A hybrid framework of feature selection and interpretability for dissolved oxygen prediction in drinking water treatment plants.

Accurate prediction of dissolved oxygen (DO) is essential for the sustainable operation of drinking water treatment plants. Conventional approaches often rely on a single feature selection method, which can result in biased or inconsistent identification of key predictors. This study proposes a sequential hybrid framework that integrates Mutual Information (MI), Mean Decrease in Impurity (MDI), Permutation Importance, and SHAP interpretability to achieve robust and transparent DO prediction. Filter-based (MI) and embedded (MDI) methods were first employed for initial relevance screening, followed by performance-based validation using Permutation Importance, while SHAP provided both global and local interpretability and reconciled ranking discrepancies. Seven influent water quality parameters were used to train Random Forest (RF) and XGBoost (XGB) models. Feature importance analysis consistently identified historical DO, water temperature, and turbidity as the dominant predictors, whereas pH and NO₂ had minimal influence. Dimensionality reduction preserved predictive accuracy while reducing model complexity by up to 70%, thereby enhancing computational efficiency. Both models demonstrated strong performance (R² = 0.928 for RF and 0.942 for XGB; RMSE < 0.27 mg/L) with narrow 95% confidence intervals. The proposed framework provides a reliable, interpretable, and cost-effective solution for real-time DO monitoring in drinking water treatment systems and offers a transferable methodology for other environmental modeling applications.