Nonlinear and direction-dependent dynamic process modelling using neural networks

The paper discusses several methods of modelling complex nonlinear dynamics using neural networks. Particular reference is made to the problem of modelling direction-dependent relationships. A typical example of this would be top product composition control in a distillation column, where it is easier (i.e. faster) to make the product less pure than it is to make it more pure by an equivalent amount. Recurrent neural networks are identified as a potential method of modelling this type of relationship. The particular architecture chosen for this example is referred to as `semirecurrent´, since only past values of the predictions of the network are fed back to the input layer. This architecture is successfully used to model direction-dependent relationships in both simulated and actual industrial process data