Search

• A general input parameter selection method is proposed, combined with an optimized wavelet de-noising ANFIS model to predict wastewater quality. • A network structure selection method is given to enhance the stability and scalability of ANFIS. • Statistical stratified sampling is employed to proportionally partition the dataset. • The experimental results confirm that the proposed model outperforms other existing AI models. The monitoring of wastewater quality is vitally important for the stability of an ecosystem. Among many machine learning techniques proposed for predicting the quality of wastewater, the adaptive neuro-fuzzy inference system (ANFIS) can achieve the best accuracy. However, due to data size limitations, the uneven distribution of randomly sampled training data can cause out-of-range prediction errors in the ANFIS model. In this study, a general-purpose input parameter selection method is proposed, combined with an optimized wavelet de-noising ANFIS model, to predict salinity parameters in wastewater discharge samples from the Las Vegas Wash, Nevada, USA. A statistical, stratified sampling strategy is used to preprocess the wastewater quality dataset. Compared with existing artificial intelligence models, the experimental results prove that the proposed model has the best performance, in which the R 2 testing value achieves 0.976, 0.975, 0.988, and 0.986 in predicting chloride, sulfate, electrical conductivity, and total dissolved solids, respectively. Image, graphical abstract [ABSTRACT FROM AUTHOR]