Data-driven process decomposition and robust online distributed modelling for large-scale processes

• Published in: International journal of systems science Vol. 49; no. 3; pp. 449 - 463
• Main Authors: Shu, Zhang, Lijuan, Li, Lijuan, Yao, Shipin, Yang, Tao, Zou
• Format: Journal Article
• Language: English
• Published: London Taylor & Francis 17.02.2018; Taylor & Francis Ltd
• Subjects: affinity propagation clustering; Algorithms; block-wise RPLS, model reduction, model-predictive control; Canonical correlation analysis; Clustering; Correlation analysis; Decomposition; Mathematical models; Modelling; On-line systems; parameter identification; process control; Process variables; Scale (ratio)
• ISSN: 0020-7721; 1464-5319
• DOI: 10.1080/00207721.2017.1406551

With the increasing attention of networked control, system decomposition and distributed models show significant importance in the implementation of model-based control strategy. In this paper, a data-driven system decomposition and online distributed subsystem modelling algorithm was proposed for large-scale chemical processes. The key controlled variables are first partitioned by affinity propagation clustering algorithm into several clusters. Each cluster can be regarded as a subsystem. Then the inputs of each subsystem are selected by offline canonical correlation analysis between all process variables and its controlled variables. Process decomposition is then realised after the screening of input and output variables. When the system decomposition is finished, the online subsystem modelling can be carried out by recursively block-wise renewing the samples. The proposed algorithm was applied in the Tennessee Eastman process and the validity was verified.