Received: 26 June 2022 Revised: 7 September 2022 Accepted: 28 September 2022

DOI: 10.1002/rnc.6409

RESEARCH ARTICLE

Adaptive control Lyapunov function based model predictive

control for continuous nonlinear systems

Yi Zheng Yanye Wang Shaoyuan Li

Department of Automation, Shanghai Jiao

Tong University, Key Laboratory of Abstract
System Control and Information A design of adaptive model predictive control (MPC) based on adaptive control
Processing, Ministry of Education of
Lyapunov function (aCLF) is proposed in this article for nonlinear continu-
China, Shanghai, China
ous systems with part of its dynamics being unknown at the starting time.
Correspondence Specifically, to guarantee the convergence of the closed-loop system with online
Yi Zheng and Shaoyuan Li, Department of
Automation, Shanghai Jiao Tong
predictive model updating, a stability constraint is designed. It limits the aCLF
University, Key Laboratory of System of the system under the MPC to be less than that under an online updated aux-
Control and Information Processing, iliary adaptive control. The auxiliary adaptive control which implements in a
Ministry of Education of China, Shanghai
200240, China. sampling-hold fashion can guarantee the convergence of the controlled system.
Email: [email protected] and The sufficient conditions that guarantee the states to be steered to a small region
[email protected]
near the equilibrium by the proposed MPC are provided. The calculation of the
Funding information proposed algorithm does not depend on the model mismatch at the starting
National Key R &D Program of China, time. And it does not require the Lyapunov function of the state of the real sys-
Grant/Award Number: 2018AAA0101701;
tem always to be reduced at each time. These provide the potential to improve
National Natural Science Foundation of
China, Grant/Award Numbers: 62073220, the performance of the closed-loop system. The effectiveness of the proposed
61833012 method is illustrated through a chemical process example.

KEYWORDS
adaptive control, Lyapunov-based control, model predictive control, nonlinear system

1 I N T RO DU CT ION

With the development of industrial technologies, industrial processes become more and more complex, and the require-
ments for the operational performance of the process and the quality of production become higher and higher. Obtaining
accurate models is important as process models are key elements of advanced model-based control systems. However, the
modeling of complex nonlinear processes continues to be a major challenge in process systems engineering. Especially, for
the system with uncertainties, or time-varying dynamics that may be caused by equipment wearing or the environment
change. It is difficult to obtain the accurate dynamic in advance for these systems.
For the control of systems with time-varying dynamics or systems whose dynamic models cannot be accurately
obtained at the starting time, many control algorithms appear in recent decades, regardless of whether they are
model-based control algorithms or model-free algorithms.1,2 An efficient and proved practical way is to incorporate the
system identification or machine learning-based technologies into the online update of the system model, for example,
the indirectly adaptive control,3,4 and the classic MIT rule,3 the machine learning-based control method,5,6 the reinforce-
ment learning-based control.7 Designing model-based adaptive control systems that utilize identification techniques or
machine learning modeling techniques to account in real-time for data with guaranteed stability becomes a hot topic in
control systems.

1254 © 2022 John Wiley & Sons Ltd. wileyonlinelibrary.com/journal/rnc Int J Robust Nonlinear Control. 2023;33:1254–1266.
 ZHENG et al.

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1255

It is worth mentioning that, in the design of the adaptive control algorithms, a tool named adaptive control Lyapunov
function (aCLF) is provided in References 8 and 9 to analyze the stabilization of the adaptive system and to design the
stabilized adaptive control.3 This method does not require that the model used in the controller converge to the real
system, the mismatch between the system model and the real system and its boundary is not used in the updating of the
control parameters.10,11
Model predictive control (MPC) optimizes the state trajectory of the system over a finite prediction horizon with input
and state constraints, and the first element of the optimal solution is applied to the controlled system at each sampling
time. It has the virtue of explicitly accommodating some essential constraints and is widely recognized as a practical
multiple variables control technology with high-performance.12-15 Since the adaptive MPC not only has the advantages of
the MPC mentioned above but also has the virtues of the adaptive control, for example, the accurate system model at the
start time is not required, and it can automatically adapt to the system dynamic and environment, the design of adaptive
MPC with guaranteed properties attracts many scholars to study.
The design of the MPC which guarantees the stabilization of the closed-loop system has been extensively studied.
Among them, Reference 16 proposed a quasi-infinite horizon MPC for nonlinear systems, where a terminal set com-
bined with the final weighting is employed to guarantee the recursive feasibility and stability of the closed-loop system.
For systems with uncertainties, Reference 17 proposed a tube-based MPC which uses a tightened constraint to guaran-
tee the recursive feasibility of MPC systems.18 The tightened constraints exclude the boundaries of possible affection
of uncertainties on states and inputs. Reference 19 proposed an MPC that uses the Lyapunov technique to ensure the
stability and feasibility of the closed-loop system when the states are bounded in the feasible region. As extensions,
Reference 20 proposed a Lyapunov-Barrier function-based MPC with the consideration of system safety. Also, some
robust methods are proposed for uncertain nonlinear systems where different types of time-delays are considered and
addressed.21-23 These might provide the help for a further extension of the analysis and synthesis for delayed nonlinear
systems.
For adaptive/learning-based MPC systems with online model updating, the stability of a closed-loop system will not
always be guaranteed with the varying of the predictive model if the parameters of MPC are designed by the initial model.
One method to ensure the stability and constraint satisfaction of the adaptive/learning-based MPC systems is to use the
maximum boundary of uncertainty to design additional constraints for MPC optimization problems.24 In the efforts of
reducing conservation of the constraints in adaptive MPC, Reference 25 proposed a method to update MPC constraints
for linear systems that benefits from the estimation of the reduced uncertainties set. For nonlinear systems, Reference 26
provided an adaptive MPC for unconstrained nonlinear systems with parametric uncertainties. Reference 27 proposed
an adaptive LMPC, in which the stability constraint is designed based on the given model mismatch level rather than the
initial mismatch, and then the constraints can be relaxed.
Except for the MPCs which update the model every sampling time, References 28 and 29 proposed event-triggered
mechanisms to keep the model mismatch always within a certain range which does not affect the convergence of the
closed-loop system. Designing an MPC that enhances the performance of nonlinear systems with guaranteeing the
stability of the closed-loop system is still an open problem to be studied.
From the above, it can be seen that designing adaptive MPC with online constraints update, except for the online
model update, will contribute to the improvement of system performance.7,25 For the online updating of constraints,
the estimation of model mismatch is required. Since the dynamic of the system is unknown until it was identified, it is
hard to accurately estimate the model mismatch. The designed control system by aCLF does not estimate the bound-
ary of model mismatch or use it in the calculation of control outputs. It does not need that the Lyapunov function of
the real system is always decreasing, which may provide a potential to improve the optimization performance, that is,
the economic performance of the closed-loop MPC system. Therefore, this article considers a design of MPC based on
the aCLF.
In the proposed MPC, the MPC law is implemented in a sampling-and-hold fashion and a constraint for the MPC
optimization problem is designed based on an online updating auxiliary adaptive control. The auxiliary adaptive control
is designed according to the aCLF. The designed constraint limits the aCLF of MPC to be less than that under the auxiliary
adaptive control at each sampling time instant. The sufficient conditions which guarantee that the states of the closed-loop
systems are steered into a small region around the equilibrium are provided. The advantages of the proposed method are
listed as follows:

