Abstract: To address the challenge that conventional analytical methods struggle to accurately determine the hydrocarbon composition of straight-run gas oil, thereby hindering detailed mechanistic analysis of its hydrocarbon reactions, a BP neural network model was developed to predict the hydrocarbon composition of waxy oil.This was based on 145 sets of collected straight-run gas oil hydrocarbon composition data and routine analytical data from industrial units. Prior to modeling, outlier removal was performed using the interquartile range (IQR) method, and feature variables were screened using Pearson correlation coefficient and maximal information coefficient (MIC) methods. The model was solved using the Levenberg-Marquardt (L-M) algorithm.Validation results indicated that, except for the prediction models of paraffin content and monocyclic aromatic content, the fitting accuracy between predicted and actual values for other BP neural network models was generally poor.To enhance prediction accuracy and model generalizability, improvements were made by optimizing the model algorithm and refining outlier detection methods. The results demonstrated that replacing the L-M algorithm with a Bayesian regularization optimization algorithm and eliminating outliers using a discriminant function method significantly improved the predictive performance of the modified BP neural network model for paraffins, naphthenes, and aromatics. The coefficient of determination (R2) between predicted and actual values increased by an average of 50.91%, while the mean absolute percentage error (MAPE), mean squared error (MSE), and mean absolute error (MAE) decreased by an average of 38.23%, 62.01%, and 36.23%, respectively.The improved model meets the accuracy requirements for predicting the hydrocarbon composition of straight-run gas oil feedstock in industrial units and is suitable for practical industrial applications.

Key words: hydrocarbon composition, BP neural network, straight-run gas oil, simulated prediction