Development of extreme learning machine radial basis function neural network models to predict residual aluminum for water treatment plants

Abstract

Two sets of input parameters were employed to develop Extreme Learning Machine Radial Basis Function (ELM-RBF) models predicting residual aluminum, in order to facilitate parametric analysis of reported physical and chemical phenomena relating to the effect of alum dosage, raw water (RW) turbidity and RW color on residual aluminum concentration. RW turbidity was identified as the dominant variable affecting the distribution of the multivariate data, condensed into two principal components using principal component analysis. Thus two sets of models were developed based on the RW turbidity value: low turbidity models and high turbidity models. The performance of all models was satisfactory, with test correlation coefficients exceeding 0.85. The shapes of the plots of the parametric analysis were satisfactory and were in line with reported phenomena. However, the numerical accuracy of the plots obtained by the parametric analysis was poor. It was noted that using data with a wider range of values for the dominant variable (RW turbidity) helped improve the parametric plots.