Improved inferential feedback control through combining multiple PCR models

A Principal Component Regression (PCR) based model is developed from process operation data for a distillation column. The top and bottom product compositions are estimated from all tray temperature measurements. The estimated product compositions are directly used in a feedback control loop. In a PCR model, the number of principal components retained is usually determined through cross validation. When different subsets of the data are used in training and testing, the resulting models may have different numbers of principal components leading to different performance on unseen data. To improve the PCR model estimation performance on unseen data, multiple PCR models are developed from the bootstrap re-samples of the original process data. The developed PCR models are then combined together. It is shown that the PCR models obtained in this way provide better estimation performance than single PCR models. When this improved PCR software sensor is used in a feedback control loop of a distillation column, the resulting control performance is much better than that from a single PCR model software sensor.