Hydrology-informed machine learning for enhancing surface runoff elucidation.

Area of Science:

• Environmental Science
• Hydrology
• Water Quality Management

Background:

• Urban runoff poses significant non-point source pollution challenges.
• Predicting runoff water quality is complex due to intricate physical processes.
• Effective prediction is vital for mitigating pollution in urban watersheds.

Purpose of the Study:

• To develop a flexible framework for enhanced prediction of urban runoff water quality.
• To integrate hydrology-hydraulic data with machine learning for improved accuracy and efficiency.
• To identify key driving forces influencing runoff water quality using interpretable analysis.

Main Methods:

• Utilized a flexible framework integrating physical-driven hydrological-hydraulic models with machine learning networks.
• Employed high-resolution measurement data from online monitoring equipment in a Pearl River Delta urban watershed.
• Applied Shapley Additive Explanations (SHAP) for interpretable analysis of driving forces.

Main Results:

• The random forest model demonstrated superior performance among tested machine learning models.
• Achieved high prediction accuracy with R² values of 0.78 (COD), 0.77 (NH3-N), and 0.81 (SS).
• SHAP analysis identified precipitation, slope, and impervious area ratio as significant factors affecting water quality.

Conclusions:

• The proposed modeling framework effectively captures the dynamic characteristics of pollutants in surface water.
• The integration of physical data and machine learning offers a robust approach to runoff water quality prediction.
• Understanding driving forces aids in developing targeted pollution mitigation strategies.