Investigating of artificial neural network potential to predict the properties of refined raw sugar beet juice by electrocoagulation process

Introduction. In this study, based on the high potential of electrocoagulation in the removal of suspended matter and also due to high energy consumption in traditional treatment, the potential of EC was modeled by artificial neural network on purification of raw beet syrup. Material and Methods. The potential of neural network in prediction of turbidity, color and purity of raw beet juice was investigated with different parameters as voltage (5, 10 and 15 volts) pH (6, 7 and 8) and time (regular time intervals from 1 to 60 min) during electrocoagulation process. ANN modeling was carried out by Neurosolution software v6 to determine the best type of transport function, learning rule, and determination of applied percentages for training, validation and testing stages based on their mean square errors, mean square normalized errors, mean absolute errors and correlation coefficients. Results and discussion. The best neural network with maximum correlation coefficient for turbidity and purity obtained in Levenberg learning law and tangent transfer function which included 8 and 17 neurons respectively. Also, the best correlation coefficient and the less mean square error for color modeling related to a network with one hidden layer and 9 neurons that learned under levenberge learning law and sigmoid transfer function. Modeling was carried out with different percentages of data for training, validation and testing that the best prediction correlation for turbidity and purity obtained when 55% of the data were used for training, 40% of them were employed for validation and 5% of the data were used for testing, whereas the best percentage of learning, validation and testing for color prediction were 60, 30 and 10, respectively. The predicted values of models had suitable correlation with experimental data, so that correlation coefficient with experimental data of turbidity, color and purity were 0.999, 0.997 and 0.990, respectively. This study also addressed the model sensitivity to input data. The most model sensitivity of the model for prediction of turbidity, color and purity was related to voltage. Conclusion. The model was able to predict the turbidity, color and purity of the syrup under various operating conditions, as the modeling data showed a high correlation with the experimental data.