“ML-DRIVEN ANALYSIS AND OPTIMIZATION OF CATALYST CHANGEOUT SCHEDULES USING REAL-TIME PROCESS DATA”

Abstract

The Fluid Catalytic Cracking Unit (FCCU) is a vital component in modern oil refineries, responsible for converting heavy hydrocarbons into high-value products such as gasoline, diesel, and LPG. The catalyst—typically a zeolite-based material (Zeolite Y)—undergoes severe thermal and chemical stress, leading to coke deposition that gradually deactivates it. Excessive coke accumulation (> 40 wt %) reduces cracking efficiency, while premature catalyst replacement incurs significant operating costs. Conventional monitoring through laboratory sampling every 6–8 hours provide only intermittent information and is prone to human error, making real-time catalyst management difficult. To restore activity, the spent catalyst is regenerated by burning off coke deposits and recirculated to the reactor. The continuous reactor–regenerator loop in the fluid catalytic cracking process of an FCCU facilitates the flow of catalyst between the cracking and regeneration zones. Maintaining coke saturation within the optimal 33–39 wt % range is essential for yield stability and process efficiency. This paper proposes a machine-learning-driven framework for continuous catalyst health monitoring and optimization of changeout schedules using real-time process data. The framework integrates three analytical layers: (1) a soft-sensing model that estimates coke saturation from refinery sensor inputs using PCA-assisted Random Forest and XGBoost regressors; (2) a lifetime forecasting model employing Gaussian Process Regression and LSTM networks to predict the remaining useful life (RUL) of the catalyst with uncertainty quantification; and (3) a prescriptive optimization layer that minimizes operating expenditure (OPEX) through mixed-integer programming and reinforcement learning. Applied to historical FCCU datasets, the proposed system achieved a predictive accuracy of R² ≈ 0.78 and reduced catalyst consumption by nearly 25% while lowering normalized OPEX by 16%. The principal contribution of this work lies in demonstrating a unified, deployable FCCU catalyst management framework that transforms manual, heuristic-based decisions into an automated, data-driven, and economically optimized control process.