Radial basis function artificial neural network (RBF ANN) as well as the hybrid method of RBF ANN and grey relational analysis able to well predict trihalomethanes levels in tap water

[Display omitted] •Linear/log-linear model was not effective in predicting THMs levels in tap water.•RBF ANN shows high performance in predicting THMs levels in tap water.•GRA can facilitate generating sound RBF ANN models with the fewer factors.•GRA and RBF ANN paved a new way to predict DBPs in actual distribution system. Many models have been developed in previous studies for predicting the formation of disinfection by-products (DBPs) in drinking water. However, most of them were linear or log-linear regression models, and generated based on simulated disinfection of source water or treated water in a laboratory other than real tap water, which shows low application potential in practice. In this study, a radial basis function artificial neural network (RBF ANN) as well as the hybrid method of RBF ANN and grey relational analysis (GRA) was proposed to predict trihalomethanes (THMs) levels in real distribution systems. A total of 64 sets of data including THMs levels (trichloromethane (TCM), bromodichloromethane (BDCM) and total-THMs (T-THMs)) and 8 water quality parameters (temperature, pH, UV absorbance at 254 (UVA254), dissolved organic carbon, bromide, residual free chlorine, nitrite and ammonia) were used to train and verify the proposed model. As compared to linear and log-linear regression models (rp = 0.254–0.659; N25 = 46–78%), RBF ANNs for THMs (TCM, BDCM and T-THMs) prediction consistently show higher regression coefficients (rp = 0.760–0.925) and prediction accuracy (N25 = 92–98%), which indicates the high capability of RBF ANN to learn the complex non-linear relationships involved THMs formation. Further analysis shows that RBF ANNs using fewer water quality parameters based on GRA still make excellent performance in THMs prediction (rp = 0.760–0.946; N25 = 92–98%). This result demonstrates that GRA can be an effective technique to facilitate the generation of sound RBF ANN models with fewer factors.