Artificial neural networks developed for prediction of dye decolorization efficiency with UV/K2S2O8 process

The present work has focused on modeling the decolorization of C.I. Direct Red 16 (DR16) using UV/K2S2O8 process. The experiments were conducted in an air bubbled batch photo-reactor with three liters capacity, and equipped with a UV-C light source of only 6 W, placed horizontally in its center. An artificial neural network (ANN) model was developed to model and predict the behavior of the process. Five important operational parameters and decolorization efficiency were introduced as the inputs and outputs of the network. Six three-layered ANNs with different component functions and also one four-layered ANN were constructed. A developed three-layered ANN provides the best results, using BFGS quasi-Newton backpropagation learning algorithm (trainbfg), “tansig” and “purelin” as transfer functions in hidden and output layers. Also, application of 9 neurons in the hidden layer and 400 iterations for the network calibration caused to satisfy network training while overfitting was prevented. For the four-layered ANN the 5-4 mode of the 9 neurons repartition is the best among all the possibilities. The K-fold cross-validation method, employed for the performance evaluation, showed high correlation coefficient (0.9968) and low mean square error (2.56 × 10−4) for testing data. Sensitivity analysis indicated order of operational parameters relative importance on the reaction as: time > [DR16]0 > [S2O8]0 > pH ≈ temperature.