• The designed aCLF-based stability constraint provides a potential for enhancing the optimization performance of the
closed-loop system.
 Downloaded from https://iranpapr.ir ZHENG et al.

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1256

• The designed MPC guarantees the convergence of the closed-loop system containing unknown dynamics if the given
conditions are satisfied.
• The model used to predict future state trajectories can be different from the model used in the stability constraint,
which can be formulated or updated by accurate machine-learning methods for performance improvement.

2 PRELIMINARIES

2.1 Notations

Notation | ⋅ | donates the Euclidean norm of a vector. A continuous function 𝛼 ∶ [0, a) → [0, ∞) is said to belong to class
 if it is strictly increasing and 𝛼 (0) = 0. The symbol Sr (x) denotes the set Sr (x) ∶= {x ∈ Rnx ∶ V1 (x) ≤ r} where V1 is
a scalar function. The symbol Ω𝜌 (x, 𝜃) denotes the set Ω𝜌 (x, 𝜃) ∶= {x ∈ Rnx , 𝜃 ∈ Rn𝜃 ∶ Va (x, 𝜃) ≤ 𝜌} where Va is a scalar
function. A function f (x) is said to be locally Lipschitz with respect to its argument x if there exists a positive constant Lxf
such that |f (x1 ) − f (x2 )| ≤ Lxf |x1 − x2 | for all x1 ≠ x2 in the given region of x and Lxf is regarded as the Lipschitz constant.
A∕B ∶= {x ∈ Rnx ∶ x ∈ A, X ∉ B} represents the subset.

2.2 Class of systems

We consider a class of nonlinear systems with unknown parameters, whose dynamics can be described by the following
state-space model

ẋ = f (x) + F(x)𝜃 + g(x)u, (1)

where 𝜃 is a vector of unknown constant parameters and 𝜃 ∈ Rn𝜃 , x ∈ Rnx denotes the vector of states variables. u ∈
 ∶= {u ∈ Rnu ∶ |u| ≤ Mu } is a bounded vector of control inputs, and  ⊂ Rnu is a convex set. The f (x), F(x), and g(x)
are locally Lipschitz functions, and f (0) = 0, F(0) = 0. The constants Mf , MF , and Mg represent the maximum of f (⋅), F(⋅)
and g(⋅). The origin is an equilibrium point of the system (1).
The aim of this article is to design an adaptive MPC to steer the states of the system (1) to a small region around the
equilibrium.

2.3 Adaptive control Lyapunov functions

Since there are unknown parameters in system (1), we consider an adaptive controller that is dependent on the estimated
parameters 𝜃.̂

Definition 1. A smooth function Va (x, 𝜃) ∶ Rnx × Rn𝜃 → R+ , positive definite, and satisfies

𝛼1,𝜃 |x| ≤ Va (x, 𝜃) ≤ 𝛼2,𝜃 |x|, (2)

is an aCLF for system (1), if there exists a positive-definite matrix Γ ∈ Rn𝜃 ×n𝜃 such that for every 𝜃 ∈  ⊂ Rn𝜃 , Va (x, 𝜃) is
a control Lyapunov function (CLF) for the system
( ( )T )
𝜕Va
ẋ = f (x) + F(x) 𝜃 + Γ + g(x)u, (3)
𝜕𝜃

that is
{ [ ( ( )T ) ]}
𝜕Va 𝜕Va
inf f (x) + F(x) 𝜃 + Γ + g(x)u < −𝛼3 |x| , (4)
u∈Rnu 𝜕x 𝜕𝜃

where 𝛼3 (⋅) is a positive constant.

 ZHENG et al.

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1257

If an aCLF Va of system(1) exists, the system can be globally adaptively stabilized. And the parameters-based controller
̂ can be constructed according to the update of 𝜃,
k(x, 𝜃) ̂ with which the global stability can be guaranteed.
The adaptive controller can be expressed as
( )T
̂ = ΓF(x)T 𝜕Va ,
̂𝜃̇ = Γ𝜏(x, 𝜃) (5a)
𝜕x
̂
u = k(x, 𝜃), (5b)

̂ = 0. 𝜃̂ denotes the estimate of the parameters, and the adaptation

where, k(⋅, ⋅) is smooth on (Rnx ⧵ {0}) × Rn𝜃 and k(0, 𝜃)
̂
law of 𝜃 can take the form of Equation (5a), where 𝜏 (⋅, ⋅) is a smooth function on Rnx × Rn𝜃 .
Remark 1. A candidate controller u = k(x, 𝜃) can be selected as the Sontag control law for system (3) to make (4) holds.
( ( 𝜕V )T ) 𝜕V (x,𝜃) 𝜕V (x,𝜃)
Refer to Reference 30, define f ′ (x, 𝜃) = f (x) + F(x) 𝜃 + Γ 𝜕𝜃a , f ′ = a𝜕x f ′ (x, 𝜃), and g = a𝜕x g(x). Then, the
controller k(x, 𝜃) = [k1 (x, 𝜃), k2 (x, 𝜃), … , knu (x, 𝜃)] can be given as
T

