Performance Prediction Modeling of Andesite Processing Wastewater Physicochemical Treatment Via Artificial Neural Network

Abstract

A holistic approach has been introduced on the treatment of andesite marble processing wastewater. The treatment was implemented by using coagulants, flocculants, and minerals. Dosage, mixing time, settling time (ST), mixing speed (MS), and pH were optimized. Up to 99% turbidity removal was achieved by using Alum, FeCl3, PACl, and 40% cationic flocculant under optimized conditions of each. Among the minerals, pumice resulted in 81% turbidity removal efficiency in a much shorter duration as compared with the natural settling conditions. The obtained experimental data was used in the model established with Artificial Neural Network (ANN)-MLP for the prediction of treatment performance (final NTU). Feature selection was performed by using statistical correlations between each experimental parameter and final NTU. ST and MS were determined as the non-correlated parameters with final NTU. The model was integrated into feature selection approach upon establishing three scenarios with datasets: NFS was the whole experimental data consisting of all variables; FS1 was the dataset without ST and dataset FS2 was built up by removing both ST and MS. It was indicated that, when using chemical coagulants, there is no need to study the effects of ST to improve treatment performance, whereas, when using flocculants, both ST and MS has no influence on the treatment performance, so there is no need to perform extra experiments for these variables. Duration of training processes of modeling for all datasets was changing in range of 45-60 s. ANN configuration with two hidden layers was the best model structure, and ST was the parameter which had minimum influence on the treatment performance. The treatment and modeling approaches suggested in this study will be useful to give practical answers to the facility.