Abstract: Catalytic cracking is a key process for converting heavy oil to lighter fractions in petroleum refining, and establishing a product prediction model for a catalytic cracking unit facilitates process optimization and the development of smart refineries. In response to the intelligent construction requirements of a domestic refinery,a gasoline yield prediction model for a catalytic cracking unit based on an optimized BP neural network was proposed. Through data cleaning and maximum information coefficient analysis, 12 input variables highly correlated with gasoline yield were selected from the original 30 variables, achieving a 60% dimensionality reduction. On this basis, 6 intelligent optimization algorithms were employed to optimize the initial weights and thresholds of a BP neural network with a 12-8-1 structure, and the predictive performance of the model was compared under different optimization algorithms. The results show that the BP neural network optimized by the dwarf mongoose optimization algorithm (DMOA-BP) yields the best prediction accuracy, with its mean absolute error, mean square error, and mean absolute percentage error all significantly lower than those of the other algorithms. Moreover, the average coefficient of determination (R²) over fourfold cross-validation reached 0.9889. Therefore, DMOA-BP was selected as the gasoline yield prediction model for the catalytic cracking unit, providing a high-accuracy, low-complexity prediction tool for intelligent production in refineries and offering valuable guidance for the optimized operation of catalytic cracking units.

Key words: catalytic cracking, correlation analysis, BP neural network, dwarf mongoose optimization algorithm, nonlinearity, data preparation