Neural-Network-Assisted FCC Kinetic Modeling for Enhanced Parameter Estimation Using the CREC Riser Simulator

Abstract

This study presents an integrated framework for developing kinetic models of vacuum gas oil (VGO) catalytic cracking under sparse experimental data conditions. Experiments were performed in the CREC Riser Simulator at different C/O (catalyst/VGO) weight ratios. The CREC Riser Simulator is a bench-scale mini-fluidized reactor capable of reproducing the short contact times and operating conditions of industrial fluid catalytic cracking (FCC) risers. Product distributions were characterized by using a five-lump scheme consisting of unconverted VGO, light cycle oil (LCO), gasoline, light gases, and coke. To address the limitations associated with sparse datasets, a feedforward neural network (FNN) was used to reconstruct continuous reaction trajectories from discrete experimental measurements. These synthetic trajectories enabled the estimation of kinetic parameters for a phenomenological five-lump reaction network that incorporates catalyst deactivation. The resulting kinetic model established was subsequently implemented in a 1D heterogeneous model of a large-scale industrial FCC riser, providing reliable predictions of VGO conversion, product selectivity, and axial temperature profiles.