Research Article  Open Access
Robust TubeBased MPC with Piecewise Affine Control Laws
Abstract
This paper presents a tubebased model predictive control (MPC) algorithm with piecewise affine control laws for discretetime linear systems in the presence of bounded disturbances. By solving the standard multiparametric quadratic programming (mpQP), the explicit piecewise affine control laws for tubebased MPC are obtained. Each control law is piecewise affine with respect to the corresponding region (one of the partitions of the feasible set). Due to the fact that the abovementioned procedures are totally offline, the online computation time is short enough for stabilizing those systems with fast dynamics. In this paper, all the involved constraint sets are assumed to be polytopes. An illustrative example is utilized to verify the feasibility and efficiency of the proposed algorithm.
1. Introduction
The concept of “tube” in model predictive control (MPC) is proposed by [1] for linear discretetime systems with bounded additive disturbance. The core idea is to design a robust invariant set for the corresponding nominal system, and then the optimization problem with tightening constraints is solved to get the optimal control action [2]. Since all the predictions of state and control variables are confined to the tighten constraint set, the real dynamics of controlled systems will never deviate the constraints under the impaction of external disturbances. After the proposition of the algorithm, it has been greatly developed. In [3], the tubebased MPC algorithm is extended to the case of linear sampleddata systems, and in [4] the tracking problem is considered. In [5], the tuning parameter is incorporated into the basic MPC strategy which enables us to move smoothly from the existing controller to a better MPC strategy, and the methodology is applied to the tubebased outputfeedback MPC case. A probabilistic tube for linear systems with probabilistic disturbances is designed in [6], which avoids the computation burden in traditional stochastic MPC. A homothetic tube MPC synthesis method is proposed by [7], which utilizes a more general parameterization of the state and control tubes based on homothety and invariance. The tubebased MPC algorithms for nonlinear systems are presented in [8–10]. In [8], a general nonlinear finite horizon optimization problem with terminal zero constraints is resolved offline once, and the optimal sequence of states and inputs are taken as the reference trajectories, then the online optimization problem is designed to tracking these trajectories with constraints satisfaction. In [9], the nonlinear models are locally linearized, and the errors between the linearized models and the true models are confined to lie in the predesigned robust tubes. In [10], a tubebased MPC algorithm for continuoustime nonlinear systems, which satisfies the Lipschitz condition, is proposed. Although the theory of robust tubebased MPC has been stimulated a lot, but far from perfect.
The optimization problems in standard MPC algorithms are in general the linear programming (LP) or linear quadratic programming (QP), and the computation time for LP/QP can not be neglected when the controlled systems have fast dynamics. The explicit MPC takes the LP or QP in standard MPC as multiparametric linear programming (mpLP) or multiparametric quadratic programming (mpQP) and solves these optimization problems absolutely offline. The basic results and development on multiparametric programming can be found in [11–14] for mpLP and [15–19] for mpQP. In this paper, we mainly consider the mpQP, which is the most general form of optimization we faced in the regular MPC. In [20], the theoretical perspectives of multiparametric programming and explicit MPC are introduced. In [21], the authors present a method to compute the explicit statefeedback control laws for both the MPC algorithm and the constrained linear quadratic regulation problem with guaranteed feasibility, stability, and optimality, in which the explicit feedback control laws are piecewise affine and continuous. In [22], the approximated explicit control laws for MPC are obtained, which utilizes the inner/outer polytopic approximation technique and the implicit doubledescription algorithm. In [23], the systematical procedures for the analytical expression of explicit control laws of linear MPC via piecewise affine function are given, which saves the online computation time and memory requirements. There are many works which have been published on finding ways to resolve the mpQP. In [17], an efficient mpQP solver which avoids unnecessary partitioning of the parameter space by directly exploring the neighborhood of initial partition is presented. In [18], an mpQP solver with a new partitioning method of the parametric (state vector) space is proposed, which avoids the unnecessary partitioning and improves the efficiency. In [19], an algorithm is proposed to revise the existing algorithms in order to make them satisfy the facettofacet property in general and guarantee that the entire parameter space is explored. Because the partitions of feasible sets are increasing exponentially with respect to the prediction horizon, the complexity reduction methods are presented by [24–26] to remove the unnecessary partitions. The only online calculation for mpQP in MPC is the point location problem (to confirm which region contains the current parameter). In [27–29], different efficient methodologies have been proposed to resolve this problem.
