Batch-Wise Self-Corrected Gaussian Process Regression for Quality Prediction in Batch Processes

Data-driven modeling has been widely studied and employed for quality prediction in batch processes. However, existing research and development of this methodology have found it not sufficiently effective in addressing challenges from the batch-to-batch uncertainty and limitation of noisy modeling data. This work proposes a quality prediction method that combines batch correlation information and dynamic data reconciliation (DDR). The method is based on Gaussian process regression and integrated with canonical correlation analysis to capture batch-to-batch characteristics through examining the correlation between batches. Furthermore, DDR is used to dynamically adjust prediction variance, reduce the impact of noisy process data, and improve the prediction accuracy. A numerical case, a batch crystallization process, and a silica modification process have verified that the proposed method is demonstrated effective performance, compared with the candidates.