Comparative modeling study demonstrates high-accuracy coagulant dosage prediction using CatBoost in water purification, highlighting UV254 as the primary driver of dosage decisions.
Key Points
Develop an interpretable machine learning framework to accurately predict optimal aluminum sulphate coagulant dosage in drinking water treatment while explaining model decision-making.
Trained and evaluated nine regression models—including multiple linear regression (MLR), random forest, AdaBoost, CatBoost, GBRT, HistGBRT, LightGBM, NGBoost, and XGBoost—using daily raw water quality parameters from the Taksebt water treatment plant in Algeria.
Evaluated performance using R, NSE, RMSE, and MAE, verified statistical significance via Diebold-Mariano and Kruskal-Wallis tests, and extracted feature importance using SHAP and LIME.
CatBoost achieved the highest predictive performance with an R of 0.922, NSE of 0.849, RMSE of 2.498 mg/L, and MAE of 1.702 mg/L, significantly outperforming MLR (R of 0.713, NSE of 0.507, RMSE of 4.510 mg/L, and MAE of 3.576 mg/L).
SHAP feature importance analysis identified UV254 as the most critical input variable, contributing 24% to dosage predictions, while COU contributed only 5%.