√
⎧ p+ p2 + 𝛾|q|4
⎪− qi qi ≠ 0,
ki (x, 𝜃) = ⎨ |q|2 (6)
⎪0 qi = 0,
⎩
[ ]
where p = f ′ Va (x, 𝜃), qi = gi Va (x, 𝜃), 𝛾 > 0 and g = g1 , g2 , … , gnu , gi (⋅) ∶ Rn → Rn . The stability of the system (3)
under the control law k(x, 𝜃) can be ensured and the detailed proof can refer to Reference 30.
̂ ⊆ O as the stability region of the closed-loop system under the adaptive controller u = k(x, 𝜃)
We denote Ω𝜌 (x, 𝜃) ̂ and
u ∈  , where O is an open neighborhood of the origin. For all (x, 𝜃), (x , 𝜃 ) ∈ Ω𝜌 and u ∈  , by the continuously differ-
′ ′

entiable property of Va (x, 𝜃) and the local Lipschitz property of f (x), F(x), g(x), there exist positive constants Lf , LF , Lg ,
M, and M𝜏 , such that

|f (x) + F(x)𝜃 + g(x)u| ≤ M, (7a)

|𝜏 (⋅, ⋅)| ≤ M𝜏 , (7b)

| 𝜕Va 𝜕Va |
| | | ′|
| 𝜕x [f (x) + F(x)𝜃 + g(x)u] − 𝜕x [f (x ) + F(x )𝜃 + g(x )u]| ≤ (Lf + LF + Lg ) |x − x | .
′ ′ ′
(7c)
| |

3 AC LF-BASED MODEL PREDICTIVE CO NTRO L D ESIGN

3.1 Sample-and-hold implementation of adaptive controller

In this article, the adaptive controller k(⋅, ⋅) with the parameter update law (5a) will be applied in the design of the con-
straints in MPC, in which the optimized MPC output takes the form of piece-wise constant for the case that the state is
sampled discretely. Since the measurement of state x(t) during the interval of two sampling time instants is not available
̂ In this subsection, we give a piece-wise
to the controller, x(t) cannot be used directly in the parameter update law 𝜏(x, 𝜃).
̂
constant parameter update law of 𝜏 (x, 𝜃), then we present the stability of the system (1) under the adaptive control
′

̂ with the new parameter update law 𝜏 ′ (x, 𝜃).

u(t) = k(x, 𝜃) ̂
Let the piece-wise constant update law of the parameter and the adaptive control at time t, t ∈ [tk , tk+1 ), tk = t0 + kΔ,
k = 0, 1, 2..., where t0 is the initial time, and Δ is the sampling period, be
( )T
̇ = Γ𝜏 ′ (x(t), 𝜃(t))
̂𝜃(t) ̂ T 𝜕Va
= ΓF(x(tk )) , (8a)
𝜕x
̂ k )).
u(t) = k(x(tk ), 𝜃(t (8b)
 ZHENG et al.

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1258

̂ k−1 ), the estimation of 𝜃(t

Given 𝜃(t ̂ k ) equals to the sum of 𝜃(t
̂ k−1 ) and the integrating of (8a) from tk−1 to tk . Let

̂ = V1 (x) + 1 𝜃̂ T Γ−1 𝜃,
Va (x, 𝜃) ̂ (9a)
2
V(x, 𝜃) ̂ + 1 𝜃Γ
̂ = Va (x, 𝜃) ̃ −1 𝜃,
̃ (9b)
2

where 𝜃̃ = 𝜃 − 𝜃,
̂ V1 is a scalar function and
𝛼1 |x| ≤ V1 (x) ≤ 𝛼2 |x|, (10a)
| 𝜕V1 |
| |
| 𝜕x | ≤ 𝛼4 |x|, (10b)
| |

where 𝛼1 , 𝛼2 , and 𝛼4 are positive constants.

Remark 2. It should be noticed that, if the derivative of Va (x, 𝜃) ̂ is less than zero until x reaches equilibrium, and x = 0
̂ ̂
when the derivative of Va (x, 𝜃) equals zero, the state x will be stabilized at the equilibrium. It is not necessary that Va (x, 𝜃)
̂
eventually equals zero when the derivative of Va (x, 𝜃) equals zero.
̂ of system (1) under the action of the piece-wise control of (8) is presented as
Then, the decreasing property of V(x, 𝜃)
follows.
̂ if there exist a positive-definite 𝛼3 , a vector 𝜃(t
Theorem 1. Provided Va is an aCLF of (1) with u(t) = k(x, 𝜃), ̂ 0 ), and constants
𝜖w > 0, 𝜌 > 𝜌̃ > 0 satisfying
1
− (𝛼3 − 2𝛼4 MF M𝜏 |Γ|Δ − 𝛼4 LF M|Γ|Δ)𝛼2−1 𝜌̃ + (𝛼4 |Γ|Δ(2MF M𝜏 + LF M) + Lf M + Lg Mu )MΔ + 3LF MΔ(2|Γ|𝜌) 2
≤ −𝜖w ∕Δ, (11)

̂ ∈ Ω𝜌 and x ∉ S𝜌̃ , the Lyapunov function V(x, 𝜃)

then, for (x, 𝜃) ̂ is decreasing during [tk , tk+1 ) and V(tk + 1) < V(tk ) under
the piece-wise adaptive control (8b) whose parameter updates according to (8a).

Proof. Under the piece-wise control law and the new update law (8), the derivative of V can be rewritten as

V̇ = V̇ a − 𝜃̃ Γ−1 𝜃̂̇
T

𝜕Va ( )
̂ + 𝜕Va Γ𝜏 ′ (x, 𝜃)̂ − 𝜃̃ T 𝜏 ′ (x, 𝜃)
̂
= f (x) + F(x)𝜃 + g(x)k(x, 𝜃)
𝜕x 𝜕 ̂
𝜃
( ( )T ) ( ( )T )
𝜕Va ̂ 𝜕Va ̂ 𝜕Va ̃ 𝜕Va 𝜕V ̂ − 𝜃̃ T 𝜏 ′ (x, 𝜃)
̂
= f (x) + F(x)𝜃 + F(x)Γ + g(x)k(x, 𝜃) + F(x) 𝜃 − Γ + a Γ𝜏 ′ (x, 𝜃)
𝜕x 𝜕 𝜃̂ 𝜕x 𝜕 𝜃̂ 𝜕 𝜃̂
( ( )T ) ( )[ ( )T ]
𝜕Va 𝜕V 𝜕V 𝜕V
f (x) + F(x)𝜃̂ + F(x)Γ + g(x)k(tk ) + 𝜃̃ − ̂ ,
a T a a
= Γ F(x)T − 𝜏 ′ (x, 𝜃) (12)
𝜕x 𝜕 𝜃̂ 𝜕 𝜃̂ 𝜕x

