Improved predictive capability of coagulation process by extreme learning machine with radial basis function

Abstract

In an effort to improve the predictive capability of artificial neural networks for the coagulation process in a water treatment plant, extreme learning machine (ELM) coupled with radial basis function (RBF) neural networks were employed. The ELM-RBF was selected to exploit the computational robustness of ELM and accuracy of RBF for sufficiently large number of data that were available from the plant. The coagulation data were divided into two categories based on low and high turbidity. The optimum number of input parameters for modeling the coagulation of low turbidity water was found to be 3, whereas the optimum number of input parameters for modeling the coagulation of high turbidity water was found to be 4. Re-selection of the number of input parameters was necessary considering that raw water alkalinity was a significant factor in improving the high turbidity model performance. The low turbidity model was capable of predicting the coagulant dosage with correlation coefficient exceeding 0.97. The high turbidity model was capable of predicting the coagulation dosage with reasonably acceptable correlation coefficient of at least 0.80.