Application of a model based predictive control scheme to a distillation column using neural networks

A demonstration of a neural network model based predictive control scheme (MBPC) of a distillation column is described in this paper. The paper identifies significant non-linearities occurring in the dynamics of the distillation column and also demonstrates the failings of a linear model based control scheme under such conditions. Four controllers are compared including two PI controllers, linear MBPC and neural network MBPC. The resultant controllers where tested for disturbance rejection, setpoint response and closed-loop control. In each case the neural network MBPC controller outperformed the other controllers by at least 25% on an integral square error test. The linear MBPC controller had double the standard deviation about setpoint achieved by the neural network. The objective of the control scheme was to control column pressure as tightly as possible but with minimal control action so that other column parameters (e.g. product composition) were not unduly disturbed at the expense of pressure control. The neural network controller outperformed the other controllers on both counts