An efficient neural network model for aiding the coagulation process of water treatment plants

Abstract

In a water treatment plant, the decision to carry out a jar test, for determining the required coagulant dosage, is made based on notable changes in treated water qualities such as treated water turbidity and color, which is essentially a reactive response to changes in water qualities. In addition, until a change that the operator deems as ‘significant’ occurs, the plant tends to use the same dosage determined previously using the jar test for an elongated period of time. In this study, artificial neural network (ANN) models were developed to proactively decide what coagulant dosages to use based on changes in raw water parameters. Use of ANNs also prevents the regular use of costly chemicals used for jar tests and enables responding to sudden changes in water qualities. The general regression neural network (GRNN) and extreme learning machine neural networks require minimal computational effort for model development as they involve minimal model parameters and their training algorithms are not iterative. The current study determines the more convenient and efficient model of the two for aiding the coagulation process. The GRNN and ELM-RBF models predicted test data with R values of 0.9737 and 0.9783, respectively. It was noted that the GRNN was prone to overfitting and the ELM-RBF model demonstrated higher generalization ability than the GRNN. Therefore, it was concluded that the ELM-RBF model was the more suitable model for the prediction of the coagulant dosage.