A physically guided and interpretable SWAT-BiLSTM framework with Bayesian optimization for bias correction in daily...

Area of Science:

• Hydrology and Water Resources
• Environmental Modeling
• Artificial Intelligence in Environmental Science

Background:

• Accurate extreme streamflow simulation is critical for flood forecasting, water management, and water quality.
• Existing process-based and data-driven models face limitations in accuracy, robustness, and interpretability, especially during extreme events.
• Addressing these challenges requires advanced modeling techniques for reliable hydrological predictions.

Purpose of the Study:

• To develop and evaluate a novel hybrid modeling framework for enhanced streamflow prediction.
• To integrate a process-based model (SWAT) with an optimized deep learning approach (BiLSTM) and an interpretability tool (SHAP).
• To improve the accuracy, stability, and interpretability of streamflow simulations, particularly under extreme hydrological conditions.

Main Methods:

• A hybrid framework integrating the Soil and Water Assessment Tool (SWAT) with a Bidirectional Long Short-Term Memory (BiLSTM) network.
• Bayesian optimization (BO) for optimizing the BiLSTM network and random forest (RF) with correlation analysis for feature selection.
• SHapley Additive Explanations (SHAP) for analyzing model behavior and providing interpretable insights into hydrological processes.

Main Results:

• The coupled models significantly outperformed standalone models, with R² and Nash-Sutcliffe Efficiency (NSE) improving by 14.7%-27.1% and 10.0%-35.0%, respectively.
• The SWAT-S-BiLSTM model demonstrated superior performance, achieving R² of 0.89 and NSE of 0.81.
• Extreme flow prediction showed substantial improvement, with relative error for the top 0.5% of flows reducing from -11.72% (SWAT) to -0.27% (SWAT-S-BiLSTM).
• SHAP analysis revealed hydrological lag effects, threshold behaviors, and nonlinear responses, clarifying how the hybrid model corrects physical model deficiencies.

Conclusions:

• The proposed hybrid framework effectively enhances streamflow prediction accuracy and stability, especially for extreme events.
• The integration of deep learning with physical models and interpretability tools offers a robust solution for hydrological forecasting.
• The framework provides valuable insights into hydrological processes and demonstrates significant potential for practical applications in water resource management.