Prediction of polymer quality in batch polymerisation reactors using robust neural networks

Abstract A technique for predicting polymer quality in batch polymerisation reactors using robust neural networks is proposed in this paper. Robust neural networks are used to learn the relationship between batch recipes and the trajectories of polymer quality variables in batch polymerisation reactors. The robust neural networks are obtained by stacking multiple nonperfect neural networks which are developed based on the bootstrap re-samples of the original training data. Neural network generalisation capability can be improved by combining several neural networks and neural network prediction confidence bounds can also be calculated based on the bootstrap technique. A main factor affecting prediction accuracy is reactive impurities which commonly exist in industrial polymerisation reactors. The amount of reactive impurities is estimated on-line during the initial stage of polymerisation using another neural network. From the estimated amount of reactive impurities, the effective batch initial condition can be worked out. Accurate predictions of polymer quality variables can then be obtained from the effective batch initial conditions. The technique can be used to design optimal batch recipes and to monitor polymerisation processes. The proposed techniques are applied to the simulation studies of a batch methylmethacrylate polymerisation reactor.