Comparative Study of Different Supervisory Control Structures Mohammad Aminul Islam Khan 1 and Syed Ahmad Imtiaz 2 and Faisal Khan 3 and M.A.A Shoukat Choudhury 4 Abstract— With the advent of computer control, supervisory controllers such as simple cascade control, model predictive con- trol (MPC), dynamic matrix control (DMC), etc are increasingly being used in process industries. In this study, performance of three such controllers namely simple cascade controller, ‘MPC cascaded to PID’ and ‘PID free MPC’ are compared on a continuous stirred tank heater (CSTH) system. In the MPC cascaded structure the flow-loops are regulated by the PID controller. On top of that a DMC manipulates the set-points of the flow-loops to control tank temperature and level. The ‘PID-free MPC’ structure uses a DMC to manipulate the valve positions directly. The study reveals that the PID-free MPC structure outperforms the cascade structure in both disturbance rejection and set-point tracking. However, the PID-free MPC structure demands more control action and has more control load. Integrated square error (ISE) is used to quantify the performance. I. I NTRODUCTION Model predictive controllers are typically used as a super- visory layer above the base level PID controller, especially in large-scale applications. This structure gained acceptance mainly because it allows the implementation of MPC with minimal changes to the existing control structure. Also, the PID layer can act as a fall back when the MPC is turned off for any reason. However, this structure does not allow the potential benefits of the MPC to be fully harnessed. In practice, it was observed that there are many incentives in breaking the PID loop and directly manipulating the valve output using the MPC. One common example is when trying to use the full valve capacity (e.g., maximize feed, maximizing cooling) it is common practice to break the PID loop and manipulate the valve directly from MPC. Recently, a software called MaxAPC from the original inventors of DMC is being marketed that uses the DMC to directly manipulate the actuator [1]. It is claimed that this controller performs better than the MPC cascaded to PID structure. Therefore, an objective investigation of the perfor- mance of these competing control structures is necessary. In this study, a simulation-based comparative study is carried *This work was not supported by any organization 1 Mohammad Khan is a graduate student of Department of Engineering and Applied Science, Memorial University of Newfoundland, St. John’s NL, Canada. maik66@mun.ca 2 Syed Ahmad Imtiaz is faculty member of Department of Engineering and Applied Science, Memorial University of Newfoundland, St. John’s NL, Canada. simtiaz@mun.ca 3 Faisal Khan is faculty member of Department of Engineering and Applied Science, Memorial University of Newfoundland, St. John’s NL, Canada. fikhan@mun.ca 4 M.A.A Shoukat Choudhury is faculty member of Department of Chem- ical Engineering, Bangladesh University of Engineering and Technology, Dhaka, Bangladesh shoukat@che.buet.ac.bd out between two control structures: MPC cascaded to PID and MPC directly manipulating the valve output. II. LITERATURE REVIEW A. Current State of PID Controller PID is a widely used control structure in the industry. Desborough and Miller estimated that 98 percent of the controllers in a median chemical plant are PID controllers [2]. Though it is widely used for its simplicity of imple- mentation, it has different limitations. The main limitation of the PID is that it has no straightforward tuning method. The impact of this fact is evident from the result reported by Van Overschee and De Moor [3]. They concluded that 80 percent of industrial PID controllers are poorly tuned; 30 percent of these PID loops operate in manual mode; and 25 percent of the PID loops in automatic mode operate under default factory settings. A control structure to overcome the drawbacks of the conventional PID controller with fixed tuning parameters, was proposed in [4], where PID gains are automatically tuned in order to keep a predefined cost function to a minimum. The applied methodology showed superior performance com- pared to PID in both set point tracking and regulatory control. Another simple but robust technique is described in [5], combining the simplicity of PID and versatility of MPC together. In this work tuning parameters are defined based on the key performance indices such as set point tracking, disturbance rejection, and the robustness and aggressiveness of the controller. The controller showed better performance in set point tracking and disturbance rejection compared to an IMC-tuned PID controller in extensive simulation studies. The potential alternatives for PID in industrial settings are investigated in [6]. Discrete-time linear MISO controller, state feedback and observers (SFO), model predictive con- troller (MPC) and fuzzy control are mentioned as potential alternatives. All alternatives showed improved performance, especially for poorly damped systems. Controllers based on SFO require a greater modeling effort, as such its use is justified only when modeling efforts are moderate. MPC is typically used as a supervisory layer to the base layer PID. The use of MPC provides a drastic improvement of set point tracking. Moreover, computational complexity is minimized in this case, as MPC executes at a slower rate, regulating the slower dynamics of the system. The PID layer reacts for the fast interactions. Pannocchia et el. proposed an offset-free constrained linear quadratic (CLQ) controller as a potential candidate to replace PID [7]. CLQ consists of three main modules based on a state-space model of the system: a state and disturbance estimator, a constrained target calculation 455