrent neural network (RNN) unit (XAJRNN layer) and fusing it with LSTM layers. Tested in two Chinese river basins, the EDL model demonstrates superior flood simulation accuracy and enhanced interpretability compared to benchmark models.

Objective

• To develop an explainable deep learning (EDL) model for flood simulation and forecasting by systematically integrating the complex runoff generation and flow routing principles of the Xinanjiang (XAJ) hydrological model into a recurrent neural network (RNN) architecture, thereby enhancing both physical interpretability and forecasting performance.

Study Configuration

• Spatial Scale:
  • Lushui River basin: approximately 3950 square kilometers (primary tributary of the middle Yangtze River, China).
  • Qingjiang River basin: approximately 17000 square kilometers (primary tributary of the middle Yangtze River, China, focusing on the basin controlled by Shuibuya Reservoir, approximately 10860 square kilometers).
• Temporal Scale:
  • Lushui River basin: Flood season data (1 May to 31 October) from 2012–2019, with a 3-hour time step.
  • Qingjiang River basin: Flood season data (1 April to 31 October) from 2012–2020, with a 6-hour time step.
  • Training periods: Lushui (2012-2016), Qingjiang (2012-2016).
  • Testing periods: Lushui (2017-2019), Qingjiang (2017-2020).

Methodology and Data

• Models used:
  • Proposed: Explainable Deep Learning (EDL) model, comprising a physics-driven XAJRNN layer integrated with Long Short-Term Memory (LSTM) layers.
  • Benchmark: Xinanjiang (XAJ) model (calibrated with Genetic Algorithm), standalone LSTM model, and XAJ-LSTM hybrid model (XAJ outputs as LSTM inputs).
• Data sources:
  • Observation data: Areal mean rainfall (from 17 gauges for Lushui, 28 for Qingjiang), pan evaporation, and inflow discharge.
  • The Thiessen polygon method was used to calculate areal mean rainfall and pan evaporation.

Main Results

• The EDL model demonstrated superior overall performance in both basins during the test period. In the Lushui River basin, it achieved a Nash-Sutcliffe Efficiency (NSE) of 0.98, a Root Mean Squared Error (RMSE) of 43.71 cubic meters per second, and a Relative Error (RE) of -2.69 %.
• In the Qingjiang River basin, the EDL model achieved an NSE of 0.92, an RMSE of 167.94 cubic meters per second, and an RE of -8.74 %. These results consistently outperformed the XAJ, LSTM, and XAJ-LSTM benchmark models in most metrics.
• For flood event simulation, the EDL model exhibited the highest stability and accuracy, with smaller Flood Peak Relative Errors (PRE) and Peak Timing Differences (ΔT) consistently close to 0 hours, indicating precise simulation of flood peak magnitudes and timings.
• The integration of physical principles into the EDL model not only improved simulation accuracy but also enhanced interpretability, offering new insights for the fusion of deep learning and hydrological models.
• The XAJRNN layer alone showed improved runoff simulation compared to the standard XAJ model but was less accurate than the full EDL model, highlighting the synergistic benefits of combining it with LSTM.

Contributions

• Developed a novel XAJRNN layer that explicitly embeds the complex runoff generation and flow routing mechanisms of the Xinanjiang (XAJ) model into a conventional Recurrent Neural Network (RNN) unit framework.
• Constructed an Explainable Deep Learning (EDL) model by tightly fusing the physics-driven XAJRNN layer with LSTM layers, enabling synchronized training and joint parameter optimization, which overcomes the parameter mismatch issues of loosely coupled hybrid models.
• Demonstrated that this tightly integrated physics-informed deep learning approach significantly enhances flood simulation accuracy, particularly in capturing peak flow magnitudes and timings, while simultaneously improving model interpretability.
• Provided a promising new methodology for integrating complex hydrological processes into deep learning architectures, advancing the field of flood forecasting and hydrological modeling.

Funding

• National Natural Science Foundation of China (grant no. U2340205)

Citation

@article{Xiang2025explainable,
  author = {Xiang, Xin and Guo, Shenglian and Li, Chenglong and Wang, Yun},
  title = {An explainable deep learning model based on hydrological principles for flood simulation and forecasting},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-7217-2025},
  url = {https://doi.org/10.5194/hess-29-7217-2025}
}