An off-line approach for output feedback robust model predictive control☆
Introduction
Model predictive control (MPC) has been considered as an effective optimal control approach for its ability to explicitly dispose of physical constraints as well as multivariable control in a systematic way. MPC has been originally developed for chemical process control, where the controlled systems often have slow dynamics and a large number of constraints on control inputs and system states [1], [2], [3], [4]. The advantages of MPC have been widely accepted both in academia [5], [6], [7], [8], [9] and other engineering industries such as automotive industries [10], [11], [12], power systems [13], [14], [15], robotics applications [16], [17], [18], brain-machine interface systems [19], [20], cyber-physical systems [21], [22], and mechatronic applications [23]. However, most of real processes have system uncertainties (i.e., model uncertainties, external disturbances and/or noises), and the models of real systems are difficult to be described. Thus, MPC techniques have to be extended to incorporate system uncertainties. The studies of MPC optimization formulations that ensure recursive feasibility and stability in the presence of system uncertainties are often referred to as robust MPC (RMPC) [7], [8], [9]. In RMPC optimization problems, the predicted system future behaviors are often related to the current system true states. In some applications of RMPC approaches, it is either impossible to directly measure all system states or economically nonviable. Hence, some RMPC schemes rely on state estimation techniques for their deployments. These motivate extensive attentions on output feedback RMPC (OFRMPC) approaches, where the synthesis of robust state estimation and observer-based robust MPC control is often considered [24], [25], [26], [27], [28], [29], [30], [31].
Linear parameter varying (LPV) systems can approximate nonlinear systems and represent uncertain systems [32], [33]. For LPV systems subject to bounded disturbances, the min-max OFRMPC approaches (e.g., state feedback robust MPC approaches in [35], [36] and OFRMPC approaches in [26], [27], [28]) consider that all the possible future closed-loop system states predicted at the current sampling time are bounded within one common robust positively invariant (RPI) set, which incorporate system uncertainties and constraint satisfactions in optimizations via closed-loop system worst case predictions. In [30], [31], the on-line OFRMPC approaches consider quasi min-max optimizations to reduce the conservatism in control input optimizations. Compared with [28], [29], [31], the OFRMPC approach in [30] synthesizes the computations on observer and controller gains, and the refreshment of the estimation error bounds in one optimization problem. The optimization problems in [26], [27], [28], [29], [30], [31] are semi-definite programming (SDP) problems and can be solved via the linear matrix inequality (LMI) tool, where the computational burden increases with system dimensions and the number of sub-models in LPV systems.
In [34], [35], [36], [37], [38], [39], the off-line state feedback RMPC approaches reduce the computational burden by searching controller gains from the off-line look-up table. The time-varying controller gains are searched according to the ellipsoids with the closest containment of system true states. Different from the aforementioned off-line state feedback RMPC approaches, the off-line observer-based OFRMPC approaches in [29], [40], [41] and the off-line dynamic OFRMPC approaches in [42], [43], [44] consider the effects of estimation error bounds or system true state bounds on robust stability and physical constraints, which complicate the procedure of off-line OFRMPC algorithms. For off-line OFRMPC approaches, the main challenges are (1) the method on how to design the off-line look-up table with nested robust invariant sets and a larger feasible region; (2) the method on searching the time-varying observer gains and controller gains; and (3) the reduction on algorithm complexities of the search method for the off-line designed look-up table. The off-line OFRMPC algorithms in [41], [43] consider polyhedral estimation error constraints. The off-line OFRMPC approaches with ellipsoidal estimation error sets are investigated in [29], [40], [42]. The off-line OFRMPC approach with both polyhedral and ellipsoidal system true state bounds is investigated in [44], where polyhedral system true states are updated for tightening the ellipsoidal system true state bounds. The observer gains in [41] and sub-observer gains in [29] are fixed, and time-varying controller gains are searched according to the real-time estimated states and estimation error bounds. The off-line OFRMPC approaches in [29], [41], [42], [43], [44] are the extensions of [40] to LPV systems with bounded disturbances and noises, where time-varying estimation error bounds or system true state bounds are considered in the closed-loop stability and physical constraints. In [29], [41], [42], [43], [44], to off-line compute different sizes of ellipsoidal feasible regions for estimated states, a series of estimated states and estimation error sets should be generated in advance. However, the methods on how to generate these estimated states and estimation error sets are not clearly discussed. Moreover, for the OFRMPC approaches in [29], [41], [42], [43], [44], the ellipsoidal feasible regions for estimated states in the look-up table cannot guarantee to be nested in theory, which are disadvantageous for searching real-time controller gains or controller parameters.
