Neurocomputing approach for real time optimisation modelling of an industrial process

Concerns neural modelling, first of a methanol-water flash system, then of a crude oil distillation column. The models were developed with different combinations of variables. Radial basis function (RBF) net models were tested and taken as the base case. Hierarchically structured net (HSNN) models and simple serial and hybrid net-model configurations were also developed. The RBF nets modelled the methanol-water system well. Different combinations of output variables affect the predictions. Grouping suitable output variables combinations in a model gave better predictions. The most difficult variable to predict was the methanol composition in the vapour outlet stream, y. More complex models were required for better prediction of this variable; results of the simple serial and hybrid ANN show a significant improvement. A hybrid of ANN and first principles model gave the best prediction. Although there was an increase in the total number of nodes, the pitfall of the model is avoided because the nets are separated. Performance of the linear-nonlinear HSNN was highly dependent on using the suitable slave input. The results show potential for further investigation of more complex network models for highly nonlinear variables. For crude oil distillation, standard RBF was able to provide a highly satisfactory model Proper grouping of related variables not only improved predictions, but also allow the complex, multivariable model to be more manageable. Since standard RBF gave sufficiently accurate predictions, developing more complex models was deemed to be unnecessary.