Latent Predictable Variance Regression for Modeling and Quality Prediction of Dynamic Chemical Engineering Processes

Dynamic latent variable (DLV) methods have received significant attention for quality prediction in chemical engineering processes due to their ability to handle high-dimensional data and complex dynamics. However, traditional DLV approaches focus on latent variable prediction and are not optimized to directly predict the original quality variables. This misalignment frequently results in a reduced prediction accuracy and computational efficiency. To address this limitation, this article proposes a novel latent predictable variance regression (LPVR) algorithm. The core innovation of LPVR lies in its objective function designed to maximize the predictable variance of latent components, which is theoretically proven to be equivalent to minimizing the prediction error of the original quality variables. Under a unified structure, LPVR sequentially extracts low-dimensional components in the order of descending predictable variance, thereby gradually minimizing prediction errors. This enables LPVR to achieve an efficient quality prediction using minimal latent variables. Additionally, the simultaneous prediction of latent variables and original quality variables enhances the model parsimony. Experimental results on a numerical case and an industrial multiphase flow benchmark demonstrate LPVR’s superior performance compared to ten representative methods, showing strong potential for real-time quality prediction in chemical engineering processes.