Control of batch end-product quality can be realized in the latent variable space using the corresponding partial least squares (PLS) model. The PLS model used for the controller was usually derived from process trajectories covering the complete batch duration. In order to predict the batch end-product quality at a selected control decision point for a new batch, missing data algorithms were needed to estimate the future trajectories of system measurements. However, errors incurred by this estimation may have a negative effect on the prediction and the following control moves, especially at the beginning of the batch when the majority of data are missing. This paper extends the original approach by using evolving window PLS models instead, where evolving window PLS models only employ system measurements up to the current time and thus the estimation of future trajectories can be avoided. The original and revised control methods implemented in the latent variable space are further compared with their counterpart model predictive control (MPC) methods implemented in the time domain through an application to a simulated fed-batch fermentation process.
|Title of host publication||IET Seminar Digest|
|Subtitle of host publication||UKACC International Conference on CONTROL 2010|
|Pages||1166 – 1171|
|Publication status||Published - 2010|