Explainable machine learning for coagulant dosage prediction in coal port wastewater treatment

Coal-laden runoff at coal ports, if inadequately treated before discharge, can compromise nearshore water quality. As the primary process for treating coal-laden runoff wastewater, coagulation dosing is largely adjusted empirically under varying influent conditions, and reliable predictive tools remain limited. To address this gap, we developed an interpretable machine-learning framework for coagulant dosage prediction. Among the evaluated algorithms, the neural network model achieved the best performance, with a test-set R² of 0.902 and a root mean squared error (RMSE) of 0.057. To explain the model, Sobol sensitivity analysis and Shapley additive explanations (SHAP) were combined to identify key factors and their interaction effects. Sobol analysis shows that target effluent turbidity and influent turbidity are dominant drivers, while temperature and electrical conductivity affect predictions through interactions. SHAP analysis further shows that target effluent turbidity and influent turbidity are the primary contributors, while the effects of temperature and conductivity are condition-dependent, supporting the presence of interactions. Overall, this study provides an accurate and interpretable tool for coagulant dosing prediction in coal-port runoff treatment and offers a "prediction-explanation" workflow to support intelligent coagulation control and compliance-oriented discharge management in coastal ports. It has significant implications for promoting the intelligent transformation of coal-laden wastewater treatment systems.