Response surface methodology (RSM), multi-layer perceptron trained by Levenberg-Marquardt (MLPLM); multi-layer perception and Sigma-Pi neural networks trained by particle swarm optimization (PSO) were used to effectively and reliably predict the performance of Classical-Fenton and Photo-Fenton processes. H2O2 doses, Fe(II) doses, and H2O2/Fe(II) rates were determined as independent variables in batch reactors. The performance of models was compared by using RMSE and MAE error criteria. The performance of models was also evaluated in terms of some properties of regression analysis and scatter that showed high linear relationship between the predictions of SPPSO and the actual removal values. As a distinctive aspect of this study, SPNN trained by PSO was used for the first time in the literature in this area and the best predictive results for almost all cases were generated. Moreover, the genetic algorithm (GA) was applied for SP-PSO model results to determine the optimum values of the study. According to the results of GA, under the optimum conditions Photo-Fenton processes had higher performance in each experiment. Thereby, SP-PSO produced satisfactory prediction results without the need for any additional experiments in the case that experimental designs are difficult or costly for wastewater treatment.