A multivariate Bayesian network analysis of water quality factors influencing trihalomethanes formation in drinking water distribution systems

Abstract

Controlling disinfection by-products formation while ensuring effective drinking water disinfection is important for protecting public health. However, understanding and predicting disinfection by-product formation under a variety of conditions in drinking water distribution systems remains challenging as disinfection by-product formation is a multifactorial phenomenon. This study aimed to assess the application of Bayesian Network models to predict the concentration of trihalomethanes, the dominant halogenated disinfection by-product class, using various water quality parameters. Naïve Bayesian and semi-naïve Bayesian models were constructed from Sydney and South East Queensland datasets across 15 drinking water distribution systems in Australia. The targeted variable, total trihalomethanes concentration, was discretised into 3 bins (<0.1 mg L−1, 0.1 – 0.2 mg L−1 and >0.2 mg L−1). The Bayesian network structures were built using water quality parameters including concentrations of individual and total trihalomethanes, disinfectant species (free chlorine, monochloramine, dichloramine, total chlorine), nitrogen species (free ammonia, total ammonia, nitrate, nitrite), and other physical/chemical parameters (temperature, pH, dissolved organic carbon, total dissolved solids, conductivity and turbidity). Seven performance parameters, including predictive accuracy and the rates of true and false positive and negative results, were used to assess the accuracy and precision of the Bayesian network models. After evaluating the model performance, the optimum models were selected to be Bayesian network augmented naïve models. These were observed to have the highest predictive accuracies for Sydney (78%) and South East Queensland (94%). Although disinfectant residuals are among the key variables that lead to trihalomethanes formation, potential concentrations of trihalomethanes in distribution systems can be more confidently predicted, in terms of probability associated with a wider range of water quality variables, using Bayesian networks. The modelling procedure developed in this work can now be applied to develop system-specific Bayesian network models for trihalomethanes prediction in other drinking water distribution systems.