Introduction

Contents

A typical industrial plant can have controllers ranging from proportional-integral-derivative controllers (PI/PID) to advanced model predictive controllers (MPC), such as dynamic matrix control (DMC) [1, 2], quadratic dynamic matrix control (QDMC) [3, 4], robust multivariable predictive control technology (RMPCT) (see review in [5, 6]), generalized predictive control (GPC) [7, 8], etc. With the goals of optimal performance, energy conservation and cost effectiveness of process operations in industry, the design of optimal controllers and controller performance assessment have received great attention in both industry and academia. Typically a ‘model’ or some sort of mathematical representation of the process and the control objective are required not only for designing suitable controllers but also for analyzing controller performance. For predictive controllers, which use a model of the system to make predictions, model identification forms the critical part of controller design. Identification aims at finding a mathematical model from the measurement record of inputs and outputs of a system [9, 10, 11]. Parametric model identification, such as of a transfer function or a state space model, involves obtaining reduced-order models of a pre-specified structure for a system that could be of a very high order and complexity. Nonparametric modeling approaches, such as impulse/step response modeling and frequency-domain based modeling, can also be found in the literature for controller design and performance analysis. Interestingly, nonparametric-model based controller design and analysis tools have been used in industry quite successfully. As a convention, identification of parametric or nonparametric models for the process is typically used as a first step in MPC design, and the models are then converted to prediction matrices as the second step before the calculation of control actions is performed. Data-driven approaches to obtain the prediction matrices used in the controller design directly from process data, and avoid the intermediate explicit model identification procedure, has become an area of active research in recent years [12, 13, 14, 15, 16, 17, 18].