Accurate and interpretable prediction of chemical oxygen demand using explainable boosting algorithms with SHAP...

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Effective water quality management relies on accurate Chemical Oxygen Demand (COD) prediction.
• COD is a key indicator of organic pollution in water bodies.

Purpose of the Study:

• To compare the performance of six ensemble boosting models for COD estimation.
• To identify the most influential water quality parameters for COD prediction.
• To assess model interpretability and uncertainty.

Main Methods:

• Six ensemble boosting models (AdaBoost, CatBoost, XGBoost, LightGBM, HistGBRT, NGBoost) were employed.
• Models were trained and validated using water quality data from two South Korean monitoring stations.
• Performance was evaluated using RMSE, MAE, R, NSE, PBIAS, and SHAP for interpretability.

Main Results:

• NGBoost demonstrated the highest accuracy at the Toilchun station (R=0.979, NSE=0.958).
• CatBoost performed best at the Hwangji station (R=0.861, NSE=0.733).
• SHAP analysis revealed Total Organic Carbon (TOC), Biochemical Oxygen Demand (BOD₅), and Suspended Solids (SS) as key predictors of COD.

Conclusions:

• Ensemble boosting models, particularly NGBoost and CatBoost, are effective for COD prediction.
• NGBoost's ability to provide probability distributions enhances the quantification of COD variability and model uncertainty.
• Understanding the influence of specific parameters like TOC, BOD₅, and SS is vital for water pollution control.