This paper considers the constrained discretetime linear systems with additive disturbances. The disturbances are assumed to be confined in a polytope. First, the standard tubebased MPC algorithm, which solves the quadratic programming (QP) online and takes initial state in the optimization as an optimized variable, is designed. Then, by transforming the optimization problem in standard tubebased MPC into the mpQP form, the optimization can be solved through mpQP solvers, which separate the parameter space into finite partitions and get a piecewise affine linear optimal solution for each partition. In this paper, the standard tubebased MPC algorithm is solved absolutely offline, and the only online calculations are to confirm which partition the current state lies in.
The remainder of this paper is structured as follows. In Section 2, the basic formulation of the controlled systems and basic definitions on polytope and polyhedron are introduced. The multiparametric programming is described in Section 3. In Section 4, the tubebased MPC algorithm with piecewise affine state feedback control laws is presented. In Section 5, the simulation results are provided to show the effectiveness of the proposed algorithm.
2. Problem Statement and Preliminaries
Consider the following discretetime linear systems with additive disturbances: where , , and are the state, input, and disturbance vector at sampling time , respectively. The system states and inputs are constrained by and the disturbance is bounded by The corresponding nominal model of (1) is given as The system (1) is supposed to satisfy the following assumptions throughout the paper.
Assumption 1. The pair of system (1) is stabilizable.
Assumption 2. The states of system (1) are measureable.
Assumption 3. , , and are compact and convex and contain the origin as interior point, respectively.
Since all the involved constraint sets in this note are confined as the convex and bounded region, that is, the polytope, the basic definitions related to polyhedron and polytope are shown to make the paper more complete and concise. For more knowledge about the polyhedron and polytope, readers can refer to [30–32].
Definition 4. A polyhedron is a convex set which originates from the intersection of a finite number of halfspaces where and and is a finite integer.
Definition 5. A polytope is the bounded polyhedron
Definition 6. The representation of a polytope , as in (6), is to depict the polytope as an intersection region of a finite number of halfspaces. The other way of statement is that the representation of a polytope is the region described by a finite number of linear inequalities ; that is, the dimensions of and are not infinite.
Definition 7. The representation of a polytope is to describe the polytope as a convex hull of its vertices: where are the vertices and is the total number of vertices.
After the basic definitions of polytope have been depicted, the related operations involved in the remaining part of the paper are listed below.
Definition 8. The Minkowski sum of two given sets and is defined as , and the Minkowski difference is .
Definition 9. The set difference of two sets and is defined as
Definition 10. Given a set , the affine map of by the mapping , where , to a set , then is defined as where and are constant matrix and vector, respectively.
Remark 11. Another problem we often faced is to find a polytope satisfying that is, to compute a polytope which is a map to . It is the inverse problem of Definition 10 and can be easily computed by using the affine map in Definition 10. A function “range” is provided by Matlab toolbox called MPT [31] to compute the in Definition 10 and a function “domain” for calculating in (10).
Due to the fact that all the variables in this paper are bounded by the polytopes, the state, input, and disturbance constraint sets are assumed to have the following form: where , , , , , and .
3. Multiparametric Quadratic Programming (mpQP)
The multiparametric programming is the linear or nonlinear programming with parameters in the objective function or/and constraints. The semiinfinite programming and multilevel programming can be cast as special multiparametric programming. The main advantage of multiparametric programming is that the explicit piecewise affine solutions with respect to the parameters can be calculated. In this section, the general formulation of multiparametric programming is introduced. We refer the readers to [11–19] and the references therein for the basic definitions and results of the multiparametric programming used in this paper.
The general formulation of multiparametric programming is as follows:where is the optimization variable and is the parameter variable.
Define the Lagrangian as where and are the Lagrange multipliers.
The firstorder KarushKuhnTucker (KKT) optimality conditions arewhere the gradient function is defined as
Remark 12. The firstorder KKT condition is a necessary condition for a solution in constrained programming to be optimal. That is to say, the solution which satisfies the firstorder KKT condition may not be the global optimal point for the constrained programming. If we want to determine whether the calculated point is global optimal point or not, the secondorder optimality condition is needed.
Although the solution satisfying the firstorder KKT condition may not be the optimal one, for the case in the following theorem, the solution is optimal.
Theorem 13. For a fixed , satisfying the firstorder KKT conditions (14a)–(14g) is the global optimal solution of (12a), if the objective function and inequality constraint functions , , in optimization problem (12a), (12b), (12c), and (12d) are convex functions and the equality constraint functions , , are linear functions. In this case, the firstorder KKT condition is the necessary and also the sufficient condition for optimality.
