MLP, ANFIS, and GRNN based real-time coagulant dosage determination and accuracy comparison using full-scale data of a water treatment plant

Abstract

Real-time determination of appropriate coagulant dosage under wide fluctuation of raw water quality in a water treatment plant (WTP) is a challenging task due to nonlinearity relation between coagulant dosage and raw water characteristics. In this research, three techniques, multilayer perceptron (MLP), adaptive neuro fuzzy inference system (ANFIS), and generalized regression neural network (GRNN), are applied to determine the coagulant dosage at Bansong drinking WTP. Each model is developed based on 8,760 historical data sets with hourly resolution for a whole year. Several statistical properties are determined to obtain the best-fit model from each method. The top performing models of each method are evaluated by external validation indices and absolute relative error according to nine turbidity zones. From the result, MLP and ANFIS models meet all conditions of validation indices, but GRNN cannot. The MLP shows the best result for high turbidity zones over 20 NTU as well as for overall performance. Meanwhile, ANFIS provides consistent results and better performance than MLP for low turbidity zones which have higher disorder of coagulant dosage data. The GRNN shows high accuracy for the highest turbidity zone which occurs during the rainy season. It is concluded that MLP, ANFIS, and GRNN can support operators effectively for real-time determination of coagulant dosage.