Prediction of treated water turbidity and its parametric behavior for the coagulation process of a water treatment plant using a joint extreme learning machine-genetic algorithm

Abstract

A new genetic algorithm was developed to improve the performance of the extreme learning machine radial basis function (ELM-RBF) neural network model to predict treated water (TW) turbidity for the coagulation process of water treatment. The genetic algorithm was constructed such that a notable improvement in the model could be achieved, within a short period of time. Two sets of models were developed based on high and low turbidity values. The genetic algorithm improved the correlation coefficient (R) and the mean squared error (MSE) of the low turbidity model from 0.76 and 2.16 × 10-4 to 0.81 and 4.85 × 10-5 respectively, in 15 minutes; whilst improving R and MSE of the high turbidity model from 0.62 and 0.0011 to 0.93 to 2.89 × 10-4 respectively, in 2 minutes. The ability of the model to capture the reported variation of TW turbidity with gradually increasing coagulant dosage was also tested. It was noted that ELM-RBF was more capable of capturing such physical and chemical phenomena better than multilayer perceptron and ELM-single layer feed-forward models. The genetic algorithm improved the ability of the model to capture the parametric behavior of TW turbidity too.