ISA Transactions xxx (xxxx) xxx

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research article

Repetitive dynamic matrix control for systems with periodic specifications

Tito L.M. Santos a ,∗, Daniel M. Lima b , Julio E. Normey-Rico c
a
Departamento de Engenharia Elétrica e de Computação (DEEC), Universidade Federal da Bahia (UFBA), Rua Aristides Novis, 02, Federação, Salvador-BA, Brazil
b
Departamento de Controle, Automação e Computação, Universidade Federal de Santa Catarina (UFSC), Blumenau, SC, Brazil
c
Departamento de Automação Sistemas (DAS), Universidade Federal de Santa Catarina (UFSC), Bairro Trindade, Caixa Postal 476, Florianópolis-SC, Brazil

ARTICLE INFO ABSTRACT

Keywords: This paper proposes a Repetitive Dynamic Matrix Control (RDMC) for systems with periodic specifications. The
Dynamic matrix control new algorithm is able to track periodic references and reject repetitive disturbances with a known period based
Repetitive control on a modified prediction error. A repetitive version of the Generalized DMC (GDMC) is also proposed such
Constrained systems
that it can be applied to control open-loop unstable systems. Only the step-response coefficients are required to
Data-driven control
describe the dynamical system such that the RDMC preserves the modeling simplicity of the Dynamic Matrix
Control (DMC). The proposed solution can be interpreted as an extension of the DMC for repetitive control
applications. A data-driven filter design is proposed in order to ensure null prediction steady-state error in
the presence of periodic disturbances even for unstable open-loop systems. Two case studies are presented to
show the usefulness of the proposed strategy for control systems with periodic specification and to illustrate
the typical advantages and drawbacks of the proposed repetitive control extension of the DMC algorithm.

1. Introduction a gantry robotic system [21] recently, but parametric transfer function
models were also used. Thus, the related MPC literature for prob-
The Dynamic Matrix Control (DMC) algorithm [1] had a significant lems with periodic specifications mainly depends on the knowledge
impact on the process control industry due to its modeling simplicity of parametric transfer functions or state-space models. Some efforts
and its ability to deal with constraints and multivariable systems [2]. have been made to simplify the identification stage of these parametric
One of the main properties of the DMC is the convolutional prediction models, as in [12,13], where a subspace repetitive predictive control
model that can be defined directly from the step response coeffi-
has been proposed. Although these solutions use a simpler state-space
cients [1–3]. This useful property for practical application has been
identification approach, a parametric state-space model is still required.
attracting the academy’s attention even to this day [4–9].
Despite the relevance of the DMC algorithm and its variations
Repetitive model predictive control is also an important research
field as the output of control systems with periodic targets and pe- in practical applications, thanks to its ability to handle constraints
riodic disturbances can achieve reference tracking without error in and multivariable systems and the usefulness of repetitive predictive
steady-state operation [10,11]. Repetitive predictive strategies have controllers to deal with periodic disturbance and targets, repetitive
been tested in several control applications such as wind tunnels [12], DMC was not presented in the literature to the best of the authors’
wind turbines [13], moving bed chromatography system [14], control knowledge. Indeed, the original DMC FIR (finite impulse response)
of a reverse flow reactor [15], grid-connected inverter [16], elec- prediction model would be based on non-convergent coefficients due
trical machine drives [17]. However, these related strategies were to the repetitive poles in such a way that the model could not be
defined from parametric descriptions such as transfer function [10] finitely approximated. This paper proposes a prediction model correc-
and state-space models [10–13,16,17]. Furthermore, if minimal state- tion to deal with the periodic evolution of the signals in a Repetitive
space representation is used and the state vector is not measurable, an DMC (RDMC) based on a modified truncated FIR prediction model.
observer should be designed in order to apply the repetitive control Motivated by [9], a filtered version of the RDMC is also proposed in
concept [18]. A cost function modification has been recently proposed
order to stabilize the prediction for integrative and unstable systems.
in [19] for the Generalized Predictive Control (GPC) and the DMC to
The proposed RDMC strategies only require the knowledge of the
deal with time-varying references, but this strategy is not able to reject
coefficients of the step response of the process. A relay-type test can be
periodic disturbances. The relevance and usefulness of repetitive con-
trollers have been experimentally illustrated in a robotic arm [20] and used to recover the step response data of unstable systems. Moreover,

∗ Corresponding author.
E-mail addresses: [email protected] (T.L.M. Santos), [email protected] (D.M. Lima), [email protected] (J.E. Normey-Rico).

https://doi.org/10.1016/j.isatra.2024.04.034
Received 14 October 2023; Received in revised form 29 April 2024; Accepted 29 April 2024
Available online 3 May 2024
0019-0578/© 2024 ISA. Published by Elsevier Ltd. All rights reserved.

Please cite this article as: Tito L.M. Santos et al., ISA Transactions , https://doi.org/10.1016/j.isatra.2024.04.034
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

a data-driven filter design based on a Quadratic Programming (QP) s.t.

optimization problem is presented where the prediction stabilization ∑
𝑗 ∑
𝑀
and the repetitive property are ensured directly from the step-response ̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑔𝑖 𝛥𝑢(𝑘 + 𝑗 − 𝑖) + (𝑔𝑖+𝑗 − 𝑔𝑖 )𝛥𝑢(𝑘 − 𝑖) + 𝑦(𝑘), (4b)
information. The proposed prediction model is able to guarantee null 𝑖=1 𝑖=1
prediction error in the presence of repetitive disturbances. Therefore, 𝑢(𝑘 + 𝑗|𝑘) = 𝛥𝑢(𝑘 + 𝑗|𝑘) + 𝑢(𝑘 + 𝑗 − 1|𝑘), (4c)
despite the relevance of other repetitive model predictive control algo-
rithms, the proposed approach is the only one that is entirely based on ̂ + 𝑗|𝑘) ∈ Y,
𝑦(𝑘 𝑗 ∈ N[1,𝑁] , (4d)
the step response coefficients as in the standard DMC strategy. Two case 𝑢(𝑘 + 𝑗|𝑘) ∈ U, 𝑗 ∈ N[0,𝑁𝑢 −1] , (4e)
studies based on an unstable Continuous Stirred-Tank Reactor (CSTR)
reactor with periodic disturbances and the tip-tracking control of a 𝛥𝑢(𝑘 + 𝑗|𝑘) ∈ 𝛥U, 𝑗 ∈ N[0,𝑁𝑢 −1] , (4f)
two-link arm with periodic reference are presented to illustrate the where 𝑦𝑟 (𝑘 + 𝑗) is the reference, 𝜆 > 0 is a design parameter that deals
usefulness of the proposed RDMC. with the trade-off between tracking error and control effort, 𝛥𝐮(𝑘) is the
The main contributions of this paper are:
vector of future control increments, and U, 𝛥U and Y are respectively
• A new DMC-based repetitive model predictive control strategy the input, the control increment and the output constraints. These
(RDMC) that is able to reject (to track) repetitive disturbances constraints have been defined for simplicity, but any linear constraint
(references) with null error in closed-loop; in the form 𝐴𝑐 𝛥𝐮 ≤ 𝑏𝑐 can be used. Moreover, relaxed constraints with
• A more general solution for repetitive model predictive control, slack variables can be considered as well.
the generalized RDMC, which can be used with unstable and The simplicity of the step-response identification approach is a
integrative processes; crucial advantage of the DMC algorithm and modified versions. On
• The establishment of formal results to show disturbance rejection the other hand, the DMC prediction model may provide a slower
and tracking properties of the RDMC; disturbance rejection if compared with other MPC strategies. Moreover,
• A data-driven approach to tune the proposed controllers; it cannot be applied to control open-loop unstable processes [4]. The
• The application of the proposed strategies to two nonlinear sim- filtered DMC [4] is able to deal with the trade-off between disturbance
ulated models and to an experimental laboratory setup. rejection and robustness, and the generalized DMC (GDMC) [9] has
been recently proposed to stabilize the prediction model of a filtered
To the best of the authors’ knowledge, this DMC-based MPC for systems DMC for open-loop unstable systems. Motivated by these recent results,
with periodic references and disturbances has not been addressed in the a modified DMC for repetitive control is presented in this work. The
literature. Furthermore, the repetitive DMC strategy is extended to deal
main objective is to achieve periodic reference tracking for stable,
with unstable and integrative systems, which is also a novelty of this
integrative, and unstable systems with repetitive disturbances.
work.
The rest of this paper is organized as follows. Section 2 presents a
Remark 1. The standard DMC is derived from a constant prediction
brief review of the main characteristics of the DMC strategy. Sections 3
error assumption (𝜂(𝑘 + 𝑗|𝑘) = 𝜂(𝑘) = 𝑦(𝑘) − 𝑦(𝑘)) ̄ with a typical
and 4 introduce the proposed control strategies and present the formal
control increment formulation (𝛥𝑢(𝑘 − 𝑖) = 𝑢(𝑘 − 𝑖) − 𝑢(𝑘 − 𝑖 − 1)) in
results. Finally, the simulation and experimental case studies and the
conclusions are respectively presented in Sections 5 and 6. order to include an integral action into the prediction model. However,
periodic disturbances require a repetitive prediction error to describe
2. Preliminaries the disturbance effect, while repetitive poles should be included in the
prediction model to achieve periodic control specifications due to the
The original DMC algorithm [1,3] is briefly revisited in this section. internal model principle.
Consider a Single-Input Single-Output (SISO) problem with no loss of
generality. The nominal convolutional step response model of the plant, Remark 2. It is assumed the system is subject to periodic disturbances
used in DMC to compute the prediction of the process output at time with a known discrete-time period given by 𝑁𝑟 .
𝑘 + 𝑗, is defined by
The following notation is used in this work: (i) 𝑦(𝑘 ̄ + 𝑗) is the
∑
∞
nominal prediction for 𝑦(𝑘 + 𝑗); (ii) 𝑦(𝑘 ̂ + 𝑗|𝑘) represents the corrected
̄ + 𝑗) =
𝑦(𝑘 𝑔𝑖 𝛥𝑢(𝑘 + 𝑗 − 𝑖), (1)
𝑖=1 prediction for 𝑦(𝑘+𝑗) computed at time 𝑘; (iii) 𝑦𝑟 (𝑘) defines the set-point
for 𝑦(𝑘); (iv) 𝜂(𝑘) and 𝜂𝑓 (𝑘) are respectively the prediction error and
where 𝛥𝑢(𝑘 − 𝑖) = 𝑢(𝑘 − 𝑖) − 𝑢(𝑘 − 𝑖 − 1) and 𝑔𝑖 , 𝑖 = 1, 2, … , ∞ represent the
step-response coefficients. In order to deal with the prediction error, filtered prediction error; (v) 𝑔𝑖 and ℎ𝑖 describe respectively the step-
a constant output prediction error model is given by 𝜂(𝑘 + 𝑗|𝑘) = response and impulse-response coefficients of the process model; (vi)
𝜂(𝑘|𝑘) = 𝑦(𝑘) − 𝑦(𝑘),
̄ 𝑗 = 0, 1, … , 𝑁, where 𝑁 is the prediction horizon, 𝑀 is the truncation parameter of the DMC convolutional model; (vii)
̄ + 𝑗) + 𝜂(𝑘 + 𝑗|𝑘) is the corrected prediction and 𝑦(𝑘) is the
̂ + 𝑗|𝑘) = 𝑦(𝑘
𝑦(𝑘 𝑁 and 𝑁𝑢 are the prediction and control horizons; (viii) 𝜆 > 0 is the
measured output of the process. Therefore, 𝑦(𝑘 ̂ + 𝑗|𝑘) can be rewritten DMC control effort weight; (ix) 𝛥𝑢(𝑘) represents a control increment,
as follows i.e. 𝛥𝑢(𝑘) = 𝑢(𝑘) − 𝑢(𝑘 − 1); (x) 𝛥𝐮(𝑘) = [𝛥𝑢(𝑘) … 𝛥𝑢(𝑘 + 𝑁𝑢 − 1)]⊤
is the vector with future control increments; (xi) 𝛥𝑟 𝑢(𝑘) represents the
∑
𝑗 ∑
∞
̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑔𝑖 𝛥𝑢(𝑘 + 𝑗 − 𝑖) + (𝑔𝑖+𝑗 − 𝑔𝑖 )𝛥𝑢(𝑘 − 𝑖) + 𝑦(𝑘). (2) repetitive control increment with period 𝑁𝑟 , i.e. 𝛥𝑟 𝑢(𝑘) = 𝑢(𝑘)−𝑢(𝑘−𝑁𝑟 );
𝑖=1 𝑖=1 (xii) 𝑞𝑖 are the repetitive coefficients of the convolutional model with
If the system is open-loop stable, then a truncation parameter 𝑀 can respect to the input 𝛥𝑟 𝑢(𝑘); (xiii) {ℎ𝑖 } is an infinite sequence defined by
be defined such that (𝑔𝑖+𝑗 − 𝑔𝑖 ) ≈ 0 for 𝑖 ≥ 𝑀 given 𝑗 ≥ 0 as discussed ℎ0 , ℎ1 , ℎ2 , . . . ; (xiv) (𝑘) = [𝑦(𝑘
̂ + 1|𝑘) … 𝑦(𝑘̂ + 𝑁|𝑘)]⊤ and 𝛥𝑟 𝐮(𝑘) =
𝑟 𝑟 ⊤
[𝛥 𝑢(𝑘) … 𝛥 𝑢(𝑘 + 𝑁𝑢 − 1)] are vectors with the future predictions and
in [22, Chapter 3]. Thus, the DMC prediction model is defined by
future repetitive control increments; (xv) 𝐘𝑁𝑟 (𝑘) and 𝛥𝑟  (𝑘) are vectors
∑
𝑗 ∑
𝑀
̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑔𝑖 𝛥𝑢(𝑘 + 𝑗 − 𝑖) + (𝑔𝑖+𝑗 − 𝑔𝑖 )𝛥𝑢(𝑘 − 𝑖) + 𝑦(𝑘). (3) with the past outputs and past repetitive control increments for free
𝑖=1 𝑖=1 response computation; (xvi) 𝐹𝑗 (𝑧) represents a suitable filter to be used
The DMC optimization problem can be expressed as follows, where for future predictions at time 𝑘 + 𝑗; (xvii) 𝑞̃𝑖,𝑗 describes the repetitive
𝑁𝑢 is the control horizon coefficients of the convolutional model filtered by 𝐹𝑗 (𝑧); (xviii) 𝑦̃𝑗 (𝑘)
[𝑁 ] represents the output 𝑦(𝑘) filtered by 𝐹𝑗 (𝑧); (xix) 𝐘(𝑘)̃ defines a vector
𝑁𝑢 −1
∑ ∑ of past filtered values of 𝑦(𝑘) with 𝐹𝑗 (𝑧) for free response computation;
min ̂ + 𝑗|𝑘)]2 +
[𝑦𝑟 (𝑘 + 𝑗) − 𝑦(𝑘 𝜆𝛥𝑢(𝑘 + 𝑗|𝑘)2 (4a)
𝛥𝐮(𝑘)
𝑗=1 𝑗=0
(xx) 𝐐 is a suitable matrix for forced response computation defined

