Development of a Machine Learning Framework for Performance Optimization in Venezuelan Petrochemical Plants

Abstract

Background: The Venezuelan petrochemical industry is facing a sustained decline in operations, with olefin plants operating at less than 30% of their nominal capacity due to the deterioration of critical infrastructure, chronic shortages of spare parts, and the absence of optimization models adapted to real operating conditions. Objective: To develop a data science-based methodological framework to revitalize operational management by integrating predictive models and optimization algorithms in piston flow reactors (PFRs), with minimal historical data requirements (≥1,500 observations). Methods: The approach uses Gradient Boosting to model the nonlinear relationship between temperature, pressure, and residence time with ethylene and propylene yield. The predictive model is coupled to a genetic algorithm using an objective function that maximizes combined olefin production by penalizing specific energy consumption (λ = 0.035). Results: The framework identifies an optimum point (845.3 °C, 0.28 MPa, 0.32 s) that increases combined production by 8.42% (40.26 to 43.65% by weight) and reduces energy consumption by 9.84% (31.5 to 28.4 GJ/ton), with a maximum deviation of 2.16% compared to rigorous simulation in Aspen HYSYS. The predictive models achieved R² = 0.942 for ethylene and R² = 0.913 for propylene. Conclusions: The approach demonstrates that reduced data volumes are sufficient to train robust predictive models, offering a viable route for data-driven management in resource-constrained contexts without significant investments in additional hardware