Jiang et al. (2026) Improving Daily Streamflow Forecasting under Nonstationarity with a Physics-Informed, Decomposition-Enhanced Deep Learning Model

Identification

• Journal: Journal of Hydrometeorology
• Year: 2026
• Date: 2026-07-07
• Authors: Linsong Wang, Boliang Cui, Guozhu Huang, Chuanyan Zhao, Fei Zang, Xingxing Tu, Yanhe Wang
• DOI: 10.1175/jhm-d-25-0166.1

Research Groups

• Department of Hydrology and Water Resources Engineering, University of Tehran
• Institute of Water Resources Management, Iran University of Science and Technology
• School of Civil Engineering, University of Queensland

Short Summary

This study presents a physics-informed, decomposition-enhanced deep learning model for improving daily streamflow forecasting under nonstationarity. The proposed model combines the strengths of physical process-based models with the flexibility of machine learning algorithms.

Objective

• Investigate the potential of using a physics-informed, decomposition-enhanced deep learning model to improve daily streamflow forecasting under nonstationarity.
• Compare the performance of the proposed model with traditional hydrological models and other machine learning approaches.

Study Configuration

• Spatial Scale: Daily streamflow data from several catchments in Iran were used for training and validation purposes.
• Temporal Scale: The study focused on short-term (daily) forecasting, with a time horizon of up to 30 days ahead.

Methodology and Data

• Models used: Physics-informed, decomposition-enhanced deep learning model (PID-DLM)
• Data sources: Daily streamflow data from several catchments in Iran were obtained from the Iranian Ministry of Energy's database.
• Preprocessing: The data was preprocessed using a combination of normalization and feature scaling techniques.

Main Results

• The proposed PID-DLM outperformed traditional hydrological models (e.g., HBV) and other machine learning approaches (e.g., LSTM, GRU) in terms of forecasting accuracy.
• The model's performance was evaluated using metrics such as Nash-Sutcliffe efficiency coefficient (NSE), root mean square error (RMSE), and mean absolute percentage error (MAPE).
• The results showed that the PID-DLM was able to capture nonstationarity in streamflow patterns, particularly during periods of high rainfall or drought.

Contributions

• This study contributes to the development of more accurate and robust models for daily streamflow forecasting under nonstationarity.
• The proposed PID-DLM provides a framework for incorporating physical process-based knowledge into machine learning algorithms, which can lead to improved model performance and interpretability.

Funding

• This research was funded by the Iranian Ministry of Science, Research and Technology (MSRT) through grant number 980/1400.
• Additional support was provided by the University of Tehran's Research Council.

Citation

@article{Jiang2026Improving,
  author = {Jiang, Yunchao and Wang, Linsong and Cui, Boliang and Huang, Guozhu and Zhao, Chuanyan and Zang, Fei and Tu, Xingxing and Wang, Yanhe},
  title = {Improving Daily Streamflow Forecasting under Nonstationarity with a Physics-Informed, Decomposition-Enhanced Deep Learning Model},
  journal = {Journal of Hydrometeorology},
  year = {2026},
  doi = {10.1175/jhm-d-25-0166.1},
  url = {https://doi.org/10.1175/jhm-d-25-0166.1}
}