Physics-Informed Machine Learning for Multiphase Hydrodynamics, Catalyst Circulation Malperformance, and Real-Time Fault Detection in Residue Fluid Catalytic Cracking (RFCC) Processes

Residue fluid catalytic cracking (RFCC) is among the most complex and economically critical conversion units in modern petroleum refining, processing heavy vacuum residue feeds under severe hydrodynamic and thermal conditions. The simultaneous presence of gas–solid multiphase flow, non-ideal catalyst circulation behavior, and thermally coupled reaction networks creates a process environment that remains inadequately described by purely data-driven or purely first-principles approaches alone. This work presents an integrated physics-informed machine learning (PIML) framework that embeds governing conservation laws, two-fluid hydrodynamic correlations, and catalyst transport phenomenology directly into neural network architectures, enabling high-fidelity process representation with limited labeled plant data. Three interconnected modules are developed: (i) a physics-informed neural network (PINN) for riser multiphase hydrodynamics, trained on a hybrid loss function combining data residuals with momentum and mass conservation constraints; (ii) a dynamic graph neural network capturing catalyst circulation malperformance modes including standpipe flooding, slide valve sticking, and stripper inefficiency; and (iii) a real-time fault detection and classification system leveraging variational autoencoder (VAE) embeddings conditioned on physics-derived latent states. The framework is validated against a commercial RFCC unit operating on Arabian Heavy residue feed, encompassing over ۱۸ months of distributed control system (DCS) historian data and ۴۷ documented fault events. Results demonstrate fault detection F۱-scores exceeding ۰.۹۴ across all malperformance categories, mean absolute percentage error (MAPE) below ۲.۱% for riser temperature profiles, and a ۷۶% reduction in false alarm rate compared to conventional multivariate statistical process control (MSPC) baselines. The proposed approach advances the state-of-the-art toward physics-consistent digital twins for refinery-scale catalytic cracking operations.

کلیدواژه ها:

Physics-informed neural networks ، Residue fluid catalytic cracking ، Multiphase hydrodynamics ، Catalyst circulation ، Fault detection ، Digital twin ، Process monitoring