Estimation of FCC feed composition from routinely measured lab properties through ANN model

Abstract

Realistic kinetic modeling of fluid catalytic cracking (FCC) units requires detailed composition of the feed stream in terms of paraffins, naphthenes and aromatics(PNA) which cannot be analyzed in a field laboratory. This paper presents an artificial neural network (ANN) model to predict detailed composition of FCC feed using routinely measured properties such as density, ASTM distillation temperatures, Conradson carbon residue (CCR) content, sulfur and total nitrogen as inputs to the model. Several feedforward-error back propagation networks with different number of neurons in hidden layers were studied using Levenberg–Marquardt (LM) training algorithm. Among different network architectures investigated, the ANN model with 8 inputs, namely density and ASTM distillation temperatures except IBP, FBP and only one neuron in the output layer to predict paraffin, naphthene and aromatic contents individually showed the best agreement with the experimental results within permissible limit. These compositions when used with a 10-lump kinetic model of FCC unit, successfully simulated plant performance for several different feeds.