Hybrid Modelling of a Polyethylene Process to Predict Polymer End-Use Properties

• Main Author: Errington, Adam
• Published: University of Newcastle upon Tyne 2008
• Subjects: 547
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.489273

This thesis presents a novel method for predicting polymer end-use properties using information from a molecular weight distribution (MWD) predicted using a hybrid modelling strategy. A mechanistic model of a dual reactor, polyethylene process has been developed to predict process information required to calculate the MWD. Process data available includes MWDs, reactor inputs and end-use properties of the polymer. The reaction kinetics fitted to the process are commercially sensitive and unavailable for modelling purposes, therefore, kinetics representative of a typical polyethylene process have been used within the model. The accuracy of predictions made by the mechanistic model is insufficient to meet the requirements of the industry. However, this thesis demonstrates that a hybrid modelling strategy, combining the mechanistic model with a non-linear empirical layer to adjust key descriptors of the molecular weight distribution, improves the prediction accuracy. It is also shown that a nonlinear empirical approach can be used to predict important polymer end-use properties from key sections of the MWD identified as having a strong influence on those properties. This contribution combines these two findings to make predictions of end-use properties directly from the hybrid modelled MWD rather than the MWD measured off-line using gel permeation chromatography. This enables predictions of both MWD and end-use properties to be made on-line and potentially incorporated into a model predictive control strategy. The results show that small differences between the hybrid MWD and the actual MWD prohibit consistent prediction of end-use properties, with only the best observations making accurate predictions. However, the predictions of end-use properties are of an accuracy comparable to black box models of the process. The incorporation of process knowledge and understanding within the mechanistic layer of the hybrid model also adds to the credibility of the predictions.