Proof. The proof can be found in [33].
Most of the optimization problems involved in linear control theory can be transformed into the mpLP or mpQP. The mpLP and mpQP correspond to the case of linear systems with linear performance cost functions and linear systems with quadratic performance cost functions, respectively. In general MPC algorithms, the mpQP is the most common optimization problem we utilized.
In this paper, we consider the following form of mpQP problem [17, 18, 21]:where , , , , , , and .
Then, from (14a), (14b), (14c), (14d), (14e), (14f), and (14g) the firstorder KKT conditions of (16a) and (16b) arewhere , , , and .
Define . By analyzing the KKT conditions (17a), (17b), (17c), and (17d), the explicit solution at a given is given by [18] where denotes the set of active constraints in (16b), , and , , and . Since the in (18) is the function of , define as the region of in which the equation of in (18) remains optimal. The region is a polytope and its representation is as follows [18]: where
To solve the mpQP problem (16a) and (16b), the feasible region of needs to be confirmed firstly. Assume that the region and is the largest ball contained in ; then it can be calculated by solving the following LP problem [21]:
In order to get the explicit piecewise affine solutions on the whole feasible set , the feasible set needs to be partitioned into a finite number of polytopes. Suppose is partitioned into partitions; then the partitions , , have to satisfy(i);(ii), , where denotes the border of partition .
Remark 14. The partitions , , are determined by finding the region defined in (19) on the whole . There now exist many methods to determine the partitions , such as the method by exploiting the facettofacet property [19, 21] and adding/withdrawing constraints from active set [18].
Theorem 15. Consider the mpQP of (16a) and (16b) and let be a polytope. Then the feasible set is convex, the optimizer is continuous and piecewise affine in each partition , and the value function is continuous, convex, and piecewise quadratic.
Proof. The proof is in [21].
Remark 16. The convexity and continuous properties in Theorem 15 are critical for the optimality of mpQP. From Theorem 25, since all the involved sets are convex, the inequality constraints in (16b) and the objective function are convex functions in each partition ; then the optimizer is the optimal solution of (16a) and (16b) in the whole region of each partition .
4. TubeBased MPC with Piecewise Affine Control Laws
In this section, a robust tubebased MPC algorithm with piecewise affine solutions is proposed. For the basic definitions and researching development on the general tubebased MPC algorithm, readers can refer to [1, 2].
For clarity, the linear nominal model of (1) is rewritten as
Since the system (22) satisfies Assumption 1, then there exists a feedback gain matrix , such that closedloop system , where , is stable.
In the sequel, the basic definition of robust invariant set is introduced by referring to [34].
Definition 17. The set is said to be a robust invariant set for uncertain linear system , if for any and any sequence of disturbances , the solution of satisfies for all .
Remark 18. In [2], a concise description of the condition is presented as , where denotes the Minkowski sum defined in Definition 8.
By assuming that the set is a polytope, that is, , the robust invariant set is chosen as the minimal robust positive invariant (mRPI) set for uncertain linear system by the outer approximation approaches in [35]. The main procedures are listed as follows.
Algorithm 19 (the procedures for finding ).
Step 0. Given , , and let .
Step 1. Set .
Step 2. Compute by the following equation:
where and is the standard basis vector in . Set .
Step 3. Compute as
Step 4. If , go back to Step 1.
Step 5. Compute as the Minkowski sum
and let be the outer approximation of robust invariant set .
Proposition 20. Given a disturbance invariant set for , let and be the state and input pair of the system and , respectively. If and , then for all .
Proof. The proof is in [1].
For the purpose of proving the stability of tubebased MPC, the general robust MPC algorithm with tighten constraints is described as follows:with the objective function where is the prediction of at the future time , predicted at time , and equals the current state . The corresponding weighting matrices , , and in (27) are chosen to be positive definite. The tighten constraint set , is defined as , . The terminal constraint set satisfies and contains the equilibrium point in its interior. The optimization (26a), (26b), (26c), (26d), and (26e) can easily be transformed into the QP problem, the solvers for which are fast and efficient in most cases.
Let the terminal cost function be a Lyapunov function; then the terminal constraints are chosen to satisfy the following stability conditions [2, 36]: C.1: , , , C.2: , .