The contribution of this paper is to design an off-line OFRMPC method for LPV systems subject to bounded disturbances and noises by generalizing the OFRMPC optimization in [30]. Firstly, a look-up table is off-line designed to save the optimized sub-observer gains with nested RPI sets and sub-controller gains with nested robust control invariant (RCI) sets. Then, based on time-varying estimation error bounds and estimated states, real-time sub-observer gains and sub-controller gains are on-line searched. Different from the off-line OFRMPC approaches in [29], [41], [42], [43], [44], the approach to design the look-up table is not required to pre-specify a set of estimated states and estimation error bounds. However, it is only required to pre-specify a reference estimation error set. Then, the nested RPI sets with the sub-observer gains for the estimation error system, and nested RCI sets with sub-controller gains for the state observer system can be iteratively computed in off-line OFRMPC optimizations. The on-line searched sub-observer gains ensure that estimation errors are convergent within non-increasing RPI sets, and the sub-controller gains steer the estimated states to the neighborhood of the origin. Different from the off-line OFRMPC approach in [29], both time-varying sub-observer gains and sub-controller gains are real-time searched. Compared with the off-line OFRMPC approach in [29], the feasible region for estimated states is larger and the conservatism in OFRMPC controller is reduced. Compared with the on-line OFRMPC approach in [30], the proposed off-line OFRMPC algorithm significantly reduces the on-line computational burden.
Denote the sets of real numbers, positive real numbers and integers as , and , respectively. A vector in -dimensional Euclidean space is denoted as . The notation represents the weight Euclidean norm of the vector , where is a symmetric weighting matrix. For two integers , denote . For two integers and , the notations and represent the minimal and maximal values of and , respectively. A matrix means that the matrix is the convex combination of the elements in the convex hull . The notations are the predicted values of at time , . The notation “” denotes a symmetric structure in LMIs. For two matrices and , the notations and , respectively, represent that the matrix is a symmetric positive semidefinite matrix and the matrix is a symmetric positive definite matrix. For a symmetric positive definite matrix and a scalar , denote the ellipsoidal set . The notation is denoted as for brevity. For two symmetric positive definite matrices and , the sets when the matrices . The Minkowski set addition of two sets is defined as . The notation represents a diagonal matrix composed of matrices and . denotes a proper dimensional identity matrix.
Section snippets
Uncertain system and physical constraints
Consider the following LPV system with disturbances and noises,where , , and are the system true state, control input, and output, respectively. The bounded disturbances and noises . The scheduling parameter and model parameters , i.e., the time-varying parameters with , where , are the
Constraints on augmented states
Definition 1 A set is an RPI set for system (7) if for all possible , system parameters in the convex hull , and such that hold for all . Theorem 1 Suppose that with and . For the given non-negative scalars and a positive scalar , if there exist non-negative scalars , a positive scalar , and matrices , , , such that Eqs. (8)–(9) are satisfied, then
Off-line OFRMPC approach
The off-line OFRMPC approach includes two parts: design the off-line look-up table that stores a sequence of nested RPI sets with sub-observer gains for estimation errors and nested RCI sets with sub-controller gains for estimated states in Algorithm 1; search the real-time sub-observer gains and sub-controller gains in Algorithm 2.
