Development, analysis and comparison of connectionist models for real-time optimisation

Abstract

This research investigates the development of different types of artificial neural networks (ANN) and grey-box ANN models for real time optimisation (RTO) of chemical processes. Therefore, must be easily computed, stable, and easily updated and maintained. Due to this, the models have a predominantly connectionist structure. All the models were developed and simulated under MATLAB environment.

Initial investigations were focused on a methanol-water flash system, which is simple, yet realistic in representing the non-linearity of multivariable chemical processes. Following this, ANN models were then developed for a crude oil distillation column, which is a more complex industrial process. Training and testing data for the network models were generated using steady state process models simulated in the Aspen Plus steady state process simulator. The sensitivity analysis feature in Aspen Plus was utilised to generate a large amount of data in a single simulation run.

Three standard ANN models were developed for the M-W flash system: multi-layer perceptrons (MLP) using backpropogation training with variable learning rate, MLP using Levenberg-Marquadt, and radial basis function networks (RBFN). Of the three standard ANN models, the RBFN was found to give the best result, and was thus selected as the base case for comparison with other models.

The RBFN models were able to model the M-W flash system well, except for y, the composition of methanol in the vapour outlet stream. Different combinations of output variables affect the predictions of the model. In general, grouping suitable output variables combinations in a network model gave better predictions.

More complex models were required for better prediction of y because of the discontinuity in y that exists in the change between the single-phase and the two-phase region. Three groups of models were developed: hierarchically structured neural network (HSNN), serial network models and hybrid ANN-first principles models (FPM). The models in all the three groups managed to improve the prediction of the base model.

Among all the models, the output-tuned HSNN model that was designed in series with a serial RBFN model provided the best prediction of y. Moreover, the structure of the output-tuned HSNN provided means for incorporating constraints into the network. The constraints used are simple and readily known information. The completely driven HSNN, was also able to perform as well, when there is an input variable that directly and strongly influenced y. In both types of HSNN, prior knowledge embedded into the network structure allowed for significant improvements over the base case.

The results of the hybrid ANN-FPM also showed a significant improvement over the base case in predicting y. Performance of the hybrid model, however, depended on the availability of accurate parameters needed by the FPM. Therefore, a hybrid ANN-FPM would require extensive prior information.

For the crude oil distillation tower, a standard RBFN was able to provide a highly satisfactory model. Proper grouping of related variables not only improved predictions, but also allowed for the complex, multivariable model to be more manageable while avoiding the "curse of dimensionality". The RBFN model for one of the crude tower sections was also able to give good predictions when tested for range and dimensional extrapolation. Since standard RBFN gave sufficiently accurate predictions, developing more complex models was deemed to be unnecessary.

This study showed that ANN and grey box ANN models have the potential to model chemical processes for RTO. Various methods exist to take advantage of readily available information that can be utilised to overcome modelling difficulties. Most significantly, since these models can be easily developed and updated, they are suitable for practical industrial applications.