Accurate prediction of coagulant dosage is essential for efficient and cost-effective water treatment. To this end, this study proposes a flexible NSGA-II optimized Stacking framework. Using the NSGA-II dual-objective optimization algorithm, the framework selects the optimal base models from a diverse model pool, which simultaneously satisfy requirements of excellent overall performance and low predictive correlation. The final Stacking model integrates MLP, Matern, and RBF Gaussian Process as base models, with Huber regression as the meta-model. The model demonstrates excellent performance, achieving R 2 0.997/0.990 on the training/test sets. SHAP Beeswarm analysis reveals that Conductivity, pH, Water Temperature, TDS, and Turbidity STE are the key predictors, collectively accounting for over 84% of the contribution. And consistent feature contribution distribution patterns between training and test sets, confirming model has strong generalization capability. SHAP Dependence analysis, enhanced with LOWESS smoothing and 95% confidence intervals, elucidates critical nonlinear relationships: Within the range of the collected features data, a U-shaped relationship exists between dosage and both Water Temperature and Turbidity STE, while Conductivity, pH, and TDS show positive correlations with dosage, which will offer valuable reference for manual decision-making in scenarios beyond the range of original dataset. Characterized by high precision, robust generalization, and clear interpretability, the proposed Stacking model serves as an effective intelligent tool for coagulant dosage determination in water treatment plants. • NSGA-II optimized Stacking framework selects diverse, high-performance base models for enhanced prediction. • Excellent coagulant dosage prediction accuracy achieved with R 2 of 0.997/0.990. • Conductivity, pH, Temperature, TDS, and Turbidity STE as dominant factors with cumulative SHAP contribution >84%. • SHAP-LOWESS-95% Confidence dependence interval reveals nonlinear dose–features relationship, providing suggestion for manual decision-making beyond model prediction abilities.
Zhao et al. (Thu,) studied this question.