̂ k )).
where k(tk ) = k(x(tk ), 𝜃(t
For t ∈ [tk , tk+1 ), according to (7a) and (7b), it can be obtained that
|x(t) − x(tk )| ≤ MΔ, (13)

|̂ ̂ k )|| ≤ |Γ|M𝜏 Δ.
|𝜃(t) − 𝜃(t (14)
| |

Considering that V̇ a (tk ) ≤ −𝛼3 (|x (tk )|) is satisfied for system (3), the definition of Va in (9a), and the properties of V1
in (10a) and (10b), the first term of the right side of (12) can be limited as
( ( )T )
𝜕Va 𝜕V
f (x) + F(x)𝜃̂ + F(x)Γ a
+ g(x)k(tk )
𝜕x 𝜕 𝜃̂
( ( )T )
𝜕Va 𝜕V
≤ f (x) + F(x)𝜃̂ + F(x)Γ a
+ g(x)k(tk ) − V̇ a (tk ) − 𝛼3 |x(tk )|
𝜕x 𝜕 𝜃̂
 ZHENG et al.

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1259

𝜕V 𝜕V 𝜕V 𝜕V
= −𝛼3 |x(tk )| + a f (x) − a (tk )f (x(tk )) + a F(x)𝜃̂ − a (tk )F(x(tk ))𝜃(t ̂ k)
𝜕x 𝜕x 𝜕x 𝜕x
( )T ( )T
𝜕Va 𝜕Va 𝜕Va 𝜕Va 𝜕V 𝜕V
+ F(x)Γ − (tk )F(x(tk ))Γ (tk ) + a g(x)k(tk ) − a (tk )g(x(tk ))k(tk )
𝜕x 𝜕𝜃̂ 𝜕x 𝜕𝜃̂ 𝜕x 𝜕x
𝜕V 𝜕V 𝜕V 𝜕V
= −𝛼3 |x(tk )| + a f (x) − a (tk )f (x(tk )) + 2 a F(x)𝜃̂ − 2 a (tk )F(x(tk ))𝜃(t ̂ k)
𝜕x 𝜕x 𝜕x 𝜕x
𝜕V 𝜕V
+ a g(x)k(tk ) − a (tk )g(x(tk ))k(tk )
𝜕x 𝜕x
𝜕Va 𝜕Va 𝜕V 𝜕V ̂ k ) + 2 𝜕Va F(x)𝜃(t
̂ k ) − 2 𝜕Va (tk )F(x(tk ))𝜃(t
= f (x) − (tk )f (x(tk )) + 2 a F(x)𝜃̂ − 2 a F(x)𝜃(t ̂ k)
𝜕x 𝜕x 𝜕x 𝜕x 𝜕x 𝜕x
𝜕V 𝜕V
+ a g(x)k(tk ) − a (tk )g(x(tk ))k(tk ). (15)
𝜕x 𝜕x

According to the Lipschitz conditions given in (7), the inequalities (10b) and (13), it has

𝜕Va 𝜕V
f (x) − a (tk )f (x(tk )) ≤ Lf MΔ, (16)
𝜕x 𝜕x
𝜕Va 𝜕V
F(x)𝜃̂ − a F(x)𝜃(t ̂ k ) ≤ 𝛼4 (|x(tk )| + MΔ)MF |Γ|M𝜏 Δ, (17)
𝜕x 𝜕x
𝜕Va 𝜕V
g(x)k(tk ) − a (tk )g(x(tk ))k(tk ) ≤ Lg Mu MΔ, (18)
𝜕x 𝜕x
𝜕Va 𝜕V
̂ k ) − a (tk )F(x(tk ))𝜃(t̂ k ) ≤ LF MΔ|𝜃(t
̂ k )|.
F(x)𝜃(t (19)
𝜕x 𝜕x

Therefore, from (15) to (19), it can be derived that

( ( )T )
𝜕Va 𝜕V
f (x) + F(x)𝜃̂ + F(x)Γ a
+ g(x)k(tk )
𝜕x 𝜕 𝜃̂
̂ k )| + (2𝛼4 MF M𝜏 |Γ|Δ + Lf M + Lg Mu )MΔ.
≤ (2𝛼4 MF M𝜏 |Γ|Δ − 𝛼3 )|x(tk )| + 2LF MΔ|𝜃(t (20)

The upper bound of the second term of (12) is deduced according to (7), which is

| ( )T | |
T 𝜕Va
|
|
|F(x(t)) ̂ k ))|| = | 𝜕Va (tk )F(x(tk )) − 𝜕Va F(x(t))|
− 𝜏(x(tk ), 𝜃(t
| 𝜕x | | |
| | | 𝜕x 𝜕x |
≤ LF MΔ, (21)

and
𝜕V
𝜃̃ − a Γ ≤ |𝜃̃ | + 𝛼4 (|x(k)| + ΔM)|Γ|.
T T
(22)
𝜕x

̂ given in (9b), and that the aCLF Va (x, 𝜃)

With consideration of the presentation of the Lyapunov function V(x, 𝜃) ̂ is a
positive-definite function, we can obtain that

1 ̃T ̃ ≤ V(x(t), 𝜃(t)).
̂
𝜃 (t)Γ−1 𝜃(t) (23)
2

Therefore, combining (12), (20)– (23), we can obtain that

V̇ ≤ (2𝛼4 MF M𝜏 |Γ|Δ − 𝛼3 )|x(tk )| + 2LF MΔ(2|Γ|V(t)) 2 + (2𝛼4 MF M𝜏 |Γ|Δ + Lf M + Lg Mu )MΔ

1

1
+ LF MΔ((2|Γ|V(t)) 2 + 𝛼4 (|x(k)| + ΔM)|Γ|)
= (2𝛼4 MF M𝜏 |Γ|Δ + 𝛼4 LF M|Γ|Δ − 𝛼3 )𝛼2−1 |x(tk )| + (𝛼4 |Γ|Δ(2MF M𝜏 + LF M) + Lf M + Lg Mu )MΔ
1
+ 3LF MΔ(2|Γ|V(t)) 2 . (24)
 ZHENG et al.

