Interpretable prediction of coagulant dosage in drinking water treatment plant based on automated machine learning and SHAP method

Abstract

Accurate prediction of coagulant dosing is critical to ensure the efficiency and cost-effectiveness of the water treatment process. This study developed an accurate prediction model for coagulant dosing in drinking water treatment plants using automated machine learning (AutoML) and enhanced its transparency with the SHAP (Shapley Additive explanation) method. Results showed that the Random Forest (RF) model developed by AutoML outperformed the best single Gradient Boosting Tree model, with a 37 % lower RMSE of 0.89, a 52 % lower MAE of 0.47, and a 5 % higher R² of 0.96. The RF model has the potential to save 10.25 % of coagulant dosage per year, specifically, 222 kg/d of PACl at a cost of 180.7 yuan/d for Yangtze River water, and 225 kg/d at a cost of 183.2 yuan/d for Luan River water, while ensuring a more stable water quality in the treated water. SHAP analysis identified conductivity, ammonia nitrogen, chemical oxygen demand, and temperature of raw water as key factors affecting PACl dosage, which significantly impacts treated water pH. The combination of AutoML and SHAP method applied to intelligent water management, can bolster water treatment efficiency and management practices, resulting in noteworthy environmental, economic, and social benefits.