Design of neural networks using genetic algorithm for on-line property estimation of crude fractionator products

The products from the crude distillation unit (CDU) of a petroleum refinery need to conform to certain standards and hence it is important that the properties, which characterize these products, be measured on-line so that necessary control action can be taken through feedback mechanism. In view of nonavailability of on-line hardware sensors, software-based sensors have been developed for prediction of product properties online. Artificial neural network (ANN) based models have been used for this purpose. To overcome the problem of designing the networks by a cumbersome trial and error procedure, a methodology based on genetic algorithm (GA) has been developed. This involves optimally designing the network architecture including number of hidden layers, number of neurons in each layer, connectivity and activation functions. For on-line prediction of any property, the network design can be completed in a matter of a few hours. Additionally GA was also used for the selection of most relevant set of input features to the neural network. The above methodology was used for designing networks to predict properties of various side-draw products from a crude fractionator. The properties include ASTM temperatures, specific gravities and Flash Points of various products. All the predicted results were generally in good agreement with lab measurements with average deviation ranging from 0.5 to 3% for the properties investigated. The methodology is quite general and can be used to design similar other nets. (c) 2006 Elsevier Ltd. All rights reserved.