Establishment of predictive machine learning models for disinfection byproduct formation during chlorination or...

In recent years, in addition to traditional chlorine disinfection, disinfection methods based on chlorine photolysis reactions using ultraviolet or xenon lamps as light sources, have received increased attention. However, both approaches inevitably generate disinfection byproducts (DBPs) that pose health risks. Monitoring DBP formation during disinfection is essential, yet it remains challenging due to time and equipment requirements. This study compiled a dataset of 4500 water samples (over 25,000 data points for DBP concentrations) with diverse natural organic matter (NOM) sources to develop predictive models under both chlorination and chlorine photolysis conditions. Target DBPs include extensively studied trihalomethanes, haloacetic acids, and unregulated DBPs. In feature engineering, reaction mechanisms related to reactive species (HO^•, Cl^•, Br^•, O₃, and ^•NO) during chlorine photolysis were incorporated. Adding input features related to these reactive species improved predictive performance. The predictive performance of twelve model algorithms was comprehensively evaluated, and tree-based ensemble machine learning methods, especially categorical boosting, achieved the best results. Furthermore, the features required for DBP subsets were redesigned, and these revised models not only maintained robust predictive performance but also reduced experimental effort. Ultimately, regression models for twenty DBPs and classification models for seven DBPs were established under chlorination, while under chlorine photolysis, regression models for thirteen DBPs and classification models for three DBPs were developed. Model interpretation showed that under chlorine photolysis, reactive species significantly impacted DBP formation by altering molecular composition of NOM precursors. Additionally, an online platform for the DBP predictive models was established to facilitate model application. This study provides a convenient approach for DBP monitoring and offers significant technical support for optimizing disinfection.