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1260

If 2𝛼4 MF M𝜏 |Γ|Δ + 𝛼4 LF M|Γ|Δ − 𝛼3 ≤ 0 which can be obtained from (11), it has

V̇ ≤ (2𝛼4 MF M𝜏 |Γ|Δ + 𝛼4 LF M|Γ|Δ − 𝛼3 )𝛼2−1 𝜌̃ + (𝛼4 |Γ|Δ(2MF M𝜏 + LF M) + Lf M + Lg Mu )MΔ

1
+ 3LF MΔ(2|Γ|𝜌) 2 . (25)

Thus, with the consideration of Equation (11), we can obtain

̇
V(x(t), ̂
𝜃(t)) ≤ −𝜖w ∕Δ. (26)

Integrating (26) from tk to tk+1 , we obtain that

̂ k+1 )) ≤ V(x(tk ), 𝜃(t

V(x(tk+1 ), 𝜃(t ̂ k )) − 𝜖w . (27)

This completes the proof. ▪

Remark 3. Theorem 1 can be fairly well established if the sampling time Δ is small enough and the aCLF Va is reasonably
selected.
Remark 4. When the ranges of 𝜃(t ̂ 0 ) and 𝜃(t
̃ 0 ) are given, which will be discussed in Section 3.3, then the range of Va (t0 ) can
be estimated according to (9). The boundary of the domain  in Definition 1 can be estimated according to the boundary
of Va (t0 ) and (9). Consequently, the Ω𝜌 can be calculated according to  and the limitation of inputs and states, the details
can refer to References 29 and 19.

3.2 MPC formulation

At the time instant tk , the following MPC optimization problem is optimized to obtain the future control input u based
on the states measurement x(tk ). To guarantee the convergence of the states of the closed-loop system, an aCLF-based
̂ presented in (8b) is employed to limit the inputs of the sys-
constraint is designed, where the adaptive controller k(x, 𝜃)
tem, then to guarantee the decreasing of the aCLF of the closed-loop system. The optimization problem of the proposed
aCLF-based MPC for system (1) is as follows

tk+N
min x̃ T (t) Q̃x (t) + uT (t) Ru (t) dt, (28a)
u∈S(Δ)∫t
k

s.t. x̃̇ (t) = f (̃x (t)) + F (̃x (t)) 𝜃 + g (̃x (t)) u (t) , (28b)

u (t) ∈  , (28c)

x̃ (tk ) = x (tk ) , (28d)

̂
𝜕Va (x, 𝜃) ̂
𝜕Va (x, 𝜃)
g (x(tk )) u (tk ) ≤ ̂ k )),
g (x (tk )) k(x(tk ), 𝜃(t (28e)
𝜕x 𝜕x

where S(Δ) is the family of piece-wise constant functions. N is the prediction horizon, Q and R are positive-definite
weight matrices applied in the cost, x̃ is the predicted trajectory, and 𝜃 is the parameter vector used in the predic-
tion model (28b). The prediction model (28b) can be updated by an arbitrary estimation method. Equation (28e)
enforces the aCLF of the closed-loop system under MPC to be less than that under the auxiliary controller. Corre-
̂ = Va (x, 𝜃)
spondingly, the V(x, 𝜃) ̂ + 1 𝜃Γ
̃ −1 𝜃̃ under MPC is less than that under the auxiliary controller, which guar-
2
antees the decreasing of V(x, 𝜃)̂ respect to the closed-loop MPC system. It should be noticed that both the aCLF
Va and V are the functions of x and 𝜃, and V(x, 𝜃) ̂ = V1 (x) + 1 𝜃̂ T Γ−1 𝜃̂ + 1 𝜃Γ
̃ −1 𝜃̃ according to (9). When the sum of
2 2
the last two terms of the above equation decrease, the V1 (x) can increase. It does not require that the Lyapunov
function for the real system decreases at each time step, which provides a possibility for the relaxing of stability
constraints.
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ZHENG et al. 1261

Denote the optimal solution to this problem by u∗ (t|tk ) for t ∈ [tk , tk+N ). At each control period, the following steps
are carried out by the proposed MPCs.

1. At each time instant tk , update the parameter of 𝜃 in the prediction (28b);

2. Calculate the 𝜃(t̂ k ) according to (8a), and k(x(tk ), 𝜃(t
̂ k )) according to (8b);
∗
3. Solve the problem (28) to get the optimal u (t|tk );
4. Sent the u∗ (tk |tk ) to the actuator of the controlled process during [tk , tk+1 );
5. Let tk ← tk+1 , and repeat the above steps in the next control period.
Remark 5. In the implementation, since the model of the real system is unknown, to obtain better performance, some
machine-learning-based modeling methods, for example, recurrent neural network, Gaussian process and so forth can
be used to learn the dynamics of the system online and replace (28b) for more accurate prediction.

3.3 Convergence

Theorem 2 shows that the aCLF-based MPC guarantees the convergence of the closed-loop system.
Theorem 2. Consider the system (1) in closed-loop under the aCLFs-based MPC designed by (28) where the auxiliary
̂ is given by (8). Let 𝜌 > 𝜌0 > 𝜌̃ > 0 satisfy
controller k(x, 𝜃)
1
𝜌0 ≤ 𝜌 − 𝜆−1 (Γ)𝛿𝜃2 , (29)
2 min

̃ 0 )| ≤ 𝛿𝜃 , then for all x (t0 ) ∈ S𝜌 , the states of the closed-loop system will
and (11), where 𝛿𝜃 is constant and satisfies |𝜃(t 0
converge to S𝜌̃ .

̂ k )) ∈ Ω𝜌 , which holds until x enters S𝜌̃ and will be proved below, the adaptive control law is a feasible
Proof. If (x(tk ), 𝜃(t
solution of MPC since it satisfies all constraints in the optimization problem (28).
̂ 0 ) such that (x(t0 ), 𝜃(t
When x(t0 ) ∈ S𝜌0 ∕S𝜌̃ , by the definitions of V1 and Va , we can always select a 𝜃(t ̂ 0 )) ∈ Ω𝜌 ⊂ Ω𝜌 ,
0
̂ 0 ) = 0).
(i.e., let 𝜃(t
According to the constraint (28e), it has that the decrease rate of Va under the MPC is less than that under the adaptive
controller, which is
V̇ a (k) ≤ 𝛼3 |xk |. (30)

Therefore, according to Theorem 1, the following inequality is always satisfied under the proposed MPC when (x, 𝜃) ∈ Ω𝜌
and x ∉ S𝜌̃ ,
V(tk ) ≤ V(tk+1 ). (31)

