Abstract

Artificial neural networks (ANN) were proposed as a multivariate experimental design tool for monitoring a photo-Fenton treatment of wastewaters containing a synthetic mixture of pesticides. ANN and Nelder-Mead simplex methods were used to find out the optimum operating parameters of a photo-Fenton pilot plant. ANN was developed to predict the most important operating parameters (e.g., the total organic carbon and the initial mineralization kinetic rate constants of the reactions), which determine the photo-catalytic degradation efficiency in photo-Fenton processes. Experimental measurements of temperature, pH, hydrogen peroxide (H2O2) consumption, initial concentration of Fe2+, and the AE were used as input data for the ANN learning. A feed-forward with one hidden layer, a Levenberg–Marquardt learning algorithm, a hyperbolic tangent sigmoidal transfer function and a linear transfer function were used to develop the ANN model. The best fitting of the training database was obtained with an ANN architecture constituted by seven neurons in the hidden layer. The simulated results were validated with experimental measurements, showing an acceptable agreement (R2 > 0.99). The ANN was subsequently coupled with a Nelder–Mead simplex method to obtain the optimum operating parameters of the photo-Fenton pilot plant. The H2O2 consumption was used as key variable for evaluating the optimization procedure. Errors less than 1% between simulated and experimental data were found. The obtained results showed that the use of ANN provides an excellent predictive performance tool with the additional capability to assess the influence of each operating parameter on the removal process of water pollutants.