Applied gradient boosting decision tree algorithms for accurate prediction of trihalomethanes: A case study in dual-sources drinking water distribution system in metropolitan

Abstract

The complex hydraulic environment in dual-source drinking water distribution system (DWDS) in metropolitan puts water safety in an unknown situation, which requires systematic water quality monitoring and controlling to reduce the water risk during water distribution process. Given that, this study conducted a one-year sampling of trihalomethanes (THMs) and identified the hydraulic junction area in a real dual-source DWDS. For efficient prediction of THMs, several predictive frameworks were established via mechanistic algorithms and machine learning (ML). The results indicated that Gradient Boosting Decision Tree (GBDT) algorithm achieved higher interpretability and generalizability in describing the relationship between various of features with R2 reaching 0.903, 0.908, 0.945 and 0.959 for trichloromethane (TCM), bromodichloromethane (BDCM), dibromomonochloromethane (DBCM) and tribromomethane (TBM), respectively. Shapley additive explanations (SHAP) analysis of GBDT models further identified water temperature, residual chlorine, water supply distance, and pH were key features affecting THMs formation in practical DWDS. Furthermore, the spatio-temporal distribution of THMs was simulated by GBDT models and identified the region might be exposed to high risk of the THMs. The simulation results indicated that the DBCM and TBM were more likely to form in the region nearly to DWTP-X, whereas in the region nearly to DWTP-Y, the TCM and BDCM were the dominant THMs species due to differences in raw water characteristics. Undoubtedly, this study provided novel insights into utilizing ML models to predict THMs levels in dual-source DWDS in the metropolitan.