The principle for choosing is to satisfy condition C.1. Then, the terminal region can be chosen as the maximum output admissible set , as in [37]. Consider the linear autonomous system with the output constraint (state constraint) , where , . The main procedures to determine are listed as follows [37, 38].
Algorithm 21 (the procedures for finding ).
Step 0. Set .
Step 1. Solve the following linear programming problem for :
where denotes the th row of the matrix .
Step 2. If , , stop. Set
Otherwise, continue.
Step 3. Set . Go back to Step 1.
By solving the optimization (26a), (26b), (26c), (26d), and (26e), the optimal control sequence is . The control input , applied to the system (1), is the first element of . Then, define where the symbol is defined to make differences between the solution of (26a), (26b), (26c), (26d), and (26e) and other optimizations which will be described later.
The step feasible set for (26a), (26b), (26c), (26d), and (26e) is defined as
Remark 22. The step feasible set for optimization (26a), (26b), (26c), (26d), and (26e), which is also called the step controllable set, is the set of initial states which can be stabilized to the target set in steps. The procedures for computing the step feasible set can be found in [32].
The robust tubebased MPC algorithm, proposed by [1, 2], is to solve the following optimization:where the parameters , , , , , and and the objective function are the same as in optimization (26a), (26b), (26c), (26d), and (26e). In this optimization, the current state does not equal , which is the prediction of current state at time . is the robust invariant set solved by Algorithm 19.
Remark 23. The optimization (32a), (32b), (32c), (32d), (32e), and (32f) can also be transformed into the QP form. The significant difference between the tubebased MPC (32a), (32b), (32c), (32d), (32e), and (32f) and the general constraint tightening MPC (26a), (26b), (26c), (26d), and (26e) is that is taken as an optimization variable in (32a), (32b), (32c), (32d), (32e), and (32f), which does not equal but satisfies the constraint . Moreover, the constraint is equivalent to . And in particular , when is symmetric. In Remark 11, the way of computing the constraint set of variable by affine map is shown.
Then, the step feasible set for (32a), (32b), (32c), (32d), (32e), and (32f) is defined as
Remark 24. The step feasible set can also be computed using the method in [32].
The robust tubebased MPC solves the optimization (32a), (32b), (32c), (32d), (32e), and (32f) repeatedly at each time instant . The optimal solutions are , and , which related to the real initial state . The control law to be applied is chosen as
Theorem 25 (stability results). The system dynamics of (1) under control input satisfies , , and the robust invariant set is robustly exponentially stable for system , , in the region of .
Proof. The proof is in [2].
In the sequel, we define , , , , and , where we can get and from Remark 11. The symbol denotes the inverse of matrix and produces an identity matrix.
In spite of the fact that the QP problem can be solved efficiently by the existing solvers, for systems with fast dynamics, the computational burden is still too huge to control these systems well. Even worse, the long computation time may make the controlled systems unstable. A viable way to tackle this problem is to use the offline approach, which solves the related optimization totally offline. The explicit regulator together with multiparametric programming is an alternative way to give a control strategy with less computation time.
Since the initial state in the optimization (32a), (32b), (32c), (32d), (32e), and (32f) is as an optimization variable, let be the augmented vector of optimization variables and .
Then, the optimization problem (32a), (32b), (32c), (32d), (32e), and (32f) can be transformed into the following mpQP:where . The matrices , , , and are defined as with
By applying the algorithm in (21a), (21b), and (21c), the feasible set of is calculated, and we have
Remark 26. Since the optimization problems (32a), (32b), (32c), (32d), (32e), and (32f) and (36a) and (36b) are the same one with different forms, the feasible set equals the feasible set .
By solving the mpQP (36a) and (36b) through algorithms in [18, 19, 21], the feasible region is assumed to be partitioned into partitions and each partition , , is a polytope. The explicit piecewise affine solution with respect to is defined as where denotes the piecewise affine function of in partition , , , and , .
Due to the fact that the optimization sequence is the combination of and , the decomposition procedures are needed to get the explicit solutions for (32a), (32b), (32c), (32d), (32e), and (32f). Define where , , , , , and , for all .
Then, the piecewise affine state feedback control laws in (32a), (32b), (32c), (32d), (32e), and (32f) are and the corresponding piecewise affine linear is calculated by where and denote the piecewise affine function of and with respect to in partition , respectively.
