The combination of multiple linear regression and adaptive neuro-fuzzy inference system can accurately predict trihalomethane levels in tap water with fewer water quality parameters

Abstract

Artificial Neural Network (ANN) models are accurate in predicting the levels of disinfection by-products (DBPs) in drinking water. However, these models are not yet practical due to the large number of parameters involved, which should take a significant amount of time and cost to detect. Developing accurate and reliable prediction models of DBPs with fewest parameters is essential in the management of drinking water safety. This study used the adaptive neuro-fuzzy inference system (ANFIS) and radial basis function artificial neural network (RBF-ANN) to predict the levels of trihalomethanes (THMs), the most abundant DBPs in drinking water. Two water quality parameters identified by multiple linear regression (MLR) models were used as model inputs, and the quality of the models was assessed based on criteria such as correlation coefficient (r), mean absolute relative error (MARE), and the percentage of predictions with absolute relative error less than 25% (NE<25%) and over than 40% (NE>40%), etc. The results showed that the ANFIS models had higher correlation coefficients (r = 0.853–0.898) and prediction accuracy (NE<25% = 91%–94%) compared to RBF-ANN models (r = 0.553–0.819; NE<25% = 77%–86%) and traditional MLR models (r = 0.389–0.619; NE<25% = 67%–77%). Conversely, the prediction error, as indicated by MARE and NE>40%, showed the opposite trend: ANFIS models (MARE = 8%–11%; NE>40% = 0–5%) < RBF-ANN models (MARE = 15%–18%; NE>40% = 5%–11%) < MLR models (MARE = 19%–21%; NE>40% = 11%–17%). The present study provided a novel approach for constructing high-quality prediction models of THMs in water supply systems using only two parameters. This method holds promise as a viable alternative for monitoring THMs concentrations in tap water, thereby contributing to the improvement of water quality management strategies.