Low-Dimensional Latent State Space Identification with Application to the Shell Control Process

Multivariable system identifications with exogenous inputs are widely encountered in real-time industrial process control and system identification experiments. This paper presents a low-dimensional latent state space identification algorithm that handles collinear variables and builds latent dynamic state space models with input-output data. The proposed method aims to project the original variables to a lower-dimensional space for optimal predictability while preserving the causal dynamics of process systems. The optimization of this algorithm is achieved through iterative procedures, where the dimension reduction and latent state space identification are performed in a unified manner. The effectiveness of the proposed algorithm is illustrated in the oil fractionator in a shell control process, compared to the traditional subspace identification method.