Predicting coagulation-flocculation process for turbidity removal from water using graphene oxide: a comparative...

Three artificial intelligence (AI) data-driven techniques, including artificial neural network (ANN), support vector regression (SVR), and adaptive neuro-fuzzy inference system (ANFIS), were applied for modeling and predicting turbidity removal from water using graphene oxide (GO). Based on partial mutual information (PIM) algorithm, pH, GO dosage, and initial turbidity were selected as the input variables for developing the models. The prediction performance of the AI-based models was compared with each other and with the response surface methodology (RSM) model, previously reported by the authors, as well. The models' estimation accuracy was assessed through statistical measures, including mean-squared error (MSE), root-mean-square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). Among the evaluated models, ANN had the highest estimation accuracy as it showed the highest R² for the validation data (0.949) and the lowest MSE, RMSE, and MAE values. Furthermore, ANN predicted 76.1% of data points with relative errors (RE) less than 10%. In contrast, the weakest prediction performance belonged to the SVR model with the lowest R² for both calibration (0.712) and validation (0.864) data. Besides, only 57.1% of the SVR's predictions were characterized by RE < 10%. The ANFIS and RSM models exhibited a more or less similar performance in terms of R² for the validation data (0.877 and 0.871, respectively) and other statistical parameters. According to the results, the ANN technique is proposed as the best option for modeling the process. Nevertheless, as the RSM technique provides valuable information about the contribution of the independent operational parameters and their complex interaction effects using the least number of experiments, simulating the process by this technique before modeling by ANN is inevitable.