2
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

from 𝑞𝑗 ; (xxi)  is a suitable matrix for free response computation Firstly, by separating the sum into two terms, that are dependent on
defined from 𝑞𝑗 ; (xxii)  describes the optimal RDMC gain, i.e.  = the future and past control actions, and by combining the result with
[1 0 … 0](𝐐⊤ 𝐐 + 𝜆𝐼)−1 𝐐⊤ ; (xxiii) ̃ is a suitable matrix for the filtered Eq. (7), the following corrected prediction is obtained
free response computation defined from 𝑞𝑗 and 𝑞̃𝑖 ; (xxiv) 𝑁𝑦 (𝑧) is the
𝑁𝑦 (𝑧) ∑
𝑗 ∑
∞
polynomial numerator of 𝑌 (𝑧) = 𝐷𝑦 (𝑧)
; (xxv) 𝐷𝑦 (𝑧) = 𝐷𝑦+ (𝑧)𝐷𝑦− (𝑧) is the ̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑞𝑖 𝛥𝑟 𝑢(𝑘 + 𝑗 − 𝑖) + 𝑞𝑖 𝛥𝑟 𝑢(𝑘 + 𝑗 − 𝑖)
𝑁𝑦 (𝑧) 𝑖=1 𝑖=𝑗+1
polynomial denominator of 𝑌 (𝑧) = , and 𝐷𝑦+ (𝑧) has the poles
𝐷𝑦+ (𝑧)𝐷𝑦− (𝑧) ∑
∞
that are inside the unit circle (stable ones), and 𝐷𝑦− (𝑧) has the remaining − 𝑞𝑖 𝛥𝑟 𝑢(𝑘 − 𝑖 + 𝑗 − 𝑚𝑗 𝑁𝑟 ) + 𝑦(𝑘 + 𝑗 − 𝑚𝑗 𝑁𝑟 ) (9)
ones. 𝑖=1
∑
𝑗 ∑
∞

3. Repetitive DMC based on periodic error correction = 𝑞𝑖 𝛥𝑟 𝑢(𝑘 + 𝑗 − 𝑖) + 𝑞𝑖+𝑗 𝛥𝑟 𝑢(𝑘 − 𝑖)

𝑖=1 𝑖=1
∑
∞
This section presents a repetitive DMC (RDMC) that can be applied − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 𝛥𝑟 𝑢(𝑘 − 𝑖) + 𝑦(𝑘 + 𝑗 − 𝑚𝑗 𝑁𝑟 ), (10)
to control open-loop stable processes as the original DMC approach. 𝑖=1−𝑗+𝑚𝑗 𝑁𝑟
The RDMC for SISO systems is defined from a repetitive control law
where the first term of the right-hand side of the previous equation is
given by
the forced response and the other terms represent the free response.
𝛥𝑟 𝑢(𝑘) = 𝑢(𝑘) − 𝑢(𝑘 − 𝑁𝑟 ), (5) The former is the part of the prediction that is dependent on the future
control action variations, and the latter is dependent on the past inputs
where 𝑁𝑟 is the desired discrete-time repetitive control period. Then,
and outputs.
the nominal convolutional model is represented by
Now note that the following identity holds
∑
∞
−𝑗+𝑚𝑗 𝑁𝑟
𝑦(𝑘)
̄ = 𝑞𝑖 𝛥𝑟 𝑢(𝑘 − 𝑖), (6) ∑
∞ ∑ ∑
∞
𝑖=1 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 = 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 + 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 . (11)
𝑖=1 𝑖=1 𝑖=1−𝑗+𝑚𝑗 𝑁𝑟
where 𝛥𝑟 𝑢(𝑘 − 𝑖) = 𝑢(𝑘 − 𝑖) − 𝑢(𝑘 − 𝑖 − 𝑁𝑟 ) and 𝑞𝑖 are suitable repetitive
coefficients. As 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 = 0 if 𝑖+𝑗 −𝑚𝑗 𝑁𝑟 < 1 because the system has been assumed
to be strictly proper (𝑞𝑖 = 0, ∀𝑖 < 1) in this type of DMC convolutional
Remark 3. The coefficients 𝑞𝑖 can be obtained from the filtered step model, the first term on the right is null, and the following identity can
−1
response of the process. That is, a filter 𝑇𝑁𝑟 (𝑧) = (1−𝑧−𝑁𝑟) is used to filter be used in Eq. (10)
(1−𝑧 )
the step signal that is then applied to the process to obtain the filtered ∑
∞ ∑
∞
∑∞ −𝑖
step response. Notice that 𝐺(𝑧) = 𝑖=0 𝑔𝑖 𝑧 is defined in the  domain − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 𝛥𝑟 𝑢(𝑘 − 𝑖) = − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 𝛥𝑟 𝑢(𝑘 − 𝑖). (12)
from the step response coefficients. The 𝑞𝑖 coefficients for all 𝑖 ≥ 1 are 𝑖=1−𝑗+𝑚𝑗 𝑁𝑟 𝑖=1
∑
computed using 𝑄(𝑧) = 𝑇𝑁𝑟 (𝑧)𝐺(𝑧) where 𝑄(𝑧) = ∞ −𝑖
𝑖=0 𝑞𝑖 𝑧 . Therefore, Thus, the free response and the forced response can be written as
the RDMC requires only the step-response information, exactly as the follows
standard DMC, which preserves the simplicity of the original modeling.
∑
𝑗 ∑
∞
̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑞𝑖 𝛥𝑟 𝑢(𝑘 + 𝑗 − 𝑖) + (𝑞𝑖+𝑗 − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 )𝛥𝑟 𝑢(𝑘 − 𝑖)
Remark 4. Note that, in this filtered model, ℎ𝑖 = 𝑞𝑖 − 𝑞𝑖−𝑁𝑟 are the 𝑖=1 𝑖=1
coefficients of the impulse response by definition. ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
forced response free response
Therefore, the repetitive convolutional model is a generalization of + 𝑦(𝑘 + 𝑗 − 𝑚𝑗 𝑁𝑟 ) (13)
the standard DMC prediction model because the latter is recovered with ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟
𝑁𝑟 = 1. free response
In order to deal with repetitive disturbances, the periodic prediction As the infinite sum in Eq. (13) is unsuitable for practical applications
error of the RDMC is defined by and the system is assumed to be BIBO stable, then a truncation horizon
𝜂(𝑘 + 𝑗|𝑘) = 𝑦(𝑘 + 𝑗 − 𝑚𝑗 𝑁𝑟 ) − 𝑦(𝑘
̄ + 𝑗 − 𝑚𝑗 𝑁𝑟 ), 𝑗 = 0, 1, … , 𝑁, (7) 𝑀 > 0 is used such that 𝑞𝑖+𝑗 − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 ≈ 0 for all 𝑖 ≥ 𝑀. Actually, if
the system is open-loop stable, then 𝑖 → ∞ ⇒ ℎ𝑖 → 0. Now notice that
where 𝑚𝑗 ∈ N is the minimum positive integer such that 𝑚𝑗 𝑁𝑟 ≥ 𝑗 𝑞𝑖+𝑗 −𝑞𝑖+𝑗−𝑁𝑟 = ℎ𝑖+𝑗 , 𝑞𝑖+𝑗 −𝑞𝑖+𝑗−2𝑁𝑟 = (𝑞𝑖+𝑗 −𝑞𝑖+𝑗−𝑁𝑟 )+(𝑞𝑖+𝑗−𝑁𝑟 −𝑞𝑖+𝑗−2𝑁𝑟 ) =
and 𝑦(𝑘) is the measured output. The variable 𝑚𝑗 is defined in order ℎ𝑖+𝑗 + ℎ𝑖+𝑗−𝑁𝑟 . Then, by induction, the following relationship holds
to guarantee that 𝑦(𝑘 + 𝑗 − 𝑚𝑗 𝑁𝑟 ) is known even if 𝑁 > 𝑁𝑟 , that is,
𝑚𝑗 −1
for a fixed 𝑁𝑟 and 𝑗 > 0, it is ensured that 𝑗 − 𝑚𝑗 𝑁𝑟 ≤ 0, ∀𝑗, 𝑚𝑗 . If ∑
the prediction horizon is defined such that 𝑁 ≤ 𝑁𝑟 , then 𝑚𝑗 = 1, 𝑞𝑖+𝑗 − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 = ℎ𝑖+𝑗−𝓁𝑁𝑟 , (14)
𝓁=0
𝑗 = 1, … , 𝑁.
Therefore, the corrected prediction is given by which guarantees that 𝑞𝑖+𝑗 − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 ≈ 0 for all 𝑖 ≥ 𝑀, 𝑗 = 1, 2, … , 𝑁
for a suitable long truncation horizon 𝑀.
∑
∞
̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑞𝑖 𝛥𝑟 𝑢(𝑘 + 𝑗 − 𝑖) + 𝜂(𝑘 + 𝑗|𝑘). (8) Then, the prediction model of the repetitive DMC is given by
𝑖=1
∑
𝑗 ∑
𝑀
Analogously to the standard DMC, the predictions 𝑦(𝑘 ̂ + 𝑗|𝑘) are rewrit- ̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑞𝑖 𝛥𝑟 𝑢(𝑘 + 𝑗 − 𝑖) + (𝑞𝑖+𝑗 − 𝑞𝑖+𝑗−𝑚𝑗 𝑁𝑟 )𝛥𝑟 𝑢(𝑘 − 𝑖)
ten with two objectives: (i) to separate the effect of the future repetitive 𝑖=1 𝑖=1
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
actions (forced response) and the past repetitive actions (free response), forced response free response
and (ii) to achieve a convergent sequence condition such that the infi-
+ 𝑦(𝑘 + 𝑗 − 𝑚𝑗 𝑁𝑟 ) (15)
nite sum can be truncated for practical implementation. The following
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟
steps are used sequentially: (i) 𝜂(𝑘 + 𝑗|𝑘) is replaced by Eq. (7), (ii) free response
the effect of past and future action are separated, (iii) all the past
actions are rewritten with respect to 𝛥𝑟 𝑢(𝑘 − 𝑖) with 𝑖 ≥ 1, (iv) and the Therefore, the prediction model can be expressed by
null elements of 𝑞𝑖 (𝑖 ≥ 0) are discarded. Then, the RDMC predictive
(𝑘) =𝐐𝛥𝑟 𝐮(𝑘) + 𝛥𝑟  (𝑘) + 𝐘𝑁𝑟 (𝑘), (16a)
equation is obtained, which is analogous to the classical DMC one,
⊤
given by Eq. (3). (𝑘) =[𝑦(𝑘
̂ + 1|𝑘) … 𝑦(𝑘
̂ + 𝑁|𝑘)] , (16b)

