Comparative study of RSM and ANN for multiple target optimisation in coagulation/precipitation process of contaminated waters: mechanism and theory. Issue 19 (30th December 2022)

• Record Type: Journal Article
• Authors: Golbaz, Somayeh; Nabizadeh, Ramin; Rafiee, Mohammad; Yousefi, Mahmood
• Is Part Of: International journal of environmental analytical chemistry. Volume 102:Issue 19(2022)
• Page Start: 8519
• Page End: 8537
• Publication Date: 2022-12-30
• Subjects: Ferric chloride -- response surface methodology -- artificial neural networks -- humic acid -- engineered Kaolin; Ecology -- Periodicals; Chemistry, Analytic -- Periodicals; Environmental chemistry -- Periodicals; Pollution -- Periodicals; 577.14
• DOI: 10.1080/03067319.2020.1849663
• Languages: English
• ISSNs: 0306-7319

Abstract

ABSTRACT: Response surface methodology (RSM) and artificial neural network (ANN) were applied to investigate the chemical coagulation/precipitation process in the treatment of water contaminated with humic acid and processed kaolin. Accordingly, the removal efficiencies of total organic carbon (TOC), turbidity, and colour were considered to assess the effectiveness of the process. The proposed models were evaluated based on determination coefficient ( R 2 ) and Mean-Square Error (MSE). Although both models could satisfactorily interpret the correlation of targets with model inputs, i.e. pH and coagulant dosage, the ANN exhibited a slightly better fit to the process data than the RSM, based on comparison between determination coefficients ( R 2 ) and Mean-Square Error (MSE). R 2, MSE, and error values of 0.990–0.999, 0.112–3.569 and 0.000–1.660% for ANN were respectively demonstrated. To investigate the predictive performance of both models, some additional experiments were subsequently carried out at obtained optimum conditions by genetic algorithm technique. TOC, colour, and turbidity removal efficiencies of 58.8, 93.0, and 100% were demonstrated at identified optimum conditions (i.e. pH 6.7 and 3 mM of ferric chloride). The results revealed a low deviation from their predicted values with maximum errors of 1.90 for RSM and 1.66 for ANN. Also, our results suggest that setting the RSM before ANN could considerably improve its prediction weaknesses.