RESEARCH ON COKING PREDICTION MODEL OF RESIDUAL OIL BASED ON DATA AUGMENTATION ANDMULTI-MODEL COMPARISON
2026, 57(7):
89-98.
Asbtract
(
82 )
PDF
References |
Related Articles |
Metrics
To enable rapid assessment of residual oil coking risks, a prediction model for To enable rapid assessment of residual oil coking risks, a prediction model for residual oil coking characteristics was developed by integrating molecular structural parameters, physicochemical properties, data augmentation, and multi-model comparison. First,the molecular structural parameters of residual oil were calculated using the Brown–Ladner method based on elemental composition, carbon residue, SARA components, and hydrogen atom types distribution data. This formed a residual oil coking characteristic prediction dataset, with key feature variables selected via Spearman correlation analysis. Subsequently, training set samples were augmented using generative adversarial network, and the optimal data augmentation factor was determined through Kolmogorov–Smirnov tests. Based on this, back-propagation neural network (BP), Gaussian kernel regression (GKR), and random forest regression (RF) models were established to predict residual oil coking characteristics, with comparative analysis of their performance. The results show that the residual oil feature parameters strongly correlated with coking yields include aromatic carbon ratio, total ring number, density, carbon residue, and resins content. The optimal data augmentation factor was found to be three times. Among the three prediction models, GKR and RF exhibited poorer prediction accuracy and generalization capabilities, while the BP neural network model demonstrated the highest prediction accuracy, with a mean absolute error of 0.1235, root mean square error of 0.1482, and a coefficient of determination of 0.8964. Furthermore, the BP model's cross-validation results exhibited the smallest error, confirming its superior stability and generalization capability.