3
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

𝛥𝑟 𝐮(𝑘) =[𝛥𝑟 𝑢(𝑘|𝑘) … 𝛥𝑟 𝑢(𝑘 + 𝑁𝑢 − 1|𝑘)]⊤ , (16c) where (𝑧𝑁𝑟 − 1) and 𝐷𝑒𝑛𝑗 (𝑧) have no common roots. This necessary
condition is important because the following polynomial identity holds
𝐘𝑁𝑟 (𝑘) =[𝑦(𝑘 + 1 − 𝑚1 𝑁𝑟 ) … 𝑦(𝑘 + 𝑁 − 𝑚𝑁 𝑁𝑟 )]⊤ , (16d)
𝑟 𝑟
𝛥  (𝑘) =[𝛥 𝑢(𝑘 − 1) … 𝛥 𝑢(𝑘 − 𝑀)] , 𝑟 ⊤
(16e) 𝑧𝑗 − 𝐹𝑗 (𝑧) = 𝐿𝑗 (𝑧)(1 − 𝑧−𝑁𝑟 ), (24)
𝑧𝑁𝑟 𝑁𝑢𝑚𝑗 (𝑧)
where 𝐐 and  are defined in Appendix A. where 𝐿𝑗 (𝑧) = is a stable filter as a consequence of the
𝐷𝑒𝑛𝑗 (𝑧)
The Repetitive DMC algorithm is formulated by the following opti- stability of 𝐹𝑗 (𝑧).
mization problem. Now, the convergence of the truncated prediction in Eq. (22) can
∑
𝑁 𝑁𝑢 −1
∑ be imposed by including an additional filter design condition, even for
min
𝑟
̂ + 𝑗|𝑘)]2 +
[𝑦𝑟 (𝑘 + 𝑗) − 𝑦(𝑘 𝜆𝛥𝑟 𝑢(𝑘 + 𝑗|𝑘)2 (17a) unstable linear systems.
𝛥 𝐮(𝑘)
𝑗=1 𝑗=0 Notice that
s.t. (𝑘) = 𝛥𝑟  (𝑘) + 𝐘𝑁𝑟 (𝑘) + 𝐐𝛥𝑟 𝐮(𝑘), (17b)
{𝑞𝑖+𝑗 − 𝑞̃𝑖,𝑗 } =−1 {𝐿𝑗 (𝑧)(𝑄(𝑧) − 𝑧−𝑁𝑟 𝑄(𝑧))} (25)
𝑢(𝑘 + 𝑗|𝑘) = 𝛥 𝑢(𝑘 + 𝑗|𝑘) + 𝑢(𝑘 + 𝑗 − 𝑁𝑟 |𝑘),
𝑟
(17c) −1
= {𝐿𝑗 (𝑧)𝐻(𝑧)}, (26)
̂ + 𝑗|𝑘) ∈ Y,
𝑦(𝑘 𝑗 ∈ N[1,𝑁] , (17d)
−𝑁 ∑∞ −𝑖 ∑∞ −𝑖
where 𝑄(𝑧) − 𝑧 𝑟 𝑄(𝑧) = 𝑖=0 (𝑞𝑖 − 𝑞𝑖−𝑁𝑟 )𝑧 = 𝑖=0 ℎ𝑖 𝑧 by the
𝑢(𝑘 + 𝑗|𝑘) ∈ U, 𝑗 ∈ N[0,𝑁𝑢 −1] , (17e) definition of 𝑞𝑖 . Alternatively, {𝑞𝑖 − 𝑞𝑖−𝑁𝑟 } = 𝐻(𝑧) where
𝛥𝑢(𝑘 + 𝑗|𝑘) ∈ 𝛥U, 𝑗 ∈ N[0,𝑁𝑢 −1] , (17f) ∑
∞
𝐻(𝑧) = ℎ𝑖 𝑧−𝑖 (27)
where, as in the DMC, 𝑁 represents the prediction horizon, 𝑁𝑢 is the 𝑖=0
control horizon, 𝑀 is the truncation horizon of the FIR prediction
is the  transform of the impulse response which defines the transfer
model, and 𝜆 > 0 is a design parameter that deals with the trade-off
function of the system by definition, but this model is not required to be
between the tracking error and the control variation aggressiveness.
known. Also, notice that the infinite sum defined by Eq. (27) converges
Notice that by definition 𝑢(𝑘 + 𝓁|𝑘) = 𝑢(𝑘 + 𝓁) if 𝓁 < 0.
to an equivalent rational transfer function with poles and zeros.
The control law is defined by
Then, from Eq. (26) and the Final Value Theorem, lim𝑖→∞ (𝑞𝑖+𝑗 −
𝑢(𝑘) = 𝛥𝑟 𝑢⋆ (𝑘|𝑘) + 𝑢(𝑘 − 𝑁𝑟 ), (18) 𝑞̃𝑖,𝑗 ) = 0 is achieved if the poles of 𝐿𝑗 (𝑧)𝐻(𝑧) are strictly inside the unit
circle and 𝐿𝑗 (1)𝐻(1) = 0. This condition is achieved if 𝐿𝑗 (𝑧) eliminates
where 𝛥𝑟 𝑢⋆ (𝑘|𝑘)
represents the optimal solution for given the 𝛥𝑟 𝑢(𝑘|𝑘) all the poles of 𝐻(𝑧) that are not strictly inside the unit circle. It is
optimization problem defined by Eq. (17). important to emphasize that the knowledge of 𝐻(𝑧) is not required
Despite the usefulness of the proposed solution, it cannot be applied because the convergence of 𝑞𝑖+𝑗 − 𝑞̃𝑖,𝑗 can be imposed from the step
to control integrative and unstable open-loop systems. Then, motivated response information.
by [9], a GDMC for repetitive control is presented in the next section.
Let 𝛼𝑢,1 , 𝛼𝑢,2 , . . . 𝛼𝑢,𝑚 be the roots of the denominator of 𝐻(𝑧) with
magnitude equal or greater than one, the convergence defined by
4. GDMC for repetitive control lim𝑖→∞ (𝑞𝑖+𝑗 − 𝑞̃𝑖,𝑗 ) = 0 is achieved if

In this section, the generalized DMC stabilizing concept is applied to (𝑧𝑁𝑟 − 1)(𝑧 − 𝛼𝑢,1 ) … (𝑧 − 𝛼𝑢,𝑚 )𝑁𝑢𝑚𝑗 (𝑧)
𝑧𝑗 − 𝐹𝑗 (𝑧) = , 𝑗 = 1, … , 𝑁. (28)
achieve a repetitive control law. The main contribution is the extension 𝐷𝑒𝑛𝑗 (𝑧)
of the repetitive DMC for unstable and integrative systems.
A filter candidate can be described by
The problem can be handled with a filtered error, which is defined
by 𝑛𝑁𝑟 +𝑚−1,𝑗 𝑧𝑁𝑟 +𝑚−1 + 𝑛𝑁𝑟 +𝑚−2,𝑗 𝑧𝑁𝑟 +𝑚−2 + ⋯ + 𝑛1,𝑗 𝑧 + 𝑛0,𝑗
𝐹𝑗 (𝑧) = , (29)
𝑧𝑁𝑟 −1 (𝑧 − 𝛽)𝑚+1
𝜂𝑓 (𝑘 + 𝑗|𝑘) = −1 {𝐹𝑗 (𝑧){𝑦(𝑘) − 𝑦(𝑘)}},
̄ (19)
where 0 < 𝛽 < 1 is a free design parameter, 𝑚 is the number of
𝐹𝑗 (𝑧) is a suitable stable causal filter. In this case, the filtered convolu- undesired roots (𝛼𝑢,1 , … , 𝛼𝑢,𝑚 ) and the parameters 𝑛𝑙,𝑗 are computed
tional model is defined from the following filtered variables1 such that condition Eq. (28) holds. Indeed, 𝛽 deals with the trade-off
{𝑞̃𝑖,𝑗 } =−1 {𝐹𝑗 (𝑧)𝑄(𝑧)}, (20) between robustness and disturbance rejection performance as in the
filtered DMC [4].
𝑦̃𝑗 (𝑘) =−1 {𝐹𝑗 (𝑧){𝑦(𝑘)}}. (21) Particular attention should be paid to the filter design in the case
The filtered convolutional model is defined from coefficients 𝑞̃𝑖,𝑗 that of unstable processes as the robustness margins can be degraded. The
are obtained from the filtered repetitive response of the process with a inherent robustness analysis can be performed as discussed in [9,
new filter 𝐹𝑗 (𝑧) = 𝑁𝑢𝑚𝑗 (𝑧)∕𝐷𝑒𝑛𝑗 (𝑧). Section 5.1].
Then, the filtered DMC prediction is simply Therefore, the filtered Repetitive DMC algorithm is formulated by
the following optimization problem.
∑
𝑗 ∑
𝑀
̂ + 𝑗|𝑘) =
𝑦(𝑘 𝑞𝑖 𝛥𝑟 𝑢(𝑘 + 𝑗 − 𝑖) + (𝑞𝑖+𝑗 − 𝑞̃𝑖,𝑗 )𝛥𝑟 𝑢(𝑘 − 𝑖) + 𝑦̃𝑗 (𝑘) . (22) ∑
𝑁 𝑁𝑢 −1
∑
𝑖=1 𝑖=1 min
𝑟
̂ + 𝑗|𝑘)]2 +
[𝑦𝑟 (𝑘 + 𝑗) − 𝑦(𝑘 𝜆𝛥𝑟 𝑢(𝑘 + 𝑗|𝑘)2 (30a)
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ 𝛥 𝐮(𝑘)
𝑗=1 𝑗=0
forced response free response
̃
s.t. (𝑘) = 𝛥𝑟  (𝑘) + 𝐘(𝑘) ̃ 𝑟 𝐮(𝑘),
+ 𝐐𝛥 (30b)
Therefore, a necessary condition to achieve the convergence of the
free response (𝑞𝑖+𝑗 − 𝑞̃𝑖,𝑗 ≅ 0, 𝑖 ≥ 𝑀, 𝑗 = 1, … , 𝑁) is 𝑢(𝑘 + 𝑗|𝑘) = 𝛥 𝑢(𝑘 + 𝑗|𝑘) + 𝑢(𝑘 + 𝑗 − 𝑁𝑟 |𝑘),
𝑟
(30c)

(𝑧𝑁𝑟 − 1)𝑁𝑢𝑚𝑗 (𝑧) ̂ + 𝑗|𝑘) ∈ Y,

𝑦(𝑘 𝑗 ∈ N[1,𝑁] , (30d)
𝑧𝑗 − 𝐹𝑗 (𝑧) = , (23)
𝐷𝑒𝑛𝑗 (𝑧) 𝑢(𝑘 + 𝑗|𝑘) ∈ U, 𝑗 ∈ N[0,𝑁𝑢 −1] , (30e)
𝛥𝑢(𝑘 + 𝑗|𝑘) ∈ 𝛥U, 𝑗 ∈ N[0,𝑁𝑢 −1] , (30f)
1
In this work with a slight abuse of notation, the unilateral -transform of ̃
where 𝐘(𝑘) = [𝑦̃1 (𝑘), … , 𝑦̃𝑁 (𝑘)] and the cost function definition is
∑∞
a general sequence such as {𝑞𝑖 } is defined by 𝑄(𝑧) = 𝑖=0 𝑞𝑖 𝑧−𝑖 . analogous to the RDMC based on the repetitive prediction error.

