Data predictive control of nonlinear process feature dynamics through latent variable behaviours

Abstract

The paper presents a new big data-based approach to control of feature dynamics of continuous nonlinear chemical/industrial processes, based on the behavioural systems theory and deep learning tools. From time-series process data, the feature dynamics of a nonlinear process are extracted using an Autoencoder (AE), a type of artificial neural network. The feature dynamics are embedded in, and can be constructed from, a linear dynamic behaviour of latent variables. The latent variable dynamic space is described by a kernel representation and linearly maps the feature variable space. A Data Predictive Control (DPC) approach is developed to optimise the feature variables by controlling the latent variable dynamics using a system behaviour framework. Behaviour-based dissipativity conditions are adopted to deal with errors that arise in the latent and feature variable spaces during neural network training. A case study is presented to illustrate the proposed approach.