Wang et al. (2025) Interpretable deep learning hybrid streamflow prediction modeling based on multi-source data fusion

Identification

• Journal: Environmental Modelling & Software
• Year: 2025
• Date: 2025-11-22
• Authors: Zhaocai Wang, Cheng Ding, Nannan Xu, Weilong Wang, Xingxing Zhang
• DOI: 10.1016/j.envsoft.2025.106796

Research Groups

• College of Information, Shanghai Ocean University, Shanghai, PR China
• Centre for Research on Environmental Ecology and Fish Nutrition of the Ministry of Agriculture, Shanghai Ocean University, Shanghai, PR China
• Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, PR China

Short Summary

This study introduces CICLAR, an enhanced interpretable deep learning hybrid model, for accurate daily streamflow and extreme flood prediction by fusing multi-source data and optimizing neural network hyperparameters. The CICLAR model significantly outperforms benchmark models, demonstrating improved accuracy in both general streamflow and extreme flood forecasting.

Objective

• To develop an enhanced interpretable deep learning hybrid model, named CEEMDAN-ISMA-CNN-LSTM-AM-RF (CICLAR), for predicting both daily streamflow and extreme flood events by integrating multi-source heterogeneous data and optimizing model components.

Study Configuration

• Spatial Scale: 11 hydrological stations located in the upstream, midstream, and downstream sections of the Jialing River in China.
• Temporal Scale: Daily streamflow predictions.

Methodology and Data

• Models used:
  • Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) for complexity reduction.
  • Improved Slime Mould Algorithm (ISMA) for hyperparameter optimization.
  • Convolutional Neural Network (CNN) - Long Short-Term Memory (LSTM) - Attention Mechanism (AM) for feature extraction and sequence learning.
  • Random Forest (RF) for non-linear summation.
  • CICLAR (CEEMDAN-ISMA-CNN-LSTM-AM-RF) as the integrated hybrid model.
• Data sources: Multi-source heterogeneous data, including remote sensing, meteorological, hydrological, and streamflow data.

Main Results

• The CICLAR model significantly outperforms other benchmark models in daily streamflow prediction.
• For the Beibei Hydrological Station, compared to the conventional Long Short-Term Memory (LSTM) model:
  • The Nash-Sutcliffe Efficiency Coefficient (NSE) of CICLAR's prediction results increased by 30 %.
  • The Mean Absolute Error (MAE) decreased by 75 %.
• For extreme flood forecasting, compared to the LSTM model:
  • The Mean Relative Error (MRE) was reduced by 0.86.
  • The Qualification Rate (QR) improved by 150 %.
• The CICLAR model demonstrates significant application value in extreme flood forecasting and water resources management.

Contributions

• Introduction of a novel interpretable deep learning hybrid model (CICLAR) that integrates multi-source data fusion, mode decomposition, hyperparameter optimization, and non-linear integration for streamflow prediction.
• Demonstrated superior performance in both daily streamflow and extreme flood prediction compared to existing benchmark models, addressing the challenges of non-linearity and complexity in hydrological forecasting.
• Provides a robust and accurate tool with significant application value for early flood warnings and efficient water resources management.

Funding

Not specified in the provided text.

Citation

@article{Wang2025Interpretable,
  author = {Wang, Zhaocai and Ding, Cheng and Xu, Nannan and Wang, Weilong and Zhang, Xingxing},
  title = {Interpretable deep learning hybrid streamflow prediction modeling based on multi-source data fusion},
  journal = {Environmental Modelling & Software},
  year = {2025},
  doi = {10.1016/j.envsoft.2025.106796},
  url = {https://doi.org/10.1016/j.envsoft.2025.106796}
}