Mihret et al. (2025) Hybrid GR4J-LSTM modeling for streamflow prediction of extreme events in data-scarce regions: Upper Blue Nile Basin, Ethiopia

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-11-22
• Authors: Temesgen T. Mihret, Fasikaw A. Zemale, Abeyou W. Worqlul, Ayenew D. Ayalew, Margaret Chen, Nicola Fohrer
• DOI: 10.1016/j.ejrh.2025.102977

Research Groups

• Faculty of Civil and Water Resources Engineering, Bahir Dar University, Ethiopia
• Department of Water Resources and Irrigation Engineering, Assosa University, Ethiopia
• International Center for Agricultural Research in the Dry Areas (ICARDA), Tunis, Tunisia
• Department of Hydrology and Water Resources Management, Institute for Natural Resource Conservation, Kiel University, Kiel, Germany
• Hydrology and Hydraulic Engineering, Vrije Universiteit Brussel, Brussels, Belgium

Short Summary

This study develops and evaluates DeepGR4J, a hybrid rainfall-runoff model combining the GR4J conceptual framework with a Long Short-Term Memory (LSTM) neural network, for streamflow prediction and extreme event simulation in data-scarce regions of the Upper Blue Nile Basin, Ethiopia. The model demonstrates superior performance and transferability compared to standalone models, effectively predicting streamflow and extreme events like floods and droughts.

Objective

• Develop a hybrid GR4J-LSTM model for streamflow prediction in the Upper Blue Nile Basin.
• Evaluate the performance of the hybrid model against standalone GR4J and LSTM models.
• Assess the transferability of the model to neighboring ungauged basins.
• Investigate the model’s capability to simulate extreme hydrological events, including floods and droughts.

Study Configuration

• Spatial Scale: Upper Blue Nile Basin, Ethiopia, focusing on the Gilgel Abay (1733 km²), Ribb (950 km²), and Gumara (1394 km²) watersheds. Elevation ranges from 1780 meters to 3700 meters above sea level.
• Temporal Scale: Daily data from 1994 to 2009 (16 years) for model calibration and validation. Specific extreme events (1996 flood and 2005 drought) were analyzed within this period.

Methodology and Data

• Models used: GR4J (Génie Rural à 4 paramètres Journalier) conceptual model, Long Short-Term Memory (LSTM) neural network, and DeepGR4J (hybrid GR4J-LSTM).
• Data sources:
  • Meteorological data: Daily rainfall and temperature from 11 gauging stations of the Ethiopian Meteorological Service Agency (EMSA) for 1994–2009. Missing data (<5%) filled using INSTAT Markov Chain simulation model. Spatially averaged using Inverse Distance Weighting (IDW).
  • Hydrological data: Daily streamflow data from Gilgel Abay, Ribb, and Gumara stations (1994–2009) from the Ethiopian Ministry of Water and Energy (MoWE). Missing data (<10%) filled using INSTAT Markov Chain simulation model.
  • Potential Evapotranspiration (PET): Estimated using the Hargreaves method.
  • Data split: 70% for model calibration and 30% for validation.
  • LSTM inputs: GR4J outputs (net precipitation (Pn), evapotranspiration (En), production store content (Ps), and percolation (Perc)), along with daily rainfall, potential evapotranspiration, and streamflow data.
  • LSTM training: Min-Max Scaling for normalization, 30-day sliding input window, 50 units in the hidden layer. Optimizers tested: Adam, RMSprop, Adagrad, and Nadam (Nadam performed best). Loss functions: Mean Squared Error (MSE) and Nash-Sutcliffe Efficiency (NSE). Dropout regularization and early stopping were applied.
  • GR4J calibration: Differential Evolution (DE) algorithm.
  • Performance metrics: Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), Percent Bias (PBias), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Squared Error (MSE).

Main Results

• The DeepGR4J hybrid model consistently outperformed both standalone GR4J and LSTM models in streamflow prediction.
• Gilgel Abay watershed (calibration/validation): DeepGR4J achieved an NSE of 0.87, KGE of 0.89, and PBias of -0.34%, with RMSE of 1.34 mm/day, MAE of 0.58 mm/day, and MSE of 1.68 mm/day.
• DeepGR4J accurately represented streamflow across all flow conditions (high, medium, and low) and demonstrated superior performance in capturing the observed Flow Duration Curve (FDC).
• Transferability to Ribb and Gumara watersheds: The model showed strong transferability, achieving NSE values of 0.91 (Ribb) and 0.92 (Gumara), and KGE values of 0.92 for both, with low PBias values (-3.13% for Ribb, -2.35% for Gumara).
• Prediction of extreme events:
  • For the 1996 flood year, DeepGR4J achieved a KGE of 0.845 and NSE of 0.897, with a PBias of -13.50% (indicating a slight underestimation of peak flows).
  • For the 2005 drought year, the model attained a KGE of 0.762 and NSE of 0.820, with a PBias of 19.02% (suggesting a modest overestimation of low flows).
  • The model effectively captured hydrological extremes, although drought simulations showed higher uncertainty.
• Ungauged downstream station (Gilgel Abay): DeepGR4J predicted an average streamflow of 3.96 mm/day, with a maximum of 29.06 mm/day and a minimum of 0.04 mm/day, aligning with previous findings on ungauged catchment contributions to Lake Tana.
• The Nadam optimizer demonstrated superior generalization capabilities during LSTM training compared to Adam, RMSprop, and Adagrad.

Contributions

• Developed and rigorously evaluated DeepGR4J, a novel hybrid GR4J-LSTM model, specifically tailored for streamflow prediction in data-scarce regions.
• Demonstrated the superior performance of DeepGR4J over traditional conceptual (GR4J) and data-driven (LSTM) models in the Upper Blue Nile Basin.
• Validated the model's transferability to hydrologically similar ungauged basins (Ribb and Gumara watersheds), addressing a critical gap in hydrological modeling for such regions.
• Successfully demonstrated the model's capability to accurately simulate extreme hydrological events (floods and droughts), which is crucial for water resource management and disaster risk reduction.
• Provided a reliable and transferable modeling approach for regions with limited hydrological data, enhancing streamflow prediction, water resource planning, and disaster risk management.
• Highlighted the effectiveness of a hierarchical training approach and the Nadam optimizer for optimizing hybrid hydrological models.

Funding

No specific funding projects, programs, or reference codes were listed in the provided paper text.

Citation

@article{Mihret2025Hybrid,
  author = {Mihret, Temesgen T. and Zemale, Fasikaw A. and Worqlul, Abeyou W. and Ayalew, Ayenew D. and Chen, Margaret and Fohrer, Nicola},
  title = {Hybrid GR4J-LSTM modeling for streamflow prediction of extreme events in data-scarce regions: Upper Blue Nile Basin, Ethiopia},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102977},
  url = {https://doi.org/10.1016/j.ejrh.2025.102977}
}