Identification and predictive control of a FCC unit using a MIMO neural model

The main aim of this work is to implement and evaluate the performance of a neural network-based model predictive control (MPC) applied to a fluid catalytic cracking (FCC) unit. The studies were carried out by dynamic simulation of a Kellogg Orthoflow F converter. The output signals were modified by random noise. From steady-state conditions, a sequence of step changes was imposed on the usual manipulated variables. Information on the process dynamics and interactions among variables is supplied by recording the responses of controlled variables. During the network training procedure, this information was readily captured by the neural model. The neural model output is composed of the four controlled variables, predicted one step ahead. Tests with unseen data showed relative errors of the output variables around 1%. This reliable neural model was then introduced into an MPC scheme, subject to process constraints. Two regulatory and a servo-regulatory problems were simulated. Both the predictions from the neural model and the optimal control calculations could be calculated rapidly, since the control horizon equals 1 or 2. Overall, simulation experiments have confirmed good regulatory and tracking properties of the proposed control system. Simulation tests with noisy measurements provide confidence that the neural model and the controller could be used in an industrial environment.