4
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

Remark 5. The multivariable filtered DMC for unstable processes with The previous lemma ensures null prediction error given repeti-
repetitive objectives can be directly obtained from the linear com- tive conditions (period 𝑁𝑟 ). Now, the periodic reference tracking is
bination of the input effects for a given output as discussed in [9, discussed. Consider a system in  domain represented by
Section 4].
𝑌 (𝑧) = 𝐻(𝑧)[𝑈 (𝑧) + 𝑊𝑢 (𝑧)] + 𝐻𝑤 (𝑧)𝑊 (𝑧),
Remark 6. The new GDMC for repetitive control can be interpreted as where 𝑊𝑢 (𝑧) and 𝐻𝑤 (𝑧)𝑊𝑦 (𝑧) describe bounded persistent repetitive
a periodic extension of [9]. If 𝑁𝑟 = 1, then the GDMC proposed in [9] disturbances with period 𝑁𝑟 . The convergence of 𝑦(𝑘) to the desired
is obtained. target 𝑦𝑟 (𝑘) where 𝑢(𝑘) is given by the generalized RDMC control law
is formally presented in the following theorem.
4.1. RDMC objectives

The main objectives of the new filtered DMC for periodic specifi- Theorem 1 (Periodic Target Tracking). Consider a system subject to
cations are: (i) to stabilize the DMC prediction for a periodic convolu- bounded repetitive disturbances with period 𝑁𝑟 . Assume that the proposed
tional model; and (ii) to achieve null prediction error in periodic steady- generalized RDMC is used to control this system such that 𝐹𝑗 (𝑧) represents
state condition. Now, these properties are discussed in the absence of the stable filters and the condition from Eq. (28) holds. If the closed-loop
active constraints in steady-state. system is BIBO stable and the constraints are not active in steady-state, then
Firstly, a repetitive signal with period 𝑁𝑟 in steady-state is defined
lim [𝑦𝑟 (𝑘) − 𝑦(𝑘)] = 0. (42)
from its -transform as follows 𝑘→∞

𝑁𝑟 (𝑧)
𝑅(𝑧) = + , (31) Proof. The internal closed-loop stability, the steady-state condition
𝐷𝑟 (𝑧)(𝑧𝑁𝑟 − 1)
without active constraints (linearity holds) and the fact that all external
where 𝑁𝑟 (𝑧) and (𝑧𝑛𝑟 − 1) have no common roots and the roots of 𝐷𝑟+ (𝑧) signals converge to a repetitive condition ensures that any signal of the
are strictly inside the unit circle. In this case, 𝑟(𝑘) = −1 {𝑅(𝑧)} is such closed-loop system can be described by 𝑋(𝑧) = 𝑁𝑥 (𝑧)∕[𝐷𝑥+ (𝑧)(𝑧𝑁𝑟 − 1)],
that
where 𝐷𝑥+ (𝑧) have all its roots inside the unitary circle.
lim 𝑟(𝑘) − 𝑟(𝑘 − 𝑁𝑟 ) = 0. (32) Thus, Lemma 1 holds and [𝑦(𝑘 + 𝑗) − 𝑦(𝑘 ̂ + 𝑗|𝑘)] = 0, 𝑗 = 1, … , 𝑁.
𝑘→∞
Now, it should be remarked that without active constraints, the
Now, if the repetitive steady-state condition is reached with a period generalized RDMC control law [9] converges to
𝑁𝑟 , then a null prediction error is achieved with the proposed RDMC
prediction model, as shown below. ̃
𝛥𝑟 𝑢(𝑘) = (𝐘𝑟 (𝑘) − 𝐘(𝑘) − ̃ 𝑗 𝛥𝑟  (𝑘)), (43)
̃
where 𝐘(𝑘)+ ̃ 𝑗 𝛥𝑟  (𝑘) is the free response, 𝐘𝑟 (𝑘) is the reference vector
Lemma 1 (Steady-State Prediction Error). Assume that 𝐹𝑗 (𝑧) are stable
and  = [1 0 … 0](𝐐⊤ 𝐐 + 𝜆𝐼)−1 𝐐⊤ where 𝐼 ∈ ℜ𝑁𝑢 ×𝑁𝑢 is an identity
filters, the condition Eq. (28) holds, and steady-state repetitive condition is
reached with a period 𝑁𝑟 such that the following limits hold matrix. The gain  is a non-null feedback vector obtained from the
optimal solution of the unconstrained optimization problem given by
lim 𝑦(𝑘) − 𝑦(𝑘 − 𝑁𝑟 ) = 0, (33) Eq. (17). Moreover, in a steady-state condition, lim𝑘→∞ ‖̃ 𝑗 𝛥𝑟  (𝑘)‖ =
𝑘→∞
0 (due to Lemma 1), and lim𝑘→∞ 𝛥𝑟 𝑢(𝑘) = 0 as a consequence of
lim 𝑢(𝑘) − 𝑢(𝑘 − 𝑁𝑟 ) = 0, (34) the closed-loop stability without active constraints is a steady-state
𝑘→∞
then, condition. Finally,

lim [𝑦(𝑘 + 𝑗) − 𝑦(𝑘

̂ + 𝑗|𝑘)] = 0, 𝑗 = 1, … , 𝑁. (35) ̃
lim 𝛥𝑟 𝑢(𝑘) = lim (𝐘𝑟 (𝑘) − 𝐘(𝑘)) = 0, (44)
𝑘→∞ 𝑘→∞ 𝑘→∞
̃
which implies in lim𝑘→∞ ‖𝐘𝑟 (𝑘) − 𝐘(𝑘)‖ = 0, or lim𝑘→∞ [𝑦𝑟 (𝑘 + 𝑗) − 𝑦(𝑘
̂ +
Proof. Firstly, the steady-state repetitive condition for a period 𝑁𝑟 can 𝑗|𝑘)] = 0, 𝑗 = 1, … , 𝑁. Now, as lim𝑘→∞ [𝑦(𝑘 + 𝑗) − 𝑦(𝑘 ̂ + 𝑗|𝑘)] = 0,
be represented by the following transforms {𝑦(𝑘)} = 𝑁𝑦 (𝑧)∕[𝐷𝑦+ (𝑧) 𝑗 = 1, … , 𝑁 (due to Lemma 1), lim𝑘→∞ [𝑦𝑟 (𝑘 + 𝑗) − 𝑦(𝑘 + 𝑗)] = 0,
(𝑧𝑁𝑟 − 1)] and {𝑢(𝑘)} = 𝑁𝑢 (𝑧)∕[𝐷𝑢+ (𝑧)(𝑧𝑁𝑟 − 1)], where 𝐷𝑦+ (𝑧) and 𝐷𝑢+ (𝑧) 𝑗 = 1, … , 𝑁. which completes the proof. □
have all their roots inside the unitary circle.
The generalized RDMC prediction is given by The same analysis can be applied to the multivariable case by
following the unconstrained solution presented in [9]. This analytical
̂ + 𝑗|𝑘) = 𝐐𝑗 𝛥𝑟 𝐮(𝑘) + ̃ 𝑗 𝛥𝑟  (𝑘) + 𝑦̃𝑗 (𝑘).
𝑦(𝑘 (36) unconstrained solution is also useful to perform a robust stability
For the convergence analysis, observe that lim𝑘→∞ 𝑢(𝑘) − 𝑢(𝑘 − 𝑁𝑟 ) = 0 analysis as discussed in Appendix B.
in order that lim𝑘→∞ 𝐐𝑗 𝛥𝑟 𝐮(𝑘) = 0 and lim𝑘→∞ ̃ 𝑗 𝛥𝑟  (𝑘) = 0 as a
consequence. Then, the prediction error (𝑒(𝑘+𝑗|𝑘) = [𝑦(𝑘+𝑗)−𝑦(𝑘+𝑗|𝑘)])
̂ 4.2. Data-driven filter design for repetitive specification
is such that
One of the main advantages of the DMC strategy comes from the fact
lim [𝑦(𝑘 + 𝑗) − 𝑦(𝑘
̂ + 𝑗|𝑘)] = lim [𝑦(𝑘 + 𝑗) − 𝑦̃𝑗 (𝑘)] (37)
𝑘→∞ 𝑘→∞ that the optimization algorithm can be entirely defined from the step-
Now, from 𝐹𝑗 (𝑧) = 𝑧𝑗 − 𝐿𝑗 (𝑧)(1 − 𝑧−𝑁𝑟 ), notice that response coefficients for a given cost function subject to constraints.
Therefore, only the step response coefficients are required to be iden-
{𝑦̃𝑗 (𝑘)} = 𝐹𝑗 (𝑧){𝑦(𝑘)} (38) tified. However, in the unstable case, the condition given by Eq. (28)
𝑗
= 𝑧 {𝑦(𝑘)} − 𝐿𝑗 (𝑧)(1 − 𝑧 −𝑁𝑟
){𝑦(𝑘)}. (39) should be verified. In order to avoid the parametric identification of
the system poles, a data-driven filter design has been proposed in [9]
As lim𝑘→∞ 𝑦(𝑘)−𝑦(𝑘−𝑁𝑟 ) = 0 and 𝐿𝑗 (𝑧) is a BIBO stable (𝐹𝑗 (𝑧) is a BIBO based on a least square minimization.
stable filter), from the final value theorem, the following convergence Now, the data-driven design for unstable systems with repetitive
holds specification is discussed. The least-square minimization can also be
lim −1 {𝐿𝑗 (𝑧)(1 − 𝑧−𝑁𝑟 ){𝑦(𝑘)}} = 0. (40) used in this problem, but in real systems, a nonlinear effect may be
𝑘→∞ present in the step response. In this case, the undesired pole is not
Therefore, the Eqs. (37), (39) and (40) can be combined as follows exactly considered due to the nonlinear distortion. However, as the
̂ + 𝑗|𝑘)] = lim [𝑦(𝑘 + 𝑗) − −1 {𝑧𝑗 {𝑦(𝑘)}}] = 0,
lim [𝑦(𝑘 + 𝑗) − 𝑦(𝑘 (41) DMC is based on a Finite Impulse Response model, the truncation error
𝑘→∞ 𝑘→∞ does not cause stability problems if |𝑔𝑖+𝑗 − 𝑔̃𝑗 | for 𝑖 > 𝑀 is significantly
which completes the proof of the lemma. □ smaller than |𝑔𝑖+𝑗 − 𝑔̃𝑗 | for 𝑖 ≪ 𝑀 because the inherent robust margins

