Multimode residual monitoring of particle concentration in flue gas from Fluid Catalytic Cracking regenerator

Abstract

The reactor-regenerator system Fluid Catalytic Cracking (FCC) highly relies on stable catalyst cycling. However, the FCC unit usually operates in multiple modes to accommodate feedstock properties, catalyst characteristics, and variations in operating parameters. Multiple operating modes exhibit different flow and separation performances within the unit, resulting in different statistical characteristics of the particle size distribution and the amount of catalyst loss under normal conditions. It brings significant challenges for process monitoring methods to accurately predict catalyst loss for multiple modes and to enhance distinguishability between transition and anomaly modes due to complex non-stationary characteristics. In this work, a Multimode Fusion Residual monitoring model using the Long Short-Term Memory encoder–decoder (MFR-LSTM) is proposed to analyze the catalyst loss in a 2.8 million t/a FCC unit. The MFR-LSTM integrates a multimode prediction module and an adaptive attention module to extract mode-specific characteristics and temporal dependencies across modes to approximate the first principles of physical/chemical relationships among different features under normal catalyst loss status. In addition, it captures the evolution from normal to fault conditions by monitoring residual variations with an adaptive threshold. Results show that the MFR-LSTM model outperforms the single model in both the stable and transition modes, with an improvement in the Root Mean Square Error (RMSE) of approximately 20% and 15%, respectively, demonstrating acceptable stability performance. Furthermore, the generalization performance of the model is confirmed with multiple particle size distribution in different modes.