Estimation of impurity and fouling in batch polymerisation reactors using stacked neural networks

A robust method for the estimation of reactive impurities and reactor fouling during the early stage of batch polymerisation using stacked neural networks is reported. Data for building neural network models are resampled using the bootstrap re-sampling technique to form several sets of training data. For each set of training data, a neural network model is developed. Predictions from individual networks are combined to form the final model prediction in order to improve model accuracy and robustness. A further benefit of bootstrap aggregated neural network is that confidence bounds for model predictions can be formulated. Stacked neural networks are used to build an inverse model of the reactor. The amounts of impurities and fouling can be worked out by comparing the predicted effective initial reaction conditions with the nominal initial conditions. The proposed techniques have been successfully applied to a pilot scale batch methyl methacrylate polymerisation reactor