Coagulation modeling using artificial neural networks to predict both turbidity and DOM-PARAFAC component removal

Abstract In this study, four different neural network models were evaluated for predicting both turbidity and dissolved organic matter (DOM) removal during the coagulation process at the Akron Water Treatment Plant (Akron, Ohio, USA). DOM was monitored and characterized using fluorescence spectroscopy and parallel factor (PARAFAC) analysis, building upon previous research which identified three unique fluorescence components (C1, C2, and C3). Neural network models were built using operational data to predict each of the fluorescence components and turbidity after coagulation based on variable raw water quality and chemical doses. Correlation coefficients between measured and model predicted values for the final turbidity, C1, C2, and C3 models on an unseen test data set were 0.91, 0.95, 0.97, and 0.51, respectively. The predictive capability of the top performing model for each parameter was evaluated using parametric analysis, external validation criteria, and the absolute relative error distribution. Results suggest that the models for settled turbidity and the three settled component scores are valid and can be used to predict the removal of individual fractions of DOM (as measured by PARAFAC components) as a function of chemical dose and raw water quality, providing the water plant the ability to simultaneously manage two key water quality treatment objectives.