Hybrid Data-driven and Process-based Modeling for Streamflow Forecasting Under Climate Change in the Upper Genale River Basin of Ethiopia

Technical Summary: Hybrid Data-driven and Process-based Modeling for Streamflow Forecasting Under Climate Change in the Upper Genale River Basin of Ethiopia

Problem Statement
Accurate streamflow forecasting is critical for sustainable water management, flood risk reduction, and irrigation planning, particularly in Sub-Saharan Africa where climate variability interacts with limited infrastructure and high dependence on rain-fed agriculture. The Upper Genale River Basin (UGRB) in southeastern Ethiopia is strategically vital for irrigation and hydropower but remains highly sensitive to climate variability. While physically based models like SWAT+ effectively represent watershed processes, they often struggle with highly nonlinear hydrological interactions and extreme events under non-stationary climate conditions. Conversely, purely data-driven machine learning (ML) models lack physical interpretability and may fail when future climatic conditions diverge significantly from historical training data. This study addresses the need for a robust forecasting framework that overcomes these individual limitations to support climate-resilient water management in the UGRB.

Methodology
The study developed a hybrid hydrological forecasting framework integrating the process-based QSWAT+ model with advanced machine learning algorithms and CMIP6 climate projections.

• Study Area and Data: The research focused on the Upper Genale Sub-River Basin (approx. 9,127 km²). The dataset included daily precipitation and temperature data (1985–2014) from five meteorological stations (supplemented by CHIRPS/CHIRTS satellite data) and daily streamflow observations (1988–2011) from the Chenemsa gauging station. Topographic, land-use, and soil data were derived from SRTM, Copernicus, and the Harmonized World Soil Database.
• Climate Projections: Twenty-two CMIP6 Global Climate Models (GCMs) were evaluated against observational data using a Comprehensive Ranking Index. CanESM5 was selected for temperature, and a six-model ensemble was used for precipitation. Projections were analyzed under SSP2-4.5 and SSP5-8.5 scenarios for near-term (2015–2044) and mid-term (2045–2074) periods, utilizing bias correction and statistical downscaling.
• Hydrological Modeling: QSWAT+ was used to simulate watershed processes, dividing the basin into 13 sub-basins and 731 Hydrologic Response Units (HRUs). The model underwent global sensitivity analysis (Sobol method) and calibration using the CALSI algorithm for the period 1988–2002, with validation from 2003–2011.
• Hybrid Framework: Four ML algorithms—Random Forest (RF), XGBoost, LightGBM, and Long Short-Term Memory (LSTM)—were implemented. These models used meteorological observations, engineered temporal features, and QSWAT+ outputs as predictors. A stacking ensemble framework (using Ridge Regression as a meta-learner) and Bayesian Model Averaging (BMA) were applied to integrate QSWAT+ simulations with ML predictions, aiming to reduce uncertainty and improve robustness.
• Analysis: Trends were assessed using the Mann–Kendall test and Sen's slope estimator. Model performance was evaluated using $R^2$, Nash–Sutcliffe Efficiency (NSE), RMSE, and Percent Bias (PBIAS).

Key Results

• QSWAT+ Performance: The standalone QSWAT+ model achieved satisfactory calibration ($NSE = 0.74$, $R^2 = 0.77$) and validation ($NSE = 0.71$, $R^2 = 0.85$) performance, indicating a reliable representation of watershed processes, though it showed slight underestimation of peak flows in the validation period.
• Hybrid Model Superiority: The integration of ML significantly enhanced predictive accuracy. The standalone ML models (particularly XGBoost) outperformed QSWAT+, achieving $NSE$ and $R^2$ values near 0.99. The hybrid stacking ensemble model achieved the highest performance with an $NSE$ and $R^2$ of 0.98, effectively capturing residual nonlinear relationships that the physical model missed. The hybrid approach also reduced systematic bias, with PBIAS values near zero.
• Climate Change Projections:
  • Temperature: Consistent warming is projected under both SSP scenarios, with minimum temperatures increasing faster than maximums. The mid-term period under SSP5-8.5 shows the most pronounced warming.
  • Precipitation: Projections show considerable variability. The ensemble mean indicates moderate rainfall reductions under SSP2-4.5 and increases under SSP5-8.5, particularly in the mid-term.
  • Hydrological Response: Under SSP2-4.5, reductions in precipitation led to decreases in runoff, water yield, and groundwater recharge. Conversely, under SSP5-8.5, increased precipitation drove higher runoff and water yield in the mid-term, though rising temperatures increased evapotranspiration across all scenarios, limiting proportional gains in water availability.
• Streamflow Trends: Future streamflow projections indicate contrasting responses. Near-term simulations suggest slight reductions, while mid-century projections under SSP5-8.5 indicate substantial increases (e.g., up to 253% increase in April flow). However, monthly variability is amplified, with some months showing significant declines. The hybrid model successfully reproduced historical variability and provided robust ensemble-averaged projections.

Significance and Claims
The paper claims that the proposed hybrid framework successfully combines the physical realism of process-based models with the predictive power of machine learning to improve streamflow forecasting in data-scarce, climate-vulnerable regions. Key contributions include:

1. Enhanced Accuracy: The hybrid approach (specifically the stacking ensemble) significantly outperformed standalone QSWAT+ and individual ML models, achieving high $NSE$ and $R^2$ values while maintaining hydrological consistency.
2. Climate Adaptation Insights: The study provides specific projections for the Upper Genale Basin, highlighting a future characterized by increased evapotranspiration, potential water scarcity under moderate emission scenarios (SSP2-4.5), and heightened flood risks under high-emission scenarios (SSP5-8.5).
3. Decision Support: The findings offer a basis for climate-resilient irrigation planning, groundwater conservation, and flood management infrastructure development. The authors emphasize that integrating hybrid hydrological-ML models into reservoir operations and water allocation planning is essential for sustainable basin management.

The study concludes that while uncertainties remain regarding precipitation projections and model structures, the hybrid framework provides a valuable tool for assessing climate change impacts and reducing prediction uncertainty in Ethiopian river basins.