Predicting multi-species disinfection byproduct formation at small catchment scale using fluorescence spectroscopy data analyzed by machine learning

In Ireland, 82% of public water supplies originate from surface water sources which often contain elevated concentrations of dissolved organic matter (DOM) from a range of allochthonous (e.g., leaf leachate, manure) and autochthonous (e.g., macrophytes, biofilms, algae) catchment sources. During disinfection, this DOM may react with chlorine to produce potentially carcinogenic disinfection byproducts (DBPs) such as trihalomethanes (THMs), haloacetic acids (HAAs) and a range of nitrogen-containing species such as haloacetonitriles (HANs) and halonitromethanes (HNMs). As a result, Ireland has the highest reported number of total THM exceedances, (e.g., concentrations in excess of 100 μg L^-1) in potable water across European Union member states. Removal of DOM precursors from raw water prior to chlorination has shown to be effective in mitigating DBP formation. However, significant infrastructural challenges remain in Ireland with many small treatment plants requiring costly upgrades. Hence, there is an urgent need for low-cost proactive monitoring tools to quantify DOM composition and concentration of source water to aid in the production of safe drinking water.

The overall aim of the present study is to better understand the spatiotemporal dynamics of DOM precursors and associated DBP formation at the scale of small river catchments (e.g., <50 km²) typical of drinking water source areas. To achieve this, we investigated two sub-catchments (34 km² and 18 km²) of the River Lee basin, Republic of Ireland, which serve water treatment plants known to be at risk of THM exceedances. High resolution field sampling and measurement of DBP precursors (DOC, DIC, DON, NH₄⁺, Cl^– and Br^–) and DOM optical properties using UV-vis and fluorescence excitation−emission matrix (EEM) spectroscopy were combined with 214 three-day batch chlorination experiments from 36 monitoring points (including 12 groundwater) from February to November 2021. A machine learning ensemble including bagging tree, generalized boosted regression and neural networks models was developed to explore and predict DBP formation potential using EEM parameters, including parallel factor analysis (PARAFAC) components and the measured DBP concentrations from the batch chlorination experiments. Therefore, we could predict with on average of 13% and 6% precision and error, respectively, the concentration of twenty DBPs produced from chlorination including four THMs, nine HAAs, four HANs, one HNM and two haloketone species. In addition, DOM molecular size distribution was measured on 25 samples by size exclusion – organic carbon – nitrogen detection (LC-OCD-OND) to explore the composition of DOM sources. Our findings highlight potential opportunities for DBP risk reduction through proactive online monitoring of source water using fluorescence EEM spectroscopy. This knowledge will help to organize appropriate mitigation strategies at the catchment level as well as aid in treatment process optimization using fluorescence EEM spectroscopy which surpasses the capabilities of traditional online UV-vis spectroscopy. Overall, the findings of our research will help to decrease the number of total THM exceedances in Ireland and better protect consumer health in relation to drinking water chemical quality around the world.