Since V(t0 ) ≤ 𝜌 obtained from (29) and (x(t0 ), 𝜃(t ̂ 0 )) ∈ Ω𝜌 , it has V(t1 ) ≤ 𝜌. Then, we have (x(t1 ), 𝜃(t1 )) ∈ Ω𝜌 .
0
Recursively using above process, the (x, 𝜃) will always remain in Ω𝜌 , until the states enter S𝜌̃ . This completes the
proof. ▪
{ }
Remark 6. Generally, the rough range of 𝜃 is given in practice. Let 𝜃 belong to a set Θ ∶= 𝜃 ∈ Rn𝜃 ∶ |𝜃 − 𝜃0 | ≤ 12 𝛿𝜃′ ,
̂ 0 ) to satisfy 𝜃(t
where 𝜃0 and 𝛿 ′ are constants. If we select 𝜃(t ̂ 0 ) ∈ Θ, it has |𝜃(t
̃ 0 )| = |𝜃 − 𝜃(t
̂ 0 )| ≤ 𝛿 ′ . Then the 𝛿𝜃 in (29) can
𝜃 𝜃
be obtained by letting 𝛿𝜃 ≥ 𝛿𝜃′ .
Remark 7. When x ∈ S𝜌̃ at time tk , the control algorithm can switch from the proposed MPC to a local control which
is designed according to the model (28b) with fixed parameter 𝜃 that has been recursively modified by real-time data
obtained during [t0 , tk ). After tk , in general, the accuracy of the model is improved. The local control can be the Sontag
control designed for (28b) with 𝜃(t) = 𝜃(tk ), which will lead to an invariant set subject to the input constraints. The more
accurate the parameters are, the smaller the size of the smallest invariant set can be. When x enters S𝜌̃ which belongs to
the attraction region of the state under the local control, the states will enter and remain within the smallest invariant set
corresponding to the local control.19 Compared to the method where the constraints are designed according to the initial
boundary of the model-plant mismatch, the advantage of the proposed MPC problem (28) is that it provides a potential
for performance improvement during the procedure of driving the state from an initial point to S𝜌̃ that can be very small
if the sampling interval is small enough.
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1262 ZHENG et al.

Remark 8. It should be pointed out that some obstacles are faced when deriving new results. Specifically, how to design the
auxiliary controller implemented in a sample-and-hold fashion, which can guarantee the convergence of the closed-loop
system. To overcome this challenge, the modified parameter updating law of (8a) is developed. Using (7c), the affection of
the sample-and-hold implementation of the controller is deduced. Then, Theorem 1 gives the sufficient condition under
which the closed-loop adaptive system will converge to a small region. The sufficient condition reflects the upper bound
of the affection of the sample-and-hold implementation to the Lyapunov function of the real system.

4 SIMULATION

The proposed method is used to control a plant composed of two continuously stirred tank reactors (CSTRs) and a flash
tank separator, as shown in Figure 1. The reactant A is fed into CSTRs, where it is converted into product B and the
desired product B can further react into side-product C. Feed streams feed into the first CSTR and the second CSTR with
the composition xA0 and the flowrate Ffi , i = 1, 2, respectively. The liquid in each tank will flow into the next tank with
flowrates of F1 and F2 , respectively. After the splitting of the distillate in the separator, the product with flowrate F3 leaves
the reaction system, and a small portion of the residue with flowrate FD is purged before being recycled to the first CSTR.
In this application, assume that reaction temperature Ti , i = 1, 2, 3, are well controlled, which are constants, by exter-
nal heat input Qi . With the notations and the values of the parameters defined in Table A1 in the Appendix, the model
for the process can be simplified as follows
( )
ẋ A1 = V1−1 Ff 1 xA0 + FR xAR − F1 xA1 − kA1 xA1 , (32a)

ẋ B1 = V1−1 (FR xBR − F1 xB1 ) + kA1 xA1 − kB1 xB1 , (32b)

( )
ẋ A2 = V2−1 F1 xA1 + Ff 2 xA0 − F2 xA2 − kA2 xA2 , (32c)

ẋ B2 = V2−1 (F1 xB1 − F2 xB2 ) + kA2 xA2 − kB2 xB2 , (32d)

ẋ A3 = V3−1 (F2 xA2 − (FR + FD ) xAR − F3 xA3 ) , (32e)

ẋ B3 = V3−1 (F2 xB2 − (FR + FD ) xBR − F3 xB3 ) , (32f)

where,
kAi = kA e−EA ∕RTi and kBi = kB e−EB ∕RTi ,

and the compositions of A, B, and C in the reflux stream from the flash tank separator can be modeled as

𝛼A xA3
xAR = ,
𝛼A xA3 + 𝛼B xB3 + 𝛼C xC3
𝛼A xB3
xBR = ,
𝛼A xA3 + 𝛼B xB3 + 𝛼C xC3
xCR = 1 − xAR − xBR ,

FIGURE 1 Illustration of the chemical process.

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ZHENG et al. 1263

0.05
0.3 ACLF-MPC
MPC1 0
MPC2
-0.05
0.2 ACLF-MPC
-0.1 MPC1

x1

x2
MPC2
0.1
-0.15

-0.2
0
-0.25
0 1 2 3 4 5 0 1 2 3 4 5
Time(s) Time(s)

ACLF-MPC 0
0.2 MPC1
MPC2
0.15
ACLF-MPC
-0.1 MPC1
0.1
x3

x4
MPC2
0.05
-0.2
0

-0.05 -0.3
0 1 2 3 4 5 0 1 2 3 4 5
Time(s) Time(s)

ACLF-MPC 0
0.2 MPC1
MPC2
0.15
-0.1
0.1 ACLF-MPC
x5

x6

MPC1
0.05 MPC2
-0.2
0

-0.05 -0.3
0 1 2 3 4 5 0 1 2 3 4 5
Time(s) Time(s)

FIGURE 2 The evolution of state x

where xC3 = 1 − xA3 − xB3 . xAi and xBi are the composition of reactant A and B in the ith vessel. Vi , i = 1, 2, 3 is the volume of
each vessel. FR the flowrate of the recycling residue. 𝛼s , s ∈ {A, B, B} is the relative volatility of s. ka and Ea ∕R, a ∈ {A, B},
are the pre-exponential value and normal actor energy for the reaction a.
Denote the state vector with x = [x1 , x2 , x3 , x4 , x5 , x6 ]T = [xA1 − xA1
s
, xB1 − xB1
s
, xA2 − xA2
s
, xB2 − xB2
s
, xA3 − xA3s
, xB3 − xB3
s T
]
and the inputs as u = [u1 , u2 ] = [Ff 1 − Ff 1 , Ff 2 − Ff 2 ] with −5 kg/s ≤ u1 , u2 ≤ 10 kg/s.
T s s T

