In this thesis a mathematical model based on artificial neural network for catalytic reformer unit (I) of Isfahan oil refinery was made. This model is able to projection the unit outlet factors included production rate and associated octane number, and also LPG productivity and gas/feed ratio, based on 15 inlet factors included feed rate and other feed's characteristics, catalyst life time, weight average bed temperature and etc. Number of 819 industrial data were used for education, validation and examination of different neural networks, and ultimately based on maximum correlation coefficient and minimum residual error, a cascade four-layer network, consisting 15 neurons in first layer, 10 neurons in 2nd layer, 7 neurons in 3rd layer and 4 neurons in last layer was selected. The neurons' transfer functions in all layers were tan-sigmoid, and Levenberg-Marquardt algorithm was used for artificial neural network education. Overall correlation coefficient for real data and selected neural network outputs is equal to 0.96942, and maximum residual error for prediction of process performance in one working period is equal to 1.2912 unit. Therefore an objective function for maximizing the difference between economical values added of unit normal condition and optimum condition was defined. Using the model and genetic algorithm, the amount of chlorine and water in recycle gas, WABT, and feed rate were optimized. The average difference between economical value added of unit normal condition and optimum condition was 6.7972 10 4 $, which means using the results, the performance of unit can be improved.
