Neural networks are now being used extensively in many areas of control Model predictive control schemes and several other techniques have adapted rapidly to encompass the non-linear predictive capabilities of neural networks and many papers have been written looking at the feasibility of these systems. Much work has also been carried out in the area of neural network learning algorithms and architecture.
This paper deals solely with the pre-treatment of real data produced by a binary methanol/water distillation column to produce a prediction of top composition using variables other than top temperature. The predictor used is a Feed Forward Artificial Neural Network with a standard back error propagation teaming algorithm. The work covered shows that with additional pretreatment of the data the neural network model produced is an improvement over a neural network model produced using the untreated data, over the column operating range. The pretreatment techniques used are easily adaptable to other data sets and systems. Techniques used involve data smoothing, frequency content analysis and a trial and error based computational technique.