Search

Due to the complex nature of processing units and the variability in raw oil properties, traditional first-principle methods are not always effective in predicting the output of these units without adapting them. Data-driven models have become popular alternatives because they are more adaptable to the variability in processing units, especially when large amounts of data are available. Feature selection is a crucial aspect of data-driven modeling, especially for processing units. It involves identifying and selecting the most important variables influencing the unit's output. The selected variables are then used as inputs to the model, which helps to reduce noise and improve model accuracy. Accordingly, this article presents a hybrid method that combines feature selection with data-driven modeling to create an accurate model for predicting the output of a Residue Fluid Catalytic Cracking (RFCC) process unit. The Metaheuristic Ant Colony Optimizer (MACO) algorithm was used to extract the best subset of features from the collected data, which were then used with the Artificial Immune Genetic Algorithm with Local Vaccination (AIGA-LV) method to select the optimal features. Finally, the Random Forest (RF) method was used to predict the products of the RFCC process unit. The data were collected from a refinery plant in central Iran for nine months. The proposed model was efficient in feature selection and flexible in handling different scenarios and products. It improved the accuracy to 99.17% at its best and improved the primary model by 3–20%. Therefore, this method could be useful for predicting the output of various industrial processes, especially the RFCC and FCC process units. [ABSTRACT FROM AUTHOR]