Faculty of Electrical Engineering Universiti Teknologi Malaysia VOL. 11, NO. 1, 2009, 8-14 ELEKTRIKA http://fke.utm.my/elektrika 8 A Heuristic Approach for Tuning Model Predictive Controller Tri Chandra S. Wibowo, Nordin Saad * and Mohd Noh Karsiti Department of Electrical and Electronics Engineering, Universiti Teknologi PETRONAS, 31750 Tronoh, Perak, Malaysia. * Corresponding author: nordiss@petronas.com.my, Tel: 605-3687835, Fax: 605-3657443 Abstract: This paper presents the results of a heuristic approach for tuning an embedded model predictive control (MPC) in real-time conditions. A gaseous pilot plant, a kind of interacting series processes, is used as the controlled process. Process modelling in this MPC uses gray-box approach, where two multi-input single-output (MISO) models are combined into a multi-input multi-output (MIMO) model. A real-time embedded control system consisting the off-the-shelf components and software tools is implemented and analysed experimentally. Four tuning parameters are investigated to illustrate the intuitively understandable MPC tuning rules. The aim here is to deliver an intuitive MPC tuning implementation that may found useful in practical applications. Keywords: Gaseous pilot plant, model predictive control (MPC), real-time tuning. 1. INTRODUCTION Model predictive control (MPC) is one of the most popular multivariable controls implemented in process industries. MPC promises several advantages, among others, its algorithm can deal with multivariable case, provides a systematic way to the treatments of constraints and taking into account of actuator limitations, and introduces feedforward control in a natural way to compensate the measurable disturbances. However, the computational algorithm of MPC is somehow more complex than the classical proportional-integral- derivative (PID) control. The controller output of MPC is based on the prediction of process output generated by the internal model, which commonly a kind of linear time-invariant (LTI) models, even though it is well known that mostly the real systems of interest are nonlinear, time-varying, may contain delays, and some variables or signals of central importance may not able to be measured directly. Moreover, MPC is a kind of discrete control systems, which produces the control signals in every sampling time. These conditions may presence some challenges in the practical implementation of MPC. When it goes to real-time implementation, the hardware embedded MPC may necessary to be used. Even though some applications are still prefer to use a dedicated personal computer (PC) for implementing MPC, however, it seems to be less appropriate due to the possibility of user intervention. In the practical process control, the control algorithms are commonly embedded in the control modules which are separated from the data recording and monitoring purposes. Hence, the embedded MPC is expected to function without user intervention, although it may require user interaction. MPC has a set of tuning parameters, which can be used to fine-tune the closed-loop response for good performance and stability. Al-Ghazzawi et al. [1] presented an on-line tuning strategy for linear MPC algorithms based on the linear approximation between the closed-loop predicted output and the MPC tuning parameters. Their work revealed that basically these parameters are adjusted via a trial and error procedure, which is a cumbersome task due to their overlapping effect and due to non-linearity brought by input constraints. This procedure can be more systematic since there are several general tuning guidelines available in the literature. Shridhar and Cooper [2] developed a novel tuning strategy for multivariable MPC, however, their approach is limited to unconstrained MPC and also requires representing the process by a first-order plus dead-time (FOPDT) model, which may not work well for higher order and/or unstable process. Trierweiler and Farina [3] presented a novel tuning strategy for multi- input multi-output (MIMO) MPC. The proposed MPC tuning procedure is based on robust performance number (RPN), which can measure the difficulties of control problem. Wojsznis et al. [4] proposed a practical tuning approach, which features simple calculation of controller design parameters. Yet this calculation of tuning parameters seems to be suitable only for certain conditions. The experimental formula to define a penalty on move factor requires the presence of dead time in MPC scans, otherwise the calculation will results the same value of penalty on move factor for all kind of process. Notably, the tuning strategies mentioned in [1-4] are validated with simulations, which can be anticipated to produce different results when it goes to real-time applications. This paper presents a practical tuning approach of MPC in the real-time control of a gaseous pilot plant. This study begins with the theory of model predictive control algorithms. Next, is the construction of a rapid prototype of an embedded MPC controller and a real-time implementation of the MPC. Following this, fine-tuning of the closed-loop responses of MPC is carried out in