Enhancing physically-based hydrological modeling with an ensemble of machine-learning reservoir operation modules...

Area of Science:

• Hydrology and Water Resources
• Environmental Modeling
• Machine Learning Applications

Background:

• Reservoir operations significantly alter river flow dynamics globally.
• Accurate representation of reservoir operations in hydrological models is challenging due to data limitations.
• Existing hydrological models often use simplified reservoir operation modules.

Purpose of the Study:

• To develop a hybrid hydrological modeling framework integrating a process-based model with a machine-learning-based reservoir operation module.
• To improve the simulation of river flow variability influenced by reservoir operations.
• To enhance the accuracy of runoff simulations in regions with significant reservoir influence.

Main Methods:

• A hybrid framework combining the Soil and Water Assessment Tool (SWAT) with a machine learning ensemble (Random Forest, Support Vector Machine, AutoGluon) for reservoir outflow prediction.
• Machine learning models used precipitation and temperature data as inputs to predict reservoir outflows.
• The framework was evaluated using the Xijiang Basin in the Pearl River Basin, China.

Main Results:

• The hybrid model demonstrated superior performance compared to traditional SWAT reservoir modules and models without dedicated reservoir modules.
• Significant improvements in Nash-Sutcliffe Efficiency (NSE) and coefficient of determination (R²) for daily runoff simulations were observed.
• The model effectively captured peak flows and dry period runoff, indicating better characterization of reservoir impacts.

Conclusions:

• The proposed hybrid modeling approach enhances the accuracy of runoff simulations by effectively characterizing reservoir operations.
• This framework offers valuable insights for improving hydrological forecasting and water resource management in reservoir-influenced river basins.
• The integration of machine learning with process-based models presents a promising direction for hydrological modeling.