Rostampour et al. (2025) Prediction and comparison of streamflow using hybrid and independent models in Zola dam basin

Identification

• Journal: Modeling Earth Systems and Environment
• Year: 2025
• Date: 2025-09-08
• Authors: Javad Rostampour, Hamed Sahranavard, Farshad Ahmadi, Mahdi Mollazadeh
• DOI: 10.1007/s40808-025-02589-4

Research Groups

• University of Birjand, Birjand, Iran, Islamic Republic of
• Shahid Chamran University of Ahvaz, Ahvaz, Iran, Islamic Republic of

Short Summary

This study models and forecasts streamflow in the Zola dam basin using independent (Extreme Learning Machine, Long Short-Term Memory) and hybrid (Wavelet, Variational Mode Decomposition) machine learning models. It demonstrates that hybrid models significantly enhance prediction accuracy, with the ELM-VMD approach achieving the best performance.

Objective

• To model and forecast streamflow in the Zola dam watershed using independent (Extreme Learning Machine - ELM, Long Short-Term Memory - LSTM) and hybrid (ELM-Wavelet, ELM-VMD, LSTM-Wavelet, LSTM-VMD) machine learning and deep learning algorithms.
• To investigate the possibility of increasing prediction accuracy through the use of decomposition methods (Wavelet function and Variational Mode Decomposition - VMD) to build hybrid models.

Study Configuration

• Spatial Scale: Zola dam watershed, covering approximately 960 square kilometers.
• Temporal Scale: Monthly data over a 41-year period (1981 to 2021).

Methodology and Data

• Models used: Extreme Learning Machine (ELM), Long Short-Term Memory (LSTM), Wavelet theory, Variational Mode Decomposition (VMD). Hybrid models: ELM-Wavelet, ELM-VMD, LSTM-Wavelet, LSTM-VMD.
• Data sources: Monthly historical streamflow and precipitation raw observations from the Zola dam watershed.

Main Results

• Hybrid models consistently increased the performance and accuracy of streamflow projections compared to independent models.
• The ELM-VMD approach exhibited the best and most accurate performance with evaluation indices: Root Mean Square Error (RMSE) = 0.34, Mean Absolute Error (MAE) = 0.26, and Kling–Gupta Efficiency (KGE) = 0.992.
• The ELM-Wavelet algorithm was the second-best performing model (RMSE = 0.539, MAE = 0.413, KGE = 0.973).
• Independent ELM models generally outperformed independent LSTM models.
• For independent ELM and LSTM models, the simultaneous use of streamflow and precipitation data improved accuracy.
• For hybrid models (ELM-Wavelet, ELM-VMD, LSTM-Wavelet, LSTM-VMD), scenarios based solely on streamflow data showed more satisfactory performance than those based on both streamflow and precipitation.

Contributions

• Provides a comprehensive comparison of independent and hybrid machine learning and deep learning models for streamflow forecasting in the Zola dam basin.
• Demonstrates the significant improvement in streamflow prediction accuracy achieved by integrating decomposition methods (Wavelet and VMD) with ELM and LSTM algorithms.
• Identifies ELM-VMD as the most effective approach for streamflow forecasting in the studied region, offering highly accurate and reliable predictions.
• Quantifies the performance enhancement of hybrid models using standard hydrological evaluation criteria (RMSE, MAE, KGE).

Funding

Not specified in the paper.

Citation

@article{Rostampour2025Prediction,
  author = {Rostampour, Javad and Sahranavard, Hamed and Ahmadi, Farshad and Mollazadeh, Mahdi},
  title = {Prediction and comparison of streamflow using hybrid and independent models in Zola dam basin},
  journal = {Modeling Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s40808-025-02589-4},
  url = {https://doi.org/10.1007/s40808-025-02589-4}
}