A Microcontroller-Based Adaptive Model Predictive Control Platform for Process Control Applications

Michael Short,Fathi Abugchem

doi:10.3390/electronics6040088

Abstract

Model predictive control (MPC) schemes employ dynamic models of a process within a receding horizon framework to optimize the behavior of a process. Although MPC has many benefits, a significant drawback is the large computational burden, especially in adaptive and constrained situations. In this paper, a computationally efficient self-tuning/adaptive MPC scheme for a simple industrial process plant with rate and amplitude constraints on the plant input is developed. The scheme has been optimized for real-time implementation on small, low-cost embedded processors. It employs a short (2-step) control horizon with an adjustable prediction horizon, automatically tunes the move suppression (regularization) parameter to achieve well-conditioned control, and presents a new technique for generating the reference trajectory that is robust to changes in the process time delay and in the presence of any inverse response. In addition, the need for a full quadratic programming procedure to handle input constraints is avoided by employing a quasi-analytical solution that optimally fathoms the constraints. Preliminary hardware-in-the-loop (HIL) test results indicate that the resulting scheme performs well and has low implementation overhead.

Highlights

The potential benefits that adaptive and predictive control schemes can bring to industrial control applications have been well documented in recent years [1,2,3,4,5]
The need for a full quadratic programming procedure to handle input constraints is avoided by employing a quasi-analytical solution that optimally fathoms the constraints
Most modern Model predictive control (MPC) schemes utilize a quadratic (2-norm) objective function, with the optimization solved on-line by quadratic programming (QP) software; much of the original work on MPC algorithms utilized linear (1-norm) objective functions which were solved on-line by linear programming (LP) software [5,6]. This transition was principally due to a number of well-documented drawbacks to the use of LP in early formulations in MPC, including possible idle/deadbeat dichotomous behaviors, the need to use iterative schemes to obtain solutions even in the unconstrained case, and potentially poor scaling in the size of the MPC problem to the corresponding

Summary

Introduction

The potential benefits that adaptive and predictive control schemes can bring to industrial control applications have been well documented in recent years [1,2,3,4,5]. Most modern MPC schemes utilize a quadratic (2-norm) objective function, with the optimization solved on-line by quadratic programming (QP) software; much of the original work on MPC algorithms utilized linear (1-norm) objective functions which were solved on-line by linear programming (LP) software [5,6]. This transition was principally due to a number of well-documented drawbacks to the use of LP in early formulations in MPC, including possible idle/deadbeat dichotomous behaviors, the need to use iterative schemes to obtain solutions even in the unconstrained case, and potentially poor scaling in the size of the MPC problem to the corresponding. The principal components of a typical adaptive MPC implementation are as shown in Figure 1 and consist of an online parameter estimator, adaption mechanism, a predictor of the process future behavior given the (known) current state, and an optimizer (which may have input, output, or state constraints to navigate through)

Methods

Results

Conclusion