A new interpretable streamflow prediction approach based on SWAT-BiLSTM and SHAP

Area of Science:

• Hydrology and Water Resources
• Environmental Science
• Artificial Intelligence

Background:

• Accurate streamflow prediction is vital for water resource management and understanding climate change impacts.
• Traditional hydrological models often have limitations in capturing complex streamflow dynamics.

Purpose of the Study:

• To enhance streamflow simulation performance by coupling conceptual hydrological models with machine learning algorithms.
• To evaluate the effectiveness of different coupled models, including SWAT-Transformer, SWAT-LSTM, SWAT-GRU, and SWAT-BiLSTM.

Main Methods:

• Developed four coupled models (SWAT-Transformer, SWAT-LSTM, SWAT-GRU, SWAT-BiLSTM) using the Soil and Water Assessment Tool (SWAT) and machine learning.
• Utilized meteorological data (precipitation, temperature, humidity, wind speed) to generate hydrological features.
• Employed machine learning to predict daily streamflow based on meteorological and hydrological features in the Sandu-River Basin.

Main Results:

• SWAT-BiLSTM demonstrated superior streamflow simulation performance, achieving R² of 0.92 and NSE of 0.91 during calibration and 0.90 during validation.
• All four coupled models outperformed the calibrated SWAT model in streamflow prediction.
• Coupled models showed minimal systematic bias (PBIAS < 10%), unlike the tendency of SWAT to underestimate streamflow.
• SHapley Additive exPlanations (SHAP) revealed precipitation as the most influential feature (29.7% global importance) and highlighted the dominance of SWAT-generated hydrological features.

Conclusions:

• Coupling conceptual hydrological models with machine learning significantly enhances streamflow prediction accuracy.
• The SWAT-BiLSTM model offers a robust and interpretable solution for daily streamflow forecasting.
• The SHAP analysis increases confidence in the coupled models' predictions and their underlying mechanisms.