Adaptive system identification of industrial ethylene splitter: A comparison of subspace identification and artificial neural networks

Abstract

The manuscript considers the problem of data-driven modeling of an ethylene splitter (from an industrial plant). The process presently operates with end composition controllers that does not work well during process transition. The objective of the present work is to investigate the use of different data-driven techniques such as subspace identification and neural network-based methods for the purpose of developing a dynamic data-driven model. To this end, first an ethylene splitter simulation model is built that replicates industrial operation. The ability of the simulation model to capture the key traits of the process dynamics are first established by comparing it with data from the plant operation. The simulation model is subsequently utilized to work as a test bed for future control purposes and to serve as an additional test of the modeling approaches. An online model adaptation scheme is developed to improve the model's prediction capabilities under new operation patterns.