The explicit piecewise affine linear control law for robust tubebased MPC at the initial point is defined as
Remark 27. Comparing (44) with (34), the control law (44) is explicit piecewise affine linear with respect to the initial state and can be computed totally offline. The piecewise affine linear control laws are prestored in the memory of computer, and when these control laws are needed to be applied online, the only measure to take is finding the partition which contains by searching in a lookup table. So the computation time has been extremely decreased, and the algorithm is fit for controlling those systems with fast dynamics.
Substituting (42) and (43) into (44) yields the concise form of (44) as
Since the explicit piecewise affine state feedback control laws are stored in a lookup table, certain searching methods are needed to confirm which partition contains the current state . The related procedures are referred to as the wellknown “point location” problem. There already exist many viable and efficient methods to figure out these point location problems, such as the binary search tree utilized in [27], the subdivision walking method in [28], and the hash tables in [29].
Theorem 28. Consider the mpQP of (36a) and (36b) with piecewise affine solutions (44); then one has the following. (i)The feasible set is convex, the optimizer is continuous and piecewise affine in each partition , and the value function is continuous, convex, and piecewise quadratic.(ii)The piecewise affine linear control law in (44) for tubebased MPC algorithm is continuous and piecewise affine in each partition .
Proof. Since (ii) in the theorem is deeply involved with (i), (i) and (ii) will be proven separately and orderly. We first prove (i). It is obvious that (i) in Theorem 28 is the direct conclusion of Theorem 15; then (i) is proven. We then prove (ii). From (i), is continuous and piecewise affine in . Also by (18), is linear with respect to in . Since , where is the identify matrix, and , where is the identify matrix, and are the combinations of ; then and are continuous piecewise linear in . Also is the combination of and ; then is continuous and piecewise affine in each .
In the sequel, the robust stability of the disturbed linear system (1) under explicit piecewise affine control law (44) is analyzed.
Theorem 29 (stability results). The system dynamics of (1) under control input satisfies , , , , and the robust invariant set is robustly exponentially stable for system , , in the region of .
Proof. In this proof, we discard the traditional idea as in [39, 40], which takes the system (1) under piecewise affine control law as a switched linear system. To prove the stability of these switched linear systems, we need to construct and then prove the asymptotical decreasing property of a series of piecewise Lyapunov functions for each separate subsystem, which is fairly complex. So we changed the way to prove that the piecewise affine control law is equivalent to for all , . Since all the constraints in (36b) are linear functions for each fixed and the objective function is a quadratic function, then both of them are convex functions. From Theorem 25, we can see that which satisfies the firstorder KKT condition is the optimal solution of (36a) and (36b) at ; then it means and are also optimal at , ; then is optimal in , . If , , the corresponding optimal solutions of (32a), (32b), (32c), (32d), (32e), and (32f) are and . Since the optimizations (36a) and (36b) and (32a), (32b), (32c), (32d), (32e), and (32f) are the same optimization with different forms and equals , combined with Theorem 28, we conclude that the piecewise affine control laws in (44) equal the control law for all . Then, by virtue of the robust stability results in Theorem 25, Theorem 29 is proven.
5. Illustrative Example
In order to verify the efficiency of the control strategy, the proposed algorithm is applied to a disturbed linear system, which is borrowed from [2]. For simplicity and comparison, the related parameters are mostly the same as [2]. The system model is described as follows: where
The bounds on the state, control, and disturbance are
The corresponding nominal model is
Choose
For the nominal model (49), the optimal statefeedback gain is calculated by solving the discretetime LQR problem. By applying the Matlab function “dlqr,” is calculated as and the corresponding positive symmetric matrix is
The robust invariant set for (1) under can be computed from Algorithm 19. is a polytope and its representation is where
Similarly, the terminal region can be calculated from Algorithm 21. The representation of is where
Choose and initial state . By utilizing the algorithm proposed in this paper, the feasible sets , , and are plotted in Figure 1, the partitions of are shown in Figure 2, and the phase trajectory with embedded robust invariant sets from the starting point is plotted in Figure 3. In order to test whether the constraints on state and input are violated or not, the Monte Carlo method is utilized and the results are shown in Figures 4 and 5. Also to verify the feasibility in the whole feasible set , different representative initial points are simulated and the corresponding phase trajectories are plotted in Figure 6. Moreover, for comparison, the partitions of the feasible sets , , are plotted in Figures 7, 8, 9, and 10. The total number of partitions of , , is shown in Table 1. And the comparison of online computation time between the method in [2] and the proposed method in this paper for each control step is given in Table 2. For simplicity, the representative partitions of and corresponding piecewise affine control laws are listed in Table 3.