The measurements of states are available at each time instants tk with sampling time Δ = 0.1 s. The parameter vector
𝜃 = [kA1 , kB1 , kA2 , kB2 ]T is unknown. The control objective is to steer the states to origin.
The prediction horizon of MPC is chosen to be N = 10 and the weight matrices for MPC are chosen as Q =
diag(50, 10, 50, 10, 10, 10) and R = diag(10−4 , 10−4 ). Set V1 in (9a) be V1 (t) = 12 xT P1 x with P1 = diag(3.6, 0.16, 1.6, 0.16,
1.6, 0.16), and set Γ = diag(0.5, 0.5, 0.5, 0.5).
̂ 0 ) = 𝜃(t0 ) = [4.5, 4, 4.5, 4]T , and the state x(t0 ) to be [0.290, −0.210, 0.206, −0.203, 0.215, −0.251]T at the
Initialize 𝜃(t
start time t0 . The parameter 𝜃̂ in constraint (28e) updates according to (8a), and the Sontag control designed for system
(3) where the functions of f , g, and F are specified by (32) is chosen as the auxiliary controller k(x(tk ), 𝜃(t ̂ k )). The extended
Kalman filter is used to estimate 𝜃 in (28b) online for the prediction of state trajectory used in the cost function.
To verify the effectiveness of the proposed method, we give the other two schemes for comparison: the MPC1 which
uses the classical LMPC control scheme and worked as a baseline, where no model mismatch is presented between the
models of MPC and the controlled system; the MPC2 where the stability constraint is designed based on the initial model
mismatch boundary and does not update online. The idea of MPC2 was presented in References 24 and 27.
The states’ trajectories are shown in Figure 2. It shows that these three methods can all achieve closed-loop stability.
Among them, MPC1 gets the best performance since there is no model error. Under the premise of parameter errors, the
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1264 ZHENG et al.

1 1

0 0

-1 -1

-2 -2
u1

u2
-3 -3
ACLF-MPC ACLF-MPC
-4 MPC1 -4 MPC1
MPC2 MPC2
-5 -5

-6 -6
0 1 2 3 4 5 0 1 2 3 4 5
Time(s) Time(s)

FIGURE 3 Manipulated inputs u1 and u2

4.5 4

4 3.5

3.5 3

3 2.5
0 1 2 3 4 5 0 1 2 3 4 5
Time(s) Time(s)
4
4.4
3.5
4.3
4.2
3
4.1
4 2.5
0 1 2 3 4 5 0 1 2 3 4 5
Time(s) Time(s)

FIGURE 4 The evolution of 𝜃̂

convergence speed of the proposed MPC is slightly lower than the baseline obtained by MPC1 where there is no model
mismatch in the predictive model. Compared with MPC2, the convergence speed of our proposed scheme is faster under
the same model mismatch at the start time. Figure 3 shows the input trajectories under the three methods, respectively.
Figure 4 gives the trajectories of 𝜃̂ under the given parameters update law. It shows that all the elements of 𝜃̂ converge
̂ the states are convergent under the proposed method as
to constants after finite time steps. With the convergence of 𝜃,
shown in Figure 2.
t ( )
The costs J(t) = ∫𝜏=t xT (𝜏)Qx(𝜏) + uT (𝜏)Ru(𝜏) d𝜏 of the closed-loop systems in the three cases are 1.76, 1.44, and 2.19,
0
respectively. Taking the performance of MPC1 as a baseline, the ratios of performance of the proposed method and that of
MPC2 to the baseline are 1.76∕1.44 = 1.22 and 2.19∕1.44 = 1.52, respectively. Compared with the performance obtained
by MPC2 which only updates the predictive model, the performance of the proposed method is closer to the baseline
than that of MPC2. There is only a decrease of 22% from the baseline under the proposed method, which is caused by the
model mismatch at the start time. It improves 32% of the baseline than that of the MPC2. This improved performance
illustrates that the proposed constraint update design provides a possibility for the improvement of performance.

5 CO N C LU S I O N

An aCLF-based MPC is developed for a class of nonlinear systems whose accurate models are unknown at the starting
time instant. A convergence constraint, which updates online according to the newly obtained data, is designed based on
the aCLF technique. The online updating constraint can guarantee that the state of the closed-loop system is driven to a
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ZHENG et al. 1265

small neighborhood of the equilibrium if the provided sufficient conditions are satisfied. An arbitrary estimation method
that can accurately predict the future state trajectory with new data arriving can be employed to further improve the per-
formance of the closed-loop system. The designed aCLF-based MPC provides a potential for the operational performance
improvement of a process. The simulation implements the proposed method to a chemical process, which illustrates the
effectiveness of the proposed method.

CONFLICT OF INTEREST
The authors declare no potential conflict of interests.

FUNDING INFORMATION
This work was supported by National Key R&D Program of China (2018AAA0101701); National Natural Science
Foundation of China (62073220, 61833012).

DATA AVAILABILITY STATEMENT

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.

ORCID
Yi Zheng https://orcid.org/0000-0001-5141-0729
Shaoyuan Li https://orcid.org/0000-0003-3427-2912

