Abstract: Aiming at the large loss problem of research octane number (RON) of refined gasoline from S Zorb unit, based on the historical data from a S Zorb unit in the past 3 years, the maximal information coefficient (MIC) and Pearson correlation coefficient were used in combination with BP neural network, 22 modeling variables were selected from 273 variables including feed oil properties, adsorbent properties, product properties and operation variables, and a RON prediction model with the structure of 21-14-1 was established and validated. The results show that the prediction model has good goodness of fit and generalization ability, and its average absolute error and determination coefficient (R²) of the model on the test set are 0.116 3 and 0.960 1, respectively. On this basis, genetic algorithm (GA) was used to optimize the best operation variables to improve the product RON. It was found that the model could effectively reduce RON loss of gasoline through optimization, and the RON loss of gasoline could be reduced by 25% through the optimization model.

Key words: catalytic cracking gasoline, S Zorb process, BP neural network, genetic algorithm, octane loss