A novel modified training of radial basis network: Prediction of conversion and selectivity in 1-hexene dimerization process

Abstract In this study, a radial basis function neural network (RBF-NN) has been designed and constructed based on coupling the orthogonal least square (OLS) method, genetic algorithm (GA) and the gradient descent algorithm. By investigation of the effective parameters of basis function spreads on the performance of the RBF-NN, the basis function spreads and centers of the proposed RBF-NN have been obtained using OLS and GA simultaneously. The proposed technique has been applied to model 1-hexene dimerization process to evaluate its performance. The effects of temperature, pressure, Weight Hourly Space Velocity (WHSV) and Time On Stream (TOS) have been investigated on 1-hexene conversion and the dimer selectivity of 1-hexene dimerization process in the proposed scheme. The results obtained from the proposed scheme have been compared with those of other three methods; namely back propagation neural network, RBF-NN based on MATLAB neural network Toolbox (newrb) and an optimized RBF-NN model designed by Bataineh et al. The comparison of the results shows that the proposed, novel scheme outperforms the other alternative models with average Root Mean Square Error (RMSE) equal to 1.06 and the correlation coefficient (R 2 ) of greater than 0.95 for output parameters.