Novel prediction of gasoline yield via extremely randomized trees-based algorithm optimized by particle swarm optimization

The fluid catalytic cracking (FCC) process plays a critical role in maximizing gasoline yield; however, its operation is highly complex due to nonlinear process interactions in feedstock composition. In the present study, a data-driven, ensemble learning framework is developed to enhance gasoline yield prediction accuracy. A three-stage feature selection strategy integrating the Variance Inflation Factor (VIF), Recursive Feature Elimination with Cross-Validation (RFECV) and Sensitivity Analysis was employed to identify the most informative variables prior to model development. Particle Swarm Optimization (PSO) was used for hyperparameter optimization of the proposed Extremely Randomized Trees-based model (XRTR). Model performance was evaluated against baseline models, and the proposed VIF–RFECV–PSO–XRTR framework demonstrated superior predictive performance for improved gasoline yield. In single train–test evaluation, the model achieved R 2 = 92.26%, RMSE = 0.3750, and MAE = 0.1008. Cross-validation further confirmed its robustness, with R 2 = 92.20% (±0.55) at k fold = 10.