Physics-Guided Machine Learning Framework for Catalyst Deactivation Modeling, Coke Formation Mechanisms, and Real-Time Performance Optimization of Continuous Catalytic Reforming (CCR) Units in Petroleum Refineries

Continuous Catalytic Reforming (CCR) units represent one of the most strategically important processes in modern petroleum refining, converting low-octane naphtha feedstocks into high-octane aromatic-rich reformate while co-producing hydrogen. Catalyst deactivation through coke deposition remains the primary limitation on unit performance, requiring careful management of the regeneration cycle and operating conditions. Traditional first-principles kinetic models, while mechanistically rigorous, are computationally prohibitive for real-time optimization and fail to generalize across the wide range of feedstock compositions and operating conditions encountered in commercial refinery practice. This paper presents a comprehensive Physics-Guided Machine Learning (PGML) framework that synergistically integrates fundamental reaction engineering principles with state-of-the-art deep learning architectures for CCR catalyst deactivation modeling and real-time unit optimization. The framework employs Physics-Informed Neural Networks (PINNs) constrained by catalytic reforming reaction stoichiometry, thermodynamic equilibrium relationships, and deactivation kinetics derived from the Voorhies correlation and modified Langmuir-Hinshelwood-Hougen-Watson (LHHW) rate expressions. A novel hybrid architecture combines Long Short-Term Memory (LSTM) networks for temporal deactivation trajectory prediction with Graph Neural Networks (GNNs) encoding the reactor network topology of the four-reactor CCR train. The framework was trained on ۴۸ months of industrial plant data from a ۳۵,۰۰۰ bpd CCR unit, comprising over ۲.۳ million time-stamped process measurements, and validated against an independent ۱۲-month dataset from a second commercial unit with differing feedstock characteristics. The PGML framework achieved a mean absolute percentage error (MAPE) of ۱.۸۷% for coke content prediction on catalyst samples, compared to ۸.۴۲% for a purely data-driven baseline and ۴.۲۱% for a traditional kinetic model. Real-time optimization trials demonstrated a ۳.۲% increase in reformate octane yield, a ۷.۱% reduction in catalyst regeneration frequency, and an estimated annual economic benefit of USD ۴.۷ million per processing unit. The framework demonstrates robust extrapolation behavior under feedstock composition shifts, attributed to the physics-based constraints enforcing thermodynamic consistency of predictions.

کلیدواژه ها:

Physics-Informed Neural Networks ، Continuous Catalytic Reforming ، Catalyst Deactivation ، Coke Formation ، Real-Time Optimization ، Deep Learning ، Process Systems Engineering