Computational modeling of a Fluid Catalytic Cracking Unit

This paper presents an in-depth comparative analysis focused on product yield forecasting for a Fluid Catalytic Cracking process (FCC) using different variations of the Long Short-Term Memory Neural Network (LSTM). We introduce an innovative Multi-Headed LSTM (MH-LSTM) that addresses challenges arising from the different time-scale dynamics of FCC subunits—specifically, the fast reactor riser and the slowe regenerator. In our approach, these subsystems are modeled independently in separate LSTM heads each used to capture the unique temporal features of its respective process. These independently learned representations are then integrated into a unified network, enabling more accurate multi-step, multivariate forecasts of FCC product yields. Results based on Mean Squared Error (MSE) and R² score indicate that the proposed MH-LSTM model not only outperforms other LSTM-based models in product yield forecasting for FCC units but also maintains robust performance across different Signal-to-Noise Ratio levels. However, this improvement comes at the expense of an increased training time.