Data-Driven Prediction of Intermediate Product Properties in Petroleum Refining

Accurate prediction of intermediate product properties is essential for optimizing petroleum refinery operations. Traditional methods relying solely on crude oil assays fail to capture dynamic crude behavior under actual refinery conditions. This paper presents a data-driven methodology that integrates real-time process data, laboratory measurements, and crude blend composition to enhance property prediction accuracy. Five modeling approaches were evaluated: a reference swing-cut model, a semiempirical model, principal component regression, gradient boosting, and neural networks. The methodology was demonstrated through two industrial case studies predicting kerosene mercaptan content and straight-run naphtha sulfur content in a crude distillation unit processing over 90 different crude oils. The semiempirical model achieved superior performance, reducing mean absolute error by 49% for kerosene mercaptans and 40% for naphtha sulfur compared to the reference model, while maintaining better generalization than complex machine learning approaches. Temporal validation revealed that simpler models with domain-informed constraints exhibited superior robustness to changing operating conditions. The approach provides interpretable coefficients that enable early detection of crude behavior deviations from assay data. These improvements facilitate better operational decision-making and establish a foundation for future reactive planning workflows.