Building inferential prediction models of batch processes using subspace identification

In this paper, a general method for converting available batch plant data into useful prediction models is proposed. A subspace identification method, which was proposed for identification of continuous systems, is used to develop a batch-to-batch correlation model from plant data. In this context, the state of the model is a holder of relevant information contained in the past batch data for predicting the behavior of current and future batches. The modeling framework naturally allows the user to capture correlations among the variables within each individual batch as well as those of successive batches, as reflected in the modeling data, and take advantage of them in prediction and control. It will be shown that the batch-to-batch model can be converted into a time transition model that can be used to predict the future behavior of the relevant variables, including the end-quality variables, in real time based on incoming measurements. Various practical issues will be addressed, such as the reduction of state dimension and incorporation of delayed laboratory measurements of quality variables.