Novel insights into halogenated carbazoles (HCZs) prediction in tap water: a comparative study of grey relational analysis-based neural networks

Abstract

In recent years, halogenated carbazoles (HCZs) have emerged as a novel class of disinfection byproducts (DBPs) detected in tap water, posing potential dioxin-like toxicity risk to human health. Therefore, it is particularly important to detect HCZs in tap water, yet this process is time-consuming and labor-intensive. To address this challenge, developing predictive models for HCZs based on water quality parameters presents an attractive alternative. Unfortunately, no prediction studies for HCZs have been reported to date. This study investigated the feasibility of using linear models, log-linear models, backpropagation neural networks (BPNN), general regression neural networks (GRNN), and radial basis function neural networks (RBFNN) to predict the occurrence of HCZs, including 3-chlorocarbazole (3-CCZ), 3-bromocarbazole (3-BCZ) and total halogenated carbazoles (ΣHCZs) in tap water. The input parameters for BPNN, GRNN, and RBFNN were selected based on grey relational analysis (GRA) results, while the linear and log-linear models were developed using stepwise regression. The results showed that linear and log-linear models were not suitable (N₂₅=0.30-0.64, r_p=0.40-0.71), and BPNN demonstrated limited prediction performance (N₂₅=0.52-0.78, r_p=0.77-0.83). GRNN excelled only in predicting ΣHCZs (N₂₅=0.98, r_p=0.96). Particularly, RBFNN exhibited good performance in predicting 3-CCZ, 3-BCZ, and ΣHCZs (N₂₅=0.70-0.89, r_p=0.89-0.94), providing valuable insights and possibilities for real-time monitoring of HCZs in tap water.