5
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

ensure robust stability in this case [9, Section 5.1], and where 𝑔̃𝑖 are
the step response coefficients filtered by 𝐹𝑗 (𝑧). This is an expected
consequence as the DMC is based on a linear model, but a nonlinear
distortion can be verified from the practical step response.
The exact data-driven design is not required for robust stability, but
it may be desired for repetitive specifications. In this case, the uncon-
strained least square minimization can be replaced by a constrained
quadratic programming problem.
For simplicity, assume that
𝑛𝑁𝑟 +𝑚−1,𝑗 𝑧𝑁𝑟 +𝑚−1 + ⋯ + 𝑛1,𝑗 𝑧 + 𝑛0,𝑗
𝐹𝑗 (𝑧) =
𝑧𝑁𝑟 −1 (𝑧 − 𝛽)𝑚+1
where 𝜃 = [𝑛𝑁𝑟 +𝑚−1,𝑗 𝑛𝑁𝑟 +𝑚−2,𝑗 … 𝑛0,𝑗 ]⊤ is a decision vector and 𝛽
is a fixed parameter. Also assume that 𝑝1 , 𝑝2 , … 𝑝𝑁𝑟 are roots of
∑ Fig. 1. Open-loop test for a step response identification.
𝑝(𝑧) = 𝑧𝑁𝑟 − 1 and define 𝑄(𝑧) = ∞ −𝑖
𝑖=0 𝑞𝑖 𝑧 . The vector of parameters 𝜃
can be obtained from
𝑁𝑏
∑ order to reject a periodic disturbance with the fundamental frequency
min (𝑞𝓁 − 𝑞̃𝓁 ) (45a)
𝜃
𝓁=𝑁𝑎 defined by 𝜔 = 0.05 rad∕s. The GDMC [9] is compared with a repetitive
𝑛1,𝑗 𝑧+𝑛0
approach where 𝐹𝑗 (𝑧) = (𝑧−0.75) 2 are the standard stabilizing filters and
s.t. (45b) 𝑛 𝑧4 +𝑛 𝑧3 +𝑛 𝑧2 +𝑛 𝑧+𝑛
−1
𝐹𝑗 (𝑧) = 4,𝑗 3,𝑗 2,𝑗
𝑧3 (𝑧−0.75)2
1,𝑗 0
define the repetitive filters for each
{𝑞̃𝓁 } =  {𝐹𝑗 (𝑧)𝑄(𝑧)}, (45c)
prediction instant 𝑗.
𝐹𝑗 (𝑧𝑖 ) − 𝑧𝑗𝑖 = 0, 𝑧𝑖 = 𝑝𝑖 , 𝑖 = 1, 2, … 𝑁𝑟 , (45d) The open-loop test presented in Fig. 1 is used to recover the ap-
𝑛𝑁𝑟 +𝑚−1,𝑗 𝑧𝑁𝑟 +𝑚−1 + ⋯ + 𝑛1,𝑗 𝑧 + 𝑛0,𝑗 proximate step response coefficients as in [9]. Notice that the process is
𝐹𝑗 (𝑧) = , (45e) open-loop unstable as the output diverges for a constant input. Further
𝑧𝑁𝑟 −1 (𝑧 − 𝛽)𝑚+1 details about the equilibrium can be found in [9,24,25]. Moreover,
𝜃 = [𝑛𝑁𝑟 +𝑚−1,𝑗 𝑛𝑁𝑟 +𝑚−2,𝑗 … 𝑛0,𝑗 ]⊤ , (45f) it should be remarked that the system is nonlinear, thus, the coeffi-
cients (𝑔𝑖 ) are an approximate representation of the step response in a
where 𝑞̃𝓁 is a consequence of 𝜃 = [𝑛𝑁𝑟 +𝑚−1,𝑗 𝑛𝑁𝑟 +𝑚−2,𝑗 … 𝑛0,𝑗 ]⊤ and 𝜃 is
neighborhood of the equilibrium.
free a decision vector. The parameters 𝑁𝑎 and 𝑁𝑏 > 𝑀 can be defined
∑ 𝑁𝑏 The GDMC filter numerator is obtained from 𝑔𝑖 with a least square
by considering the criterion  (𝑁𝑎 , 𝑁𝑏 ) = 𝓁=𝑁 (𝑞𝓁 − 𝑞̃𝓁 ) as a figure
𝑎 minimization problem and the filter parameters of the Repetitive GDMC
of merit. A search grid can be used from 𝑁𝑎 and 𝑁𝑏 where a smaller
are computed from 𝑞𝑖 by solving the QP optimization problem as in
 (𝑁𝑎 , 𝑁𝑏 ) indicates an improved fit. More details can be found in [9].
Eq. (45). The last 20 samples of the response presented in Fig. 1
(interval between 200 s and 300 s) in the optimization horizon of
Remark 7. An adaptive filter approach [23] can be used to modify the data-driven design have been used to compute the numerators of
the filter online in order to reject periodic disturbances with a variable 𝐹𝑗 (𝑧−1 ). Now notice that (𝑘) = 𝐐𝛥𝑟 𝐮(𝑘) + 𝛥 ̃ 𝑟  (𝑘) = 𝐆𝛥𝐮(𝑘) + 𝛥
̃ (𝑘)
frequency nature. However, this type of solution is not in the scope of if 𝑁𝑟 = 1. Thus, in order to compare the convergence of the FIR
this work. predictor defined by 𝛥 ̃ 𝑟  , the infinity norm by column of ̃ (denoted
by ‖∶,𝑖 ‖∞ ) are shown in Fig. 2. As expected, the last elements of
̃
5. Case studies ‖̃ ∶,𝑖 ‖∞ are significantly smaller than the initial ones with respect
to 𝑖 such that the residual uncertainty is handled by the inherent
The differences between the standard and repetitive DMC algo- robustness of the GDMC strategies. The proposed data-driven strategies
rithms are presented in the following simulated case studies. Firstly, have effectively imposed the convergence of the FIR coefficients. This
an unstable open-loop problem is presented to illustrate the generality is a required behavior in order to guarantee that the infinite sum
of the proposed strategy [9]. Then, the tip-tracking control problem truncation is reasonable and has a minor effect with respect to the
with an inherently generated disturbance is presented to discuss the prediction error.
potential benefits of a repetitive strategy. The closed-loop response for set-point tracking and periodic dis-
turbance rejection are presented in Figs. 3 and 4. A sawtooth input
5.1. Continuous Stirred-Tank Reactor (CSTR) control — unstable case disturbance with amplitude ±0.1 and fundamental frequency 𝜔𝑟 =
0.05 rad∕s is applied by 𝜌(𝑡) between 300 s and 650 s. Both strategies are
The problem is defined from a continuously stirred tank reactor for able to track the set point and to ensure BIBO stability in the presence of
carrying out an enzymatic reaction [24,25]. The inlet concentration constraints, but the sawtooth disturbance has a persistent effect in the
𝐶𝑓 (𝑡) and well mixed output concentration 𝐶𝑜 (𝑡) are related by standard DMC. The Repetitive DMC effectively achieved the expected
𝑄∗𝑓 disturbance rejection in the presence of repetitive disturbances despite
𝑑𝐶𝑜 (𝑡) 𝑘1 𝐶𝑜 (𝑡)
= ∗ (𝐶𝑓 (𝑡) + 𝜌(𝑡) − 𝐶𝑜 (𝑡)) − , (46) the nonlinearity and the constraints in an unstable equilibrium point.
𝑑𝑡 𝑉𝑐 (𝑘2 𝐶𝑜 (𝑡) + 1)2
It is interesting to notice a small transient oscillation just after the
where 𝜌(𝑡) represents an additive input disturbance and the nominal setpoint change. It is caused by the plant-model mismatch that changes
operating parameters [25] are 𝑘1 = 10 L∕s, 𝑘2 = 10 L∕mol, 𝑉𝑐∗ = 1 L, the closed-loop poles and zeros. The closed-loop dynamics depend on
𝑄∗𝑓 = 0.03333 L∕s, 𝐶𝑓∗ = 3.288 mol∕L, and 𝐶𝑜∗ = 1.316 mol∕L. In the both the controller and the process models, hence, a new controller
absence of input-disturbance, then 𝜌(𝑡) ≡ 0. The manipulated variable imposes a different closed-loop response.
and the process variable are respectively assumed to be 𝑢(𝑘) = 𝐶𝑓 (𝑘𝑇𝑠 )− The effect of the choice of the filter pole (originally 𝛽 = 0.75)
𝐶𝑓∗ and 𝑦(𝑘) = 𝐶𝑜 (𝑘𝑇𝑠 ) − 𝐶𝑜∗ , where 𝐶𝑓∗ and 𝐶𝑜∗ represent the equilibrium on the closed-loop is now illustrated in a new simulation. Also, a
point and the sampling period is 𝑇𝑠 = 5 s. Simulink Band-Limited White Noise with noise power given by 10−6
Input and output constraints are given, respectively, by −0.05 ≤ and sampling period 𝑇𝑠 = 5 s was added to the output measurement
𝑦(𝑘𝑇𝑠 ) ≤ 0.12, and −0.1 ≤ 𝑢(𝑘𝑇𝑠 ) ≤ 0.1. The general design parameters in order to improve the discussion. In order to simulate the problem
are 𝑁 = 15, 𝑁𝑢 = 5, 𝜆 = 0.1, and 𝑀 = 40. Moreover, 𝑁𝑟 = 4 is used in with measurement noise, 𝑦𝑚 (𝑘) = 𝑦(𝑘) + 𝜁(𝑘) replaces 𝑦(𝑘) in the RDMC

6
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

Fig. 2. Convergence analysis of the FIR prediction model.

Fig. 5. Closed-loop response comparison with multiple filter designs of the concentra-
tion control based on the Repetitive GDMC with a sawtooth disturbance from 300 s until
650 s with amplitude ±0.1 and frequency 0.05 rad/s. A measurement noise (Simulink
Band-Limited White noise) with noise power given by 10−6 and sampling period 𝑇𝑠 = 5 s
was also included in this test.

Fig. 3. Closed-loop response of the concentration control based on the standard GDMC
with a sawtooth disturbance from 300 s until 650 s with amplitude ±0.1 and frequency
0.05 rad/s.

Fig. 6. Schematic representation of the two-link arm.

by the filter zeros that depend on the process and the repetitive period.

5.2. Position control of a non-minimum phase two-link arm

The position control for a two-link arm problem was experimentally

Fig. 4. Closed-loop response of the concentration control based on the Repetitive tested in [26]. In this work, a simulation study is presented, but the
GDMC with a sawtooth disturbance from 300 s until 650 s with amplitude ±0.1 and
source of uncertainties are: (i) the mismatch between the nonlinear
frequency 0.05 rad/s.
simulation model and linear MPC prediction model, and (ii) the quan-
tization error induced by the encoder simulation. The problem consists
of trajectory control for the tip of a two-link robotic arm in order to
optimization problem, where 𝜁(𝑘) is generated by the Limited White perform a pick-and-place task. The schematic representation of the two-
Noise source. link arm is represented in Fig. 6, where 𝑙1 and 𝑙2 represent the arm
In this case, the band-limited white noise definition provides peak lengths, (a) is an actuated joint controlled by a DC motor, (b) is a
values around ±0.001 that represent approximately 1% of the set-point passive joint with a torsional spring, 𝜃1 (𝑡) and 𝜃2 (𝑡) are the absolute
value and the results in closed-loop are shown in Fig. 5. The difference angles, 𝑇𝑚 (𝑡) is the torque provided by the DC motor, and 𝑤1,place
between the results with 𝛽 = 0.7 and 𝛽 = 0.8 is negligible in this exam- represents a desired displacement with respect to the origin in the axis
ple. The main difference is verified with 𝛽 = 0.9 during the interval 𝑤1 that defines the position where the place task is executed.
between 200 s and 300 s as the plant-model mismatch has a slower The control problem was introduced in [26,27]. The output position
transient effect in closed loop. Anyway, all options present the expected is defined by 𝑦(𝑘) = 𝑙1 sin(𝜃𝑒,1 (𝑘𝑇𝑠 )) + 𝑙2 sin(𝜃𝑒,2 (𝑘𝑇𝑠 )) (𝑦(𝑘) = 𝑤1,𝑡𝑖𝑝 (𝑘𝑇𝑠 )
results despite the measurement noise, plant-model mismatch and the in the axis 𝑤1 ) with 𝑇𝑠 = 0.1 s and 𝑦(𝑘) is a discrete-time variable.
sawtooth disturbance. It should be emphasized that the closed-loop The variation with respect to the axis 𝑤2 was not considered in [26],
response with respect to this design parameter is indirectly modified but it could be handled independently with an additional controlled