Simulation example
The LPV system parameters derived from references [29], [30] corresponding to system (1)–(2) are
Conclusions
An off-line OFRMPC approach for constrained LPV systems subject to bounded disturbances and noises is investigated. The guarantee of the nested RPI sets with sub-observer gains and nested RCI sets with sub-controller gains are simultaneously off-line optimized and saved in the look-up table. The searched time-varying sub-observer gains ensure that real-time estimation errors are bounded and convergent in the time-varying RPI sets. The searched time-varying sub-controller gains steer the system
Declaration of Competing Interest
Authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (48)
- et al.
Fast nonlinear model predictive control: formulation and industrial process applications
Comput. Chem. Eng.
(2013) Model predictive control: recent developments and future promise
Automatica
(2014)- et al.
An outlook on robust model predictive control algorithms: reflections on performance and computational aspects
J. Process Control
(2018) - et al.
Use of model predictive control for experimental microgrid optimization
Appl. Energy
(2014) - et al.
Distributed stochastic model predictive control for cyber-physical systems with multiple state delays and probabilistic saturation constraints
Automatica
(2021) - et al.
Real-time nonlinear MPC and MHE for a large-scale mechatronic application
Control Eng. Pract.
(2015) - et al.
Simultaneous nonlinear model predictive control and state estimation
Automatica
(2017) - et al.
Observer-based output feedback robust MPC via zonotopic set-membership state estimation for LPV systems with bounded disturbances and noises
J. Franklin Inst.
(2020) - et al.
An efficient off-line formulation of robust model predictive control using linear matrix inequalities
Automatica
(2003) - et al.
An ellipsoidal off-line model predictive control strategy for linear parameter varying systems with applications in chemical processes
Syst. Control Lett.
(2012)
An ellipsoidal off-line MPC scheme for uncertain polytopic discrete-time systems
Automatica
-step off-line MPC design of nonhomogeneous Markov jump systems: a suboptimal case
J. Franklin Inst.
Robust output feedback model predictive control using off-line linear matrix inequalities
J. Process Control
A synthesis approach for output feedback robust constrained model predictive control
Automatica
Off-line approach to dynamic output feedback robust model predictive control
Syst. Control Lett.
Off-line output feedback robust MPC with general polyhedral and ellipsoidal true state bound
J. Franklin Inst.
Asymptotically necessary and sufficient conditions for stability and performance in fuzzy control: applications of Polya’s theorem
Fuzzy Sets Syst.
An industry perspective on MPC in large volumes applications: potential benefits and open challenges
IFAC-PapersOnline
Nonlinear model-predictive control for industrial processes: an application to wastewater treatment process
IEEE Trans. Ind. Electron.
Model predictive control (MPC) and its current issues in chemical engineering
Chem. Eng. Commun.
Model Predictive Control
Handbook of Model Predictive Control
Robust model predictive control: reflections and opportunities
J. Control Decis.
The development of model predictive control in automotive industry: a survey
Proceeding of the IEEE International Conference on Control Applications, Dubrovnik, Croatia
Cited by (3)
Solving a reliability-performance balancing problem for control systems with degrading actuators under model predictive control framework
2022, Journal of the Franklin InstituteCitation Excerpt :Also, it is important to note that no researches have been conducted on the RPBP under the MPC framework, and MPC is very likely to be widely applied in industrial processes and aerospace engineering in the future. On the other hand, MPC has the conceptual simplicity and the inherent ability to handle complex systems with multiple constraints, as acknowledged in many literatures [31–35]. For example, Chai et al. [34] constructed two different MPC schemes to solve the reconnaissance trajectory tracking problem with control, state and path constraints, and proposed a two-nested gradient method to decrease the computational demand.
A Survey of Output Feedback Robust MPC for Linear Parameter Varying Systems
2022, IEEE/CAA Journal of Automatica Sinica