Optimization and Prediction of Petrochemical Process Performance by Machine Learning: A Case Study on Catalytic Cracking
Year: 2025
Type: article
Abstract: Background: The fluid catalytic cracking (FCC) process is critical in modern refineries, as it converts heavy petroleum fractions into high-value products such as gasoline and olefins. However, its highly nonlinear behavior and sensitivity to multiple operating variables make traditional optimization methods inefficient and inflexible. There is a critical gap in the literature on integrated machine learning (ML) frameworks that combine supervised and unsupervised techniques for holistic FCC performance prediction and operational analysis. Objective: This study seeks to develop and validate a comprehensive ML-based framework for predicting FCC performance and uncovering hidden operational patterns, enabling real-time decision making and process optimization. Methods: Historical operational data (12 500 records, 15 variables, including temperature, pressure and catalyst-oil ratio) were used to train and evaluate supervised models (Random Forest, Neural Networks, SVM) for regression (gasoline yield) and classification (operational stability), along with unsupervised techniques (K-means clustering, PCA) for pattern discovery and dimensionality reduction. Model performance was evaluated by R², RMSE, F1-score, AUC-ROC and 5-step cross-validation. Results: The MLP Regressor model achieved the highest predictive accuracy for gasoline yield (R² = 0.8771), while Random Forest was the best classifier for operational stability (F1-Score = 0.9815). Unsupervised analysis revealed limitations: the PCA required 16 principal components to retain 95% of the variance, and K-means clustering (with k=4) obtained an average silhouette score of 0.0440, indicating an absence of significant natural clustering in the data. Conclusions: The integrated machine learning framework demonstrates great potential for real-time FCC optimization, cost reduction and energy efficiency improvement in petrochemical operations, thanks to the high accuracy of supervised models and interpretability techniques (SHAP, Sobol) that identify critical variables such as regenerator temperature and catalyst/oil ratio.
Authors: Kelvin Cabrera, Wakefield Cabrera
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4415522909