7
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

Table 1 as the system is open-loop stable and the RDMC with periodic error
System parameters.
correction is used.
Parameters Units Values The responses of the approximately linear scenario are shown in
N m
𝑘𝑠 rad
0.1772 Fig. 7 where the position reference is imposed to be 0 after 10 s. In
𝐽𝓁 kg m2 2.70 × 10−5 this problem, both strategies are able to track the desired target even
𝐽𝑑 kg m2 1.08 × 10−5
𝑚1 kg 0.050
when a constant reference is considered. The DMC aggressive design
𝑚2 kg 0.021 provides a closed-loop bandwidth that is large enough to preserve the
𝑙1 m 0.170 main frequencies of the reference signal, but a small tracking error is
𝑙2 m 0.155 visible.
𝐽1 kg m2 4.82 × 10−4
It should be remarked that repetitive control strategies implicitly
𝐽2 kg m2 1.68 × 10−4
𝑚𝑒𝑛𝑐 kg 0.100 impose additional closed-loop gain requirements in the multiples of
𝑚𝑏 kg 0.025 the fundamental frequency defined by 𝜔𝑟 = 2𝜋∕(𝑁𝑟 𝑇𝑠 ). Hence, ro-
m
𝑔 s2
9.81 bustness margins may be reduced around these frequencies and noise
N m s
𝑐11 rad
1.20 × 10−2 amplification in closed-loop can also be observed. Another drawback
N m s
𝑐12 rad
4.00 × 10−4 of the repetitive strategies can be seen from the control signal during
N m s
𝑐21 rad
4.00 × 10−4 a constant steady-state condition as depicted in Fig. 7(a) (between
N m s
𝑐22 rad
3.00 × 10−4 10 s and 20 s). A residual periodic control signal is naturally filtered
by the low-pass behavior of the system. Hence, its output effect is
numerically negligible with respect to the process variable. In this
case, 𝛥𝑟 𝑢(𝑘) = 0 → 𝑢(𝑘) = 𝑢(𝑘 − 𝑁𝑟 ) does not eliminate the periodic
prismatic joint. The angles 𝜃𝑒,1 (𝑡), 𝜃𝑒,2 (𝑡) are measured by encoders
residue which is a typical consequence of the controller’s repetitive
with a quantization error given by 𝑁𝑝𝑟𝑟 = 500 (number or pulses per
poles. Thus, a standard integral control approach is recommended if the
revolution).
control objective is defined only by constant steady-state specifications
The nonlinear model is described by
because 𝛥𝑢(𝑘) = 0 imposes a constant steady-state condition for control
[ ( )] [ ] problems with constant references and disturbances. On the other hand,
1
𝐽𝓁 + 𝐽𝑑 + 𝐽1 + 𝑚2,𝑒𝑞 𝑙12 𝑚 𝑙 𝑙 cos 𝜃1 (𝑡) − 𝜃2 (𝑡)
2 2 1 2
𝜃̈1 (𝑡)
1 ( ) the main advantage is that a null output tracking error is achieved
𝑚 𝑙 𝑙 cos 𝜃1 (𝑡) − 𝜃2 (𝑡) 𝐽2 ̈
𝜃2 (𝑡)
2 2 1 2 despite the periodic reference and plant-model mismatch effect as
[ ( ) ] [ ][ ]
1 ̇ 2 shown in Fig. 7(b) (between 5 s and 10 s).
𝑚 𝑙 𝑙 sin 𝜃1 (𝑡) − 𝜃2 (𝑡) 𝜃2 (𝑡) 𝑐 𝑐12 𝜃̇ 1 (𝑡)
+ 21 2 1 2 ( ) + 11
− 2 𝑚2 𝑙1 𝑙2 sin 𝜃1 (𝑡) − 𝜃2 (𝑡) 𝜃̇ 1 (𝑡)2 𝑐21 𝑐22 𝜃̇ 2 (𝑡)
[( ) ] [ ][ ] [ ] 5.2.1. Comparative analysis — enlarged position excursion
1
𝑚 + 𝑚2,𝑒𝑞 𝑔𝑙1 sin (𝜃1 (𝑡))
2 1
𝑘𝑠 −𝑘𝑠 𝜃1 (𝑡) 1 The tip-tracking control problem is considered with the same design
+ + = 𝑇 (𝑡),
1
𝑚 𝑔𝑙 sin (𝜃 (𝑡)) −𝑘𝑠 𝑘𝑠 𝜃2 (𝑡) 0 𝑚 parameters, but a larger tracking range is requested to analyze the
2 2 2 2
periodic disturbance rejection performance due to plant-model mis-
(47)
match. Four strategies are considered with this new reference scenario:
where the parameters of the reduced scale prototype [26] are presented the DMC, the RDMC, the repetitive GPC (RGPC), and a disturbance-
in Table 1. observer repetitive control [21]. The RGPC is based on the following
It should be emphasized that a parametric model was used [26,27] autoregressive model
in order to define the MPC control law, but only a step response is used 𝐴(𝑧−1 )(1 − 𝑧−𝑁𝑟 )𝑌 (𝑧) = 𝛥𝑟 (𝑧)𝐵(𝑧−1 )𝑧−1 𝑈 (𝑧) + 𝜖(𝑧), (48)
in this work. A modified DMC for systems with time-varying references
was applied to this problem in [19], but the periodic disturbance ̃ −1 ) = 𝐴(𝑧−1 )(1 − 𝑧−𝑁𝑟 ), 𝐵(𝑧−1 ) are parametric polynomial
where 𝐴(𝑧
rejection property does not hold because only the nominal feedforward models that should be known, and 𝜖(𝑧) is an external unknown distur-
control increment reference is modified. bance. The related disturbance-observer nonminimal state-space repet-
The parameterization of the periodic reference is presented in Ap- itive control (DO-NMSS-RC) [21] is based on a state-space model that
pendix C for simplicity, but only the pick position (𝑤1,pick ), the place also demands the knowledge of 𝐴(𝑧−1 ) and 𝐵(𝑧−1 ).2 Moreover, an
position (𝑤1,place ) and the pick-and-place task interval (𝑇Task ) are nec- additional disturbance observer should be designed in [21], which is
essary to define the periodic reference. Two scenarios are considered not a simple task as it depends on the disturbance model. The DO-
to emphasize the usefulness of the repetitive control with nonlinearity NMSS-RC state-feedback gain is given by a Linear Quadratic Regulator
and a periodic target. In the first one, the desired positions are given (LQR) control as in [21]. For comparison purposes, the RGPC cost
by 𝑤1,pick = (𝑙1 + 𝑙2 ) sin(−40𝜋∕180), 𝑤1,place = (𝑙1 + 𝑙2 ) sin(40𝜋∕180), function and the LQR objective are defined with the same RDMC
and the overall amplitude is 𝐴 = 2(𝑙1 + 𝑙2 ) sin(40𝜋∕180). Then, the tip weightings. In order to simplify the observer design and to achieve a
tracking amplitude excursion is enlarged to 𝐴 = 2(𝑙1 + 𝑙2 ) sin(80𝜋∕180) robust closed-loop response despite the uncertainties, the disturbance
with 𝑤1,pick = (𝑙1 + 𝑙2 ) sin(−80𝜋∕180), and 𝑤1,place = (𝑙1 + 𝑙2 ) sin(80𝜋∕180) model is described by 𝜇(𝑧) = (1 − 𝑧−1 )(1 − 2 cos(2𝜋∕𝑁𝑟 )𝑧−1 + 𝑧−2 )𝑧−1 𝜈(𝑧)
as in [21], where 𝜇(𝑘) is an input disturbance to be estimated and 𝜈(𝑧)
in order to show the periodic disturbance effect caused by an increased
is an external signal. Notice that a higher order model such as 𝜇(𝑧) =
angle variation range. The torque is bounded by −0.40 𝑁𝑚 ≤ 𝑇𝑚 (𝑡) ≤
(1 − 𝑧−𝑁𝑟 )𝑧−1 𝜈(𝑧) with 𝑁𝑟 = 20, for instance, significantly increases the
0.40 𝑁𝑚 as imposed by [26].
observer complexity and potentially reduces the robustness margins.
The DMC and the Repetitive DMC with periodic error correction
The chosen observer gain is 𝐾𝑜 = [0.2144 0.2156 0.2062]⊤ such that
(Section 2) are compared with respect to the proposed scenario. 𝑁 =
further details can be found in [21]. In this case, a larger tracking range
15, 𝑁𝑢 = 5, 𝑀 = 100 with 𝑇𝑠 = 0.1 s (sampling period) were
is requested in order to increase the nonlinear effect. Once more, an
used in both strategies with 𝜆 = 0.01 for the DMC and 𝜆 = 0.5 for
the RDMC. An aggressive design is required for the DMC otherwise a
significant steady-state tracking error is verified in all scenarios because 2
In this case, the continuous-time transfer function model was obtained
the reference is periodic. This type of tuning with reduced robustness from the linearization of the two-link arm model. The continuous-time transfer
margins is not necessary in the RDMC case. The step response model −19.91𝑠2 −15.29𝑠+8.38𝑒04
function is given by 𝑃 (𝑠) = 𝑠4 +4.47𝑠3 +1521𝑠2 +3611𝑠+8.048𝑒04
. Then, it is discretized
is obtained from the average step responses with amplitude 𝑇𝑚 (𝑡) = with the zero-order hold method to obtain 𝐴(𝑧−1 ) and 𝐵(𝑧−1 ). Notice that all
±0.1𝑁𝑚. Notice that the GDMC design is not necessary in this case the parameters of Table 1 are assumed to be known.

8
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

Fig. 7. Comparison of the approximately linear scenario with reduced tip position excursion — 𝐴 = 2(𝑙1 + 𝑙2 ) sin(40𝜋∕180). The position reference is switched to 0 after 10 s. Solid
lines represent the references and the dashed lines are the tip position output in the left figures.

Table 2 In this scenario, an important tracking error is observed with the

Comparison of the steady-state performance with measurement noise.
DMC in steady-state, but this problem is effectively handled by the
Criterion DMC RDMC RGPC DO-NMSS
RDMC. As expected, the periodic disturbance contribution (𝑇𝑟 = 20 s)
𝑟𝑚𝑠 0.052 0.015a 0.021 0.027 is eliminated from the output as expected. On the other hand, the DMC
T.V. 3.691 2.382 2.485 2.318a
is not able to deal with the periodic disturbances and a significant
a Best index value. steady-state error is present, which increases the mean squared error
index. The DO-NMSS only rejects the first harmonic, which provides
an interesting trade-off between a reduced control variation and the
additive measurement noise is imposed by 𝑦𝑚 (𝑘) = 𝑦(𝑘) + 𝜁(𝑘), where mean-squared error. The RGPC result is similar to the RDMC, but the
𝜁(𝑘) is generated by the Limited White Noise source with noise power sensitivity to noise is slightly higher due to its input disturbance model
given by 10−4 and 𝑇𝑠 = 0.1 s (sampling interval). Hence, 𝑦𝑚 (𝑘) replaces in comparison with the RDMC output disturbance model. In summary,
𝑦(𝑘) in the RDMC optimization problem. some minor performance variations can be observed due to the distur-
The comparative results are shown in Figs. 8 (DMC and RDMC) bance model differences. However, the proposed RDMC strategy is an
and 9 (RGPC and DO-NMSS). The standard DMC is not able to track interesting alternative to the related repetitive MPC algorithms because
the reference due to the time-varying effect of the plant-model mis- no parametric model (state-space or transfer function) is required, and
match. Moreover, the DMC high-gain design significantly degrades the the step-response description is commonly used in practice.
control signal in the presence of noise. The RDMC and RGPC have It is important to highlight that repetitive disturbances may be
similar performances, while the DO-NMSS is less sensitive to noise as generated from external sources, but they are also the consequence
a consequence of the proposed disturbance observer, but the tracking of time-varying references with a significant plant-model mismatch, as
performance is slightly degraded. This trade-off is expected due to the illustrated in this case study.
standard closed-loop design compromises. Finally, the sensitivity of the steady-state tracking error with respect
The mean squared steady-state tracking error and the steady-state to the control effort design parameter (𝜆) is illustrated in Fig. 10 for
Total Variation are respectively defined from 𝑘𝑇𝑠 = 10.1 until 𝑘𝑇𝑠 = 20 DMC and RDMC. The enlarged excursion range is considered with 𝜆 =
as follows 0.1 and 𝜆 = 0.2. It can be noticed that the DMC steady-state tracking
√
√ 200 error is highly sensitive to the design parameter because the closed-
√ ∑ [𝑦 (𝑘) − 𝑦(𝑘)]2
𝑟𝑚𝑠 = √ 𝑟
, (49) loop bandwidth is directly related to 𝜆. On the other hand, the RDMC
𝑘=101
100 preserves a reduced tracking error because the closed-loop gains of the
∑
200 sensitivity function around the resonance frequency are imposed by the
T.V. = |𝑢(𝑘) − 𝑢(𝑘 − 1)|. (50) controller poles.
𝑘=101 The steady-state performance comparison is shown in Table 3. As
These performances criteria are detailed in Table 2. It should be re- expected from Fig. 10 and the theoretical properties of the RDMC,
marked that the RDMC, the RGPC and the DO-NMSS are interesting the sensitivity of the steady-state tracking error with respect to the
alternatives to deal with periodic disturbances. However, RGPC and design parameter is reduced if compared with the DMC. The proposed
DO-NMSS demand an identification of the input–output model while approach is useful for rejecting repetitive disturbances and tracking pe-
the RDMC can be directly implemented from the coefficients of the step riodic references. Stable and unstable systems can be handled by using
response. only the information of the step-response coefficients and the controller

9
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

Fig. 8. Comparison of the nonlinear scenario with enlarged tip position excursion - 𝐴 = 2(𝑙1 + 𝑙2 ) sin(80𝜋∕180) with measurement noise. Solid lines represent the references and the
dashed lines are the tip position output in the left figures.

