Inverse internal model control of an ethylene polymerisation reactor using artificial neural networks.

Abstract

An artificial neural network is a mathematical black-box modelling tool. This tool can be used to model complex non-linear multivariable processes. In attempting to create an inverse process model of an industrial linear low density polyethylene reactor, several interesting results were encountered. Both time-invariant algebraic and time-invariant dynamic models could adequately represent the process, provided an identified 50-minute time lag was taken into account. A novel variation of the traditional IMC controller was implemented which used two inverse neural network process models. This was named Inverse Internal Model Control (IIMC). This controller was initially tested on a real multivariable pump-tank system and showed promising results. The IIMC controller was adapted to an on-line version for the polymer plant control system. The controller was run in open loop mode to compare the predictions of the controller with the actual PID ratio controllers. It was hoped that by incorporating neural network models into the controller, they would take the non-linearity and coupling of the variables into account, which the present PID controllers are unable to do. The existing PID controllers operate on separate loops involving the two main feeds (co-monomer and hydrogen) to the reactor, which constitute aspects of the control system in which the scope for advanced control exists. Although the control loop was not closed, the groundwork has been laid to implement a novel controller that could the operation of the plant.