REFERENCES
1. Zhang Y, Huang Y, Chen Z, Li G, Liu Y. A novel learning based model predictive control strategy for plug-in hybrid electric vehicle. IEEE
Trans Transp Electr. 2021.
2. Li Z, Huang M, Zhu J, Gui W, Jiang ZP, Du W. Learning-based adaptive optimal control for flotation processes subject to input constraints.
IEEE Trans Control Syst Technol. 2022.
3. Åström KJ, Wittenmark B. Adaptive Control. Courier Corporation; 2013.
4. Hoseini SM, Farrokhi M, Koshkouei AJ. Robust adaptive control of nonlinear non-minimum phase systems with uncertainties. Automat-
ica. 2011;47(2):348-357.
5. Afram A, Janabi-Sharifi F, Fung AS, Raahemifar K. Artificial neural network (ANN) based model predictive control (MPC) and
optimization of HVAC systems: a state of the art review and case study of a residential HVAC system. Energ Build. 2017;141:96-113.
6. Maiworm M, Limon D, Findeisen R. Online learning-based model predictive control with Gaussian process models and stability
guarantees. Int J Robust Nonlinear Control. 2021;31(18):8785-8812.
7. Fisac JF, Akametalu AK, Zeilinger MN, Kaynama S, Gillula J, Tomlin CJ. A general safety framework for learning-based control in
uncertain robotic systems. IEEE Trans Autom Control. 2018;64(7):2737-2752.
8. Liu L, Gao T, Liu YJ, Tong S, Chen CLP, Ma L. Time-varying IBLFs-based adaptive control of uncertain nonlinear systems with full state
constraints. Automatica. 2021;129:109595.
9. Krstić M, Kokotović PV. Control Lyapunov functions for adaptive nonlinear stabilization. Syst Control Lett. 1995;26(1):17-23.
10. Chen M, Ge SS, Ren B. Adaptive tracking control of uncertain MIMO nonlinear systems with input constraints. Automatica.
2011;47(3):452-465.
11. Zhao X, Shi P, Zheng X, Zhang L. Adaptive tracking control for switched stochastic nonlinear systems with unknown actuator dead-zone.
Automatica. 2015;60:193-200.
12. Christofides PD, Scattolini R, Muñoz de la Peña D, Liu J. Distributed model predictive control: a tutorial review and future research
directions. Comput Chem Eng. 2013;51:21-41.
13. Qin SJ, Badgwell TA. A survey of industrial model predictive control technology. Control Eng Pract. 2003;11(7):733-764.
14. Liu J, Muñoz de la Peña D, Christofides PD. Distributed model predictive control of nonlinear process systems. AICHE J.
2009;55(5):1171-1184.
15. Masero E, Frejo JRD, Maestre JM, Camacho EF. A light clustering model predictive control approach to maximize thermal power in solar
parabolic-trough plants. Sol Energy. 2021;214:531-541.
16. Chen H, Allgöwer F. A quasi-infinite horizon nonlinear model predictive control scheme with guaranteed stability. Automatica.
1998;34(10):1205-1217.
17. Mayne DQ, Kerrigan EC, Van Wyk E, Falugi P. Tube-based robust nonlinear model predictive control. Int J Robust Nonlinear Control.
2011;21(11):1341-1353.
18. Rawlings JB, Mayne DQ, Diehl M. Model Predictive Control: Theory, Computation, and Design. Vol 2. Nob Hill Publishing Madison; 2017.
19. Mhaskar P, El-Farra NH, Christofides PD. Stabilization of nonlinear systems with state and control constraints using Lyapunov-based
predictive control. Syst Control Lett. 2006;55(8):650-659.
20. Wu Z, Albalawi F, Zhang Z, Zhang J, Durand H, Christofides PD. Control Lyapunov-Barrier function-based model predictive control of
nonlinear systems. Automatica. 2019;109:108508.
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

10991239, 2023, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/rnc.6409 by Technische Universitaet Brauns, Wiley Online Library on [18/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1266 ZHENG et al.

21. Hu J, Zhang H, Yu X, Liu H, Chen D. Design of sliding-mode-based control for nonlinear systems with mixed-delays and packet losses
under uncertain missing probability. IEEE Trans Syst Man Cybern Syst. 2021;51(5):3217-3228.
22. Ding B. Robust model predictive control for multiple time delay systems with polytopic uncertainty description. Int J Control.
2010;83(9):1844-1857. doi:10.1080/00207179.2010.498058
23. Li H, Jin B, Yan W. Distributed model predictive control for linear systems under communication noise: algorithm, theory and
implementation. Automatica. 2021;125:109422. https://www.sciencedirect.com/science/article/pii/S0005109820306245
24. Aswani A, Gonzalez H, Sastry SS, Tomlin C. Provably safe and robust learning-based model predictive control. Automatica.
2013;49:1216-1226.
25. Lorenzen M, Cannon M, Allgöwer F. Robust MPC with recursive model update. Automatica. 2019;103:461-471.
26. Zhu B, Xia X. Lyapunov-based adaptive model predictive control for unconstrained non-linear systems with parametric uncertainties. IET
Control Theory Appl. 2016;10:1937-1943.
27. Wan R, Li S, Zheng Y. Model predictive control for nonlinear systems with time-varying dynamics and guaranteed Lyapunov stability. Int
J Robust Nonlinear Control. 2021;31(2).
28. Wu Z, Rincon D, Christofides PD. Real-time adaptive machine-learning-based predictive control of nonlinear processes. Ind Eng Chem
Res. 2020;59(6):2275-2290.
29. Zheng Y, Li S, Wan R, Wang Y. Economic Lyapunov-based model predictive control with event-triggered parametric identification. Int
J Robust Nonlinear Control. 2022;32(1):205-226.
30. Lin Y, Sontag ED. A universal formula for stabilization with bounded controls. Syst Control Lett. 1991;16(6):393-397.

How to cite this article: Zheng Y, Wang Y, Li S. Adaptive control Lyapunov function based model predictive
control for continuous nonlinear systems. Int J Robust Nonlinear Control. 2023;33(2):1254-1266. doi:
10.1002/rnc.6409

APPENDIX

The notations and the values of the parameters of the system (32) are shown as follows.
T A B L E A1 Process variables14
Symbol Value Units
V1 Volume of vessel 1 7.5 m3
V2 Volume of vessel 2 7.5 m3
V3 Volume of vessel 3 5.0 m3
kA Pre-exponential values for reaction 1 0.760 1/s
kB Pre-exponential values for reaction 2 0.588 1/s
EA ∕R Normal actor energy for reaction 1 −100 K
EB ∕R Normal actor energy for reaction 2 −150 K
xA0 Mass fraction of A in external streams 1 wt(%)
𝛼A Relative volatility of A 3.5 -
𝛼B Relative volatility of B 1.1 -
𝛼C Relative volatility of C 0.5 -
s
xA1 Steady mass fraction of A in vessel 1 0.540 wt(%)
s
xB1 Steady mass fraction of B in vessel 1 0.293 wt(%)
s
xA2 Steady mass fraction of A in vessel 2 0.503 wt(%)
s
xB2 Steady mass fraction of B in vessel 2 0.298 wt(%)
s
xA3 Steady mass fraction of A in vessel 3 0.269 wt(%)
s
xB3 Steady mass fraction of B in vessel 3 0.367 wt(%)
T1 Temperature in vessel 1 450.3 K
T2 Temperature in vessel 2 447.6 K
T3 Temperature in vessel 3 453.5 K
Ffs1 Steady feed stream of vessel 1 5 kg/s
Ffs1 Steady feed stream of vessel 1 5 kg/s