Data-driven intelligent modeling framework for the steam cracking process

[Display omitted] Steam cracking is the dominant technology for producing light olefins, which are believed to be the foundation of the chemical industry. Predictive models of the cracking process can boost production efficiency and profit margin. Rapid advancements in machine learning research have recently enabled data-driven solutions to usher in a new era of process modeling. Meanwhile, its practical application to steam cracking is still hindered by the trade-off between prediction accuracy and computational speed. This research presents a framework for data-driven intelligent modeling of the steam cracking process. Industrial data preparation and feature engineering techniques provide computational-ready datasets for the framework, and feedstock similarities are exploited using k-means clustering. We propose LArge-Residuals-Deletion Multivariate Adaptive Regression Spline (LARD-MARS), a modeling approach that explicitly generates output formulas and eliminates potentially outlying instances. The framework is validated further by the presentation of clustering results, the explanation of variable importance, and the testing and comparison of model performance.