Fig. 9. Comparison of the nonlinear scenario with enlarged tip position excursion - 𝐴 = 2(𝑙1 + 𝑙2 ) sin(80𝜋∕180) with measurement noise. Solid lines represent the references and the
dashed lines are the tip position output in the left figures.

design requires no additional identification procedure. Furthermore, it Table 3

should be emphasized that all cases have been tested with a nonlinear Table with a comparison of the steady-state tracking error (𝑟𝑚𝑠 ).
simulation model. 𝜆 𝑟𝑚𝑠 (DMC) 𝑟𝑚𝑠 (RDMC)
0.1 0.0693 0.0042
5.3. Experimental case study 0.2 0.1022 0.0044

The experimental temperature control of a Negative Temperature

Coefficient (NTC) sensor is analyzed in this laboratory case study. In

10
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

Fig. 10. Analysis of the design sensitivity with enlarged tip position excursion - 𝐴 = 2(𝑙1 + 𝑙2 ) sin(80𝜋∕180). Solid thicker lines represent the references. The parameters 𝜆 = 0.1 and
𝜆 = 0.2 are indicated in the legend.

Fig. 11. Comparison of the closed-loop responses of the NTC temperature control case. In the left figures, solid lines represent the references and the dashed lines are the
temperature measures.

the single sensor configuration, a temperature variation is imposed on in Scilab with the LabVIEW to Scilab Gateway. The sampling time is
the NTC such that this time-varying temperature is used to estimate 𝑇𝑠 = 0.2 ms and the digital filter (𝐹 (𝑧) = 𝑧(1−0.95)
(𝑧−0.95)
) is implemented to
the desired measurand. For instance, it can be used in fluid velocity filter the current command in order to avoid the high control signal
estimation [28]. oscillations caused by the significant measurement noise. The filter
The laboratory setup is composed of a controlled current source effect is considered in the convolutional step-response model.
(𝐼𝑠 (𝑡)), which is used to heat the NTC sensor due to its dissipated The nonlinear-model for control can be defined by
power. The experiment uses an EPCOS miniature sensor with bendable 𝑑𝑦(𝑡) 𝐺 𝐴 𝐵𝑡ℎ
= − 𝑡ℎ 𝑦(𝑡) + 𝑡ℎ 𝑒 𝑦(𝑡) 𝑣(𝑡) + 𝜖(𝑡), (51)
wires [29]. The current source range varies between 1∕3.3 mA and 𝑑𝑡 𝐶𝑡ℎ 𝐶𝑡ℎ
10∕3.3 mA in this work. The control interface is implemented in Lab- where 𝜖 is an external disturbance signal; 𝑦(𝑡) is the sensor temperature,
VIEW, and the DMC and RDMC predictive controllers are implemented 𝑣(𝑡) = 𝐼𝑠2 (𝑡) is the squared manipulated sensor current, 𝐺𝑡ℎ and 𝐶𝑡ℎ , 𝐴𝑡ℎ ,

11
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

and 𝐵𝑡ℎ are the system parameters. This model is presented to illustrate Acknowledgments
𝐵𝑡ℎ
the nonlinearity imposed by the term 𝑒 𝑦(𝑡) 𝑣(𝑡), but the step response
is directly used for control. It should be remarked that 𝐴𝑡ℎ , 𝐵𝑡ℎ , 𝐺𝑡ℎ , This research received funding from CNPq Conselho Nacional de De-
and 𝐶𝑡ℎ are available in the NTC datasheet [29]. However, 𝐺𝑡ℎ and senvolvimento Científico e Tecnológico, Brazil under grants 308741/2021-
the external disturbance 𝜖(𝑡) depend on the external temperature, fluid 8, 304032/2019-0 and 406477/2022-1. This study was financed in
velocity, and other ambient conditions [28] such that the equilibrium
part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior
point and the linearized model cannot be directly obtained by the
– Brazil (CAPES) – Finance Code 001. The authors acknowledge the
datasheet parameters.
support provided by the CNPq and CAPES.
No parameter knowledge is required in this case study since both
DMC and RDMC use a step response approach. The DMC and RDMC
parameters are 𝑁 = 15, 𝑁𝑢 = 5, 𝑀 = 300, and 𝑁𝑟 = 100. The control Appendix A. Prediction matrices
weights for the RDMC and DMC are 𝜆 = 10−3 and 𝜆 = 10−6 , respectively.
The DMC one needs to have a more aggressive design, as previously
discussed. The parameterization of the reference is similar to the two-
A.1. RDMC based on repetitive error
link case (5th) law of motion, but 𝑇𝑡𝑎𝑠𝑘 = 10 s (𝑁𝑟 = 2𝑇𝑡𝑎𝑠𝑘 ∕𝑇𝑠 ) and
the reference range varies between 305 K and 307 K. The comparative
closed-loop results are shown in Fig. 11 where the controlled operation The prediction matrices can be formulated as follows
starts at 10 s.
These experimental results are relevant because, once more, the (𝑘) = 𝐐𝛥𝑟 𝐮(𝑘) + 𝛥𝑟  (𝑘) + 𝐘𝑁𝑟 (𝑘) (A.1)
DMC and the RDMC have provided interesting closed-loop responses
with
despite the multiple sources of uncertainties and measurement noise.
The steady-state tracking error is slightly better in the RDMC case as (𝑘) = [𝑦(𝑘 ̂ + 𝑁|𝑘)]⊤ ,
̂ + 1|𝑘) … 𝑦(𝑘 (A.2)
its disturbance model considers the periodic nature of the tracking 𝑟 𝑟 𝑟 ⊤
𝛥 𝐮(𝑘) = [𝛥 𝑢(𝑘|𝑘) … 𝛥 𝑢(𝑘 + 𝑁𝑢 − 1|𝑘)] , (A.3)
error. The steady-state T.V. index and the means squared error have
been computed for the interval between 100 s and 200 s (500 and ⊤
𝐘𝑁𝑟 (𝑘) = [𝑦(𝑘 + 1 − 𝑚1 𝑁𝑟 ) … 𝑦(𝑘 + 𝑁 − 𝑚𝑁 𝑁𝑟 )] , (A.4)
1000 samples). The mean squared errors are 0.712 (DMC), and 0.389 𝑟 𝑟 𝑟 ⊤
𝛥 (𝑘) = [𝛥 𝑢(𝑘 − 1) … 𝛥 𝑢(𝑘 − 𝑀)] (A.5)
(RDMC) while the T.V. values are 4.441 (DMC), and 4.755 (RDMC).
These results show that the steady-state tracking error was reduced as follows:
with the RDMC, but an increased control activity is a price to pay in
this experimental case. However, note that the reduction of the mean ⎡ 𝑞1 0 … 0⎤
⎢𝑞 𝑞1 … 0⎥
squared error provided by the RDMC is approximately 45%, and the 𝐐=⎢ 2 ⎥, (A.6)
increase in control activity is only 7%, showing a very good trade-off ⎢⋮ ⋮ ⋱ ⋮⎥
⎢𝑞 𝑞𝑁−1 … 𝑞𝑁−𝑁𝑢 +1 ⎥⎦
between performance and control effort. ⎣ 𝑁

6. Conclusions ⎡ 𝑞2 − 𝑞2−𝑚 𝑁 𝑞3 − 𝑞3−𝑚1 𝑁𝑟 … 𝑞𝑀+1 − 𝑞𝑀+1−𝑚1 𝑁𝑟 ⎤

⎢ 𝑞 −𝑞 1 𝑟 𝑞4 − 𝑞4−𝑚2 𝑁𝑟 … 𝑞𝑀+1 − 𝑞𝑀+2−𝑚2 𝑁𝑟 ⎥⎥
=⎢ 3 3−𝑚2 𝑁𝑟
A new repetitive DMC strategy is proposed to extend repetitive ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢𝑞 − 𝑞 𝑞𝑁+1 − 𝑞𝑁+1−𝑚𝑁 𝑁𝑟 … 𝑞𝑀+𝑁 − 𝑞𝑀+𝑁−𝑚𝑁 𝑁𝑟 .⎥⎦
control properties to the well-known DMC algorithm. The novel DMC- ⎣ 𝑁 𝑁−𝑚𝑁 𝑁𝑟
based strategy is effective in controlling stable, unstable, and integra- (A.7)
tive processes with periodic specification because repetitive disturbance
rejection can be imposed and periodic reference tracking can be en-
sured. Thus, the DMC applications can be extended to deal with control A.2. RDMC based on filtered error
problems with periodic specifications. The implementation is simplified
if compared with related repetitive model predictive controllers be-
cause only the step-response coefficients are required, although it uses In this case, prediction matrices can be formulated as follows
more filters to compute the future prediction error. The advantages of (𝑘) = 𝐐𝛥𝑟 𝐮(𝑘) + 𝛥 ̃
̃ 𝑟  (𝑘) + 𝐘(𝑘) (A.8)
the proposed RDMC over the standard DMC were confirmed in two case
studies and an experimental pilot plant, corroborating the theoretical with
results presented in this paper. The repetitive control of systems subject
to disturbances with time-varying periods is an interesting topic for (𝑘) = [𝑦(𝑘 ̂ + 𝑁|𝑘)]⊤ ,
̂ + 1|𝑘) … 𝑦(𝑘 (A.9)
future research. 𝑟 𝑟 𝑟
𝛥 𝐮(𝑘) = [𝛥 𝑢(𝑘|𝑘) … 𝛥 𝑢(𝑘 + 𝑁𝑢 − 1|𝑘)] , ⊤
(A.10)
̃
𝐘(𝑘) = [𝑦̃1 (𝑘) … 𝑦̃𝑁 (𝑘)]⊤ , (A.11)
CRediT authorship contribution statement
𝑟 𝑟 𝑟 ⊤
𝛥 (𝑘) = [𝛥 𝑢(𝑘 − 1) … 𝛥 𝑢(𝑘 − 𝑀)] , (A.12)
Tito L.M. Santos: Writing – original draft, Investigation, Fund-
ing acquisition, Formal analysis, Conceptualization. Daniel M. Lima: as follows:
Writing – review & editing, Investigation, Conceptualization. Julio E. ⎡ 𝑞1 0 … 0⎤
Normey-Rico: Writing – review & editing, Supervision, Investigation, ⎢𝑞 𝑞1 … 0⎥
Funding acquisition, Conceptualization. 𝐐=⎢ 2 ⎥, (A.13)
⎢⋮ ⋮ ⋱ ⋮⎥
⎢𝑞 𝑞𝑁−1 … 𝑞𝑁−𝑁𝑢 +1 ⎥⎦
⎣ 𝑁
Declaration of competing interest
⎡ 𝑞2 − 𝑞̃1 𝑞3 − 𝑞̃2 … 𝑞𝑀+1 − 𝑞̃𝑀 ⎤
The authors declare that they have no known competing finan- ⎢ ⎥
𝑞 − 𝑞̃1 𝑞4 − 𝑞̃2 … 𝑞𝑀+1 − 𝑞̃𝑀 ⎥
=⎢ 3
̃ . (A.14)
cial interests or personal relationships that could have appeared to ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
influence the work reported in this paper. ⎢𝑞 − 𝑞̃ 𝑞𝑁+1 − 𝑞̃2 … ⎥
𝑞𝑀+𝑁 − 𝑞̃𝑀 ⎦
⎣ 𝑁 1

