eng Science and Education Publishing Journal of Automation and Control 2015-01-13 3 1 10 17 10.12691/automation-3-1-2 AUTOMATION2015312 article Masih Vafaee Ayouri masih.vafaee@gmail.com 1 Mehdi Shahbazian 1 Bahman Moslemi 2 Mahboobeh Taheri 3 Department of Instrumentation and Automation Engineering, Petroleum University of Technology, Ahwaz, Iran Department of Basic Science, Petroleum University of Technology, Ahwaz, Iran Senior expert in R&D, Sarkhoon & Qeshm Gas treating Company, Bandar Abbas, Iran Radial Basis Function Neural Network (RBFNN) is considered as a good applicant for the prediction problems due to it's fast convergence speed and rapid capacity of learning, therefore, has been applied successfully to nonlinear system identification. The traditional RBF networks have two primary problems. The ?rst one is that the network performance is very likely to be affected by noise and outliers. The second problem is about the determination of the parameters of hidden nodes. In this paper, a novel method for robust nonlinear system identification is constructed to overcome the problems of traditional RBFNNs. This method based on using Support Vector Regression (SVR) approach as a robust procedure for determining the initial structure of RBF Neural Network. Using Genetic Algorithm (GA) for training SVR and select the best parameters as an initialization of RBFNNs. In the training stage an Annealing Robust Learning Algorithm (ARLA) has been used for make the networks robust against noise and outliers. The next step is the implementation of the proposed method on the Hydrocarbon Debutanizer unit for prediction of n-butane (C4) content. The performance of the proposed method (ARLA-RBFNNs) has been compared with the conventional RBF Neural Network approach. The simulation results show the superiority of ARLA-RBFNNs for process identification with uncertainty. http://pubs.sciepub.com/automation/3/1/2/automation-3-1-2.pdf robust system identification RBF Neural Networks hydrocarbon debutanizer unit support vector regression genetic algorithm annealing robust learning algorithm