A trustworthy framework for predicting gasoline octane loss using sparse autoencoder and stacking ensemble learning in petrochemical processes

• SAE compression reduces input redundancy by 91.6% while retaining key information. • SAE+StackingEL achieves R²=0.9578, MSE=0.0021, outperforming RF by ∼10% R². • Explainable AI (PI, PDP) identifies key features and nonlinear RON loss drivers. In response to the growing demand for reliable and interpretable artificial intelligence (AI) in petrochemical and process industries, this study presents an intelligent framework for predicting Research Octane Number (RON) loss during gasoline refining. The proposed approach integrates a Sparse Autoencoder (SAE) for deep feature extraction with a Stacking Ensemble Learning (StackingEL) model for regression, effectively enhancing predictive accuracy and robustness in high-dimensional, noisy industrial datasets. Real-world operational data from a commercial petrochemical refinery were employed to train and evaluate the model. Through SAE-based dimensionality reduction, 367 original process variables were compressed into 31 compact deep features. Using these optimized representations, the SAE+StackingEL model achieved superior performance with an R² of 0.9578, MSE of 0.0021, and MedAE of 0.0280-representing approximately a 10% improvement in R² and a 68% reduction in MSE compared to the best single model (Random Forest: R²=0.8694, MSE=0.0066). Permutation Importance (PI) analysis identified spent adsorbent RON and sulfur content of treated gasoline as the most influential predictors, while Partial Dependence Plots (PDPs) revealed their nonlinear impact on octane preservation. These insights enhance model interpretability and offer actionable guidance for refining operations. The proposed trustworthy AI-driven modeling strategy supports process optimization, reduces emissions, and advances sustainable gasoline production, aligning with Industry 4.0 objectives.