12
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

Appendix B. Inherent robustness analysis the task. The general parameters are 𝐴 = 𝑤1,place − 𝑤1,pick , 𝑓0 = 𝑤1,pick ,
3 ), 𝑓 4 5
𝑓3 = 10𝐴∕(𝑇task 4 = −15𝐴∕(𝑇task ), 𝑓5 = 6𝐴∕(𝑇task ). In order to
The inherent robustness analysis for the case without active con- start from the reference at the origin, a shifted target is used with
straints can be studied as in [9]. The generalized RDMC version is 𝑓𝑠 (𝑡) = 𝑓 (𝑡 + 𝑇task ∕2) such that 𝑓𝑠 (0) = 0. Finally, 𝑓𝑠 (𝑡) is sampled
discussed due to its generality, but the same formulation can be applied for discrete-time implementation where the reference given by 𝑦𝑟 (𝑘) =
to the unfiltered RDMC. 𝑓𝑠 (𝑘𝑇𝑠 ), 0 ≤ 𝑘 < 𝑇task ∕𝑇𝑠 , 𝑁𝑒𝑛𝑑 = 2𝑇task ∕𝑇𝑠 , and 𝑦𝑟 (𝑘 + 𝑁𝑒𝑛𝑑 ) = 𝑦𝑟 (𝑘). In
In this work, first define the nominal truncated model given by this work, 𝑇task = 1 s, 𝑇𝑠 = 0.1, the continuous-time repetitive period is
𝑇𝑟 = 2 s and 𝑁𝑟 = 𝑇𝑟 ∕𝑇𝑠 = 20.
∑
𝑀
𝐻(𝑧) = (𝑞𝑖 − 𝑞𝑖−𝑁𝑟 )𝑧−𝑗 . (B.1)
𝑗=1 References
Then assume that 𝐻(𝑧) is the unknown transfer function that describes
[1] Cutler CR, Ramaker BL. Dynamic matrix control — A computer control al-
the process such that the plant-model mismatch is represented by gorithm. In: Proceedings joint automatic control conference, vol. 1, 1980, p.
72.
𝐻(𝑧) = 𝐻(𝑧)[1 + (𝑧)], (B.2)
[2] Qin SJ, Badgwell TA. A survey of industrial model predictive control technology.
Control Eng Pract 2003;11:733–64. http://dx.doi.org/10.1016/S0967-0661(02)
where (𝑧) is a multiplicative uncertainty.
00186-7.
Now notice that 𝑌 (𝑧) = 𝐻(𝑧)[1 + (𝑧)]𝑈 (𝑧) or alternatively [3] Garcia CE, Morshedi A. Quadratic programming solution of dynamic matrix
control (QDMC). Chem Eng Commun 1986;46:73–87. http://dx.doi.org/10.1080/
𝑌 (𝑧) =𝐻(𝑧)𝑈 (𝑧) + 𝑊𝑚 (𝑧), (B.3) 00986448608911397.
𝑊𝑚 (𝑧) =𝐻(𝑧)(𝑧)𝑈 (𝑧). (B.4) [4] Lima DM, Normey-Rico JE, Santos TLM. Temperature control in a solar collector
field using filtered dynamic matrix control. ISA Trans 2016;62:39–49. http:
Then, by combining Eqs. (A.8) (Appendix A) and (43), observe that the //dx.doi.org/10.1016/j.isatra.2015.09.016.
[5] Klopot T, Skupin P, Metzger M, Grelewicz P. Tuning strategy for dynamic matrix
optimal unconstrained control law is given by
control with reduced horizons. ISA Trans 2018;76:145–54. http://dx.doi.org/10.
1 ̃ ̃ 𝑟  (𝑧)] 1016/j.isatra.2018.03.003.
𝑈 (𝑧) = [𝐘𝑟 (𝑧) − 𝐘(𝑧) − 𝛥 (B.5)
1 − 𝑧−𝑁𝑟 [6] Zhang D, Hu Y, Gao Y. Optimization control of a 330 MW drum boiler unit
1 based on DMC algorithm and DEB strategy. ISA Trans 2021. http://dx.doi.org/
= ̃
[𝐘𝑟 (𝑧) − 𝐘(𝑧)] ̃ (𝑧),
−  (B.6) 10.1016/j.isatra.2021.10.027.
1 − 𝑧−𝑁𝑟
[7] Peccin VB, Lima DM, Flesch RC, Normey-Rico JE. Fast constrained dynamic
where  (𝑧) = {[𝑢(𝑘 − 1), 𝑢(𝑘 − 2), 𝑢(𝑘 − 𝑀)]⊤ }. The following transfer matrix control algorithm with online optimization. Rev Iberoam Autom Inform
matrices are used to simplify the Small-Gain Theorem analysis Ind 2022. http://dx.doi.org/10.4995/riai.2022.16619.
[8] Shi Y, Lin R, Wu X, Zhang Z, Sun P, Xie L, et al. Dual-mode fast DMC
⎡𝐹1 (𝑧)⎤ ⎡ 𝑧−1 (𝑧) ⎤ algorithm for the control of ORC based waste heat recovery system. Energy
𝐅(𝑧) = ⎢ ⋮ ⎥ , 𝐓(𝑧) = ⎢ ⋮ ⎥ . (B.7) 2022;244:122664. http://dx.doi.org/10.1016/j.energy.2021.122664.
⎢ ⎥ ⎢ −𝑀 ⎥ [9] Santos TL, Normey-Rico JE. A generalised dynamic matrix control for unstable
⎣𝐹𝑛 (𝑧)⎦ ⎣𝑧 (𝑧)⎦
processes based on filtered predictions. ISA Trans 2023;136:297–307. http://dx.
Therefore, the control law Eq. (B.6) can be described by doi.org/10.1016/j.isatra.2022.11.010.
[10] Wang L, Freeman CT, Chai S, Rogers E. Predictive-repetitive control with
̃
(1 + 𝐓(𝑧)) −1
constraints: From design to implementation. J Process Control 2013;23:956–67.
𝑈 (𝑧) = −𝑁
[𝐘𝑟 (𝑧) − 𝐅(𝑧)𝑌 (𝑧)] (B.8)
1−𝑧 𝑟 http://dx.doi.org/10.1016/j.jprocont.2013.03.012.
[11] Wang L. Tutorial review on repetitive control with anti-windup mechanisms.
For robustness analysis, 𝐘𝑟 (𝑧) can be assumed to be zero with no loss
Annu Rev Control 2016;42:332–45. http://dx.doi.org/10.1016/j.arcontrol.2016.
of generality such that 09.016.
̃ −1 [12] Frederik J, Kröger L, Gülker G, van Wingerden J-W. Data-driven repetitive
(1 + 𝐓(𝑧))
𝑈 (𝑧) = − −𝑁
𝐅(𝑧)𝑌 (𝑧) (B.9) control: Wind tunnel experiments under turbulent conditions. Control Eng Pract
1−𝑧 𝑟
2018;80:105–15. http://dx.doi.org/10.1016/j.conengprac.2018.08.011.
= − 𝐶𝑒𝑞 (𝑧)𝑌 (𝑧). (B.10) [13] Navalkar ST, van Wingerden JW, van Solingen E, Oomen T, Pasterkamp E, van
Kuik GAM. Subspace predictive repetitive control to mitigate periodic loads on
̃
(1+𝐓(𝑧)) −1
Notice that the equivalent controller (𝐶𝑒𝑞 (𝑧) = has 𝐅(𝑧)) large scale wind turbines. Mechatronics 2014;24:916–25. http://dx.doi.org/10.
1−𝑧−𝑁𝑟 1016/j.mechatronics.2014.01.005.
the expected repetitive poles. [14] Natarajan S, Lee JH. Repetitive model predictive control applied to a simulated
Finally, from Eq. (B.3), the following closed-loop system for robust- moving bed chromatography system. Comput Chem Eng 2000;24:1127–33. http:
ness analysis is obtained //dx.doi.org/10.1016/S0098-1354(00)00493-2.
[15] Balaji S, Fuxman A, Lakshminarayanan S, Forbes JF, Hayes RE. Repetitive model
𝑌 (𝑧) =(1 + 𝐻𝐶𝑒𝑞 (𝑧))−1 𝑊𝑚 (𝑧), (B.11) predictive control of a reverse flow reactor. Chem Eng Sci 2007;62:2154–67.
http://dx.doi.org/10.1016/j.ces.2006.12.082.
𝑊𝑚 (𝑧) = − 𝐻(𝑧)(𝑧)𝐶𝑒𝑞 (𝑧)𝑌 (𝑧), (B.12) [16] Lunardi A, Conde E, Assis J, Meegahapola L, Fernandes DA, Sguarezi Filho AJ.
Repetitive predictive control for current control of grid-connected inverter under
where the robust stability criterion is simply distorted voltage conditions. IEEE Access 2022;10:16931–41. http://dx.doi.org/
10.1109/ACCESS.2022.3147812.
‖(𝑧)‖∞ ‖𝐻(𝑧)𝐶𝑒𝑞 (𝑧)(1 + 𝐻𝐶𝑒𝑞 (𝑧))−1 ‖∞ < 1, (B.13) [17] Liu Y, Cheng S, Ning B, Li Y. Robust model predictive control with simplified
repetitive control for electrical machine drives. IEEE Trans Power Electron
This unconstrained analysis can be used to evaluate the inherent ro-
2019;34:4524–35. http://dx.doi.org/10.1109/TPEL.2018.2857837.
bustness of a given RDMC controller without active constraints. [18] Houtzager I, van Wingerden J-W, Verhaegen M. Rejection of periodic wind
disturbances on a smart rotor test section using lifted repetitive control. IEEE
Appendix C. Periodic reference parameterization — tip-tracking Trans Control Syst Technol 2013;21:347–59. http://dx.doi.org/10.1109/TCST.
problem and experimental case study 2011.2181171.
[19] Pereira BSS, Santos TLM, Araújo JM. Receptance-based model predictive control
of multibody systems with time-varying references. Meccanica 2024;59:33–48.
A periodic reference was defined in [26]. The chosen motion law http://dx.doi.org/10.1007/s11012-023-01737-3.
is given by 𝑓 (𝑡) = 𝑓0 + 𝑓3 𝑡3 + 𝑓4 𝑡4 + 𝑓5 𝑡5 , 0 ≤ 𝑡 < 𝑇task ∕2, and [20] Liuping Wang CTF, Rogers E. Disturbance observer-based predictive repetitive
𝑓 (𝑡) = 𝑓0 + 𝑓3 (𝑇task − 𝑡)3 + 𝑓4 (𝑇task − 𝑡)4 + 𝑓5 (𝑇task − 𝑡)5 , 𝑇task ∕2 ≤ 𝑡 < 𝑇task . control with constraints. Internat J Control 2022;95:1060–9. http://dx.doi.org/
The constants 𝑓0 , 𝑓3 , 𝑓4 and 𝑓5 are computed from the pick (𝑤1,pick ) 10.1080/00207179.2020.1839674.
[21] Wang L, Freeman CT, Rogers E, Young PC. Disturbance observer-based repet-
and the place positions (𝑤1,place ) and the task interval between the itive control system with nonminimal state-space realization and experimental
pick and the place instant (𝑇task ). The overall period is 𝑇𝑟 = 2𝑇task evaluation. IEEE Trans Control Syst Technol 2023;31:961–8. http://dx.doi.org/
because the tip should return to the pick position in order to restart 10.1109/TCST.2022.3202299.

13
T.L.M. Santos et al. ISA Transactions xxx (xxxx) xxx

[22] Camacho E, Bordons C. Model predictive control. Berlin: Springer; 1999. [27] Richiedei D, Trevisani A. Simultaneous active and passive control for eigen-
[23] Tito L. M. Santos BPMS, Uzêda L. Multivariable filtered smith predictor for structure assignment in lightly damped systems. Mech Syst Signal Process
systems with sinusoidal disturbances. Internat J Systems Sci 2017;48:2182–94. 2017;85:556–66. http://dx.doi.org/10.1016/j.ymssp.2016.08.046.
http://dx.doi.org/10.1080/00207721.2017.1309591. [28] Ferreira RPC, Freire RCS, Deep CS, de Rocha Neto JS, Oliveira A. Hot-wire
[24] Liou C-T, Yu-Shu C. The effect of nonideal mixing on input multiplicity in a anemometer with temperature compensation using only one sensor. IEEE Trans
CSTR. Chem Eng Sci 1991;46:2113–6. http://dx.doi.org/10.1016/0009-2509(91) Instrum Meas 2001;50:954–8.
80168-X