Prediction model for increasing propylene from FCC gasoline secondary reactions based on Levenberg-Marquardt algorithm coupled with support vector machines

Levenberg-Marquardt (LM) algorithm was adopted to optimize the multiple parameters of the support vector machines (SVM) model to overcome the difficulty in selecting the parameters of SVM and to fit relational expression of high nonlinearity. Strategy of dividing the training data into working data to train SVM and the testing data so as to avoid over-fitting was performed. Comparison of the proposed LM/SVM method with three reported hybridized SVM approaches (GA/SVM, SM/SVM and SQP/SVM) was also carried out. The new method was applied in modelling for the prediction of propylene by secondary reactions of FCC gasoline. Best performance of LM/SVM employing polynomial kernel was demonstrated. Good agreement between predicted results and experimental data suggests that the LM/SVM method is successfully developed and the SVM model for increasing propylene is well established. Finally, sequential quadratic programming (SQP) algorithm was employed to optimize the operation conditions of FCC gasoline secondary reaction for maximizing the propylene yield. The obtained optimization conditions are consistent with experimental data and reported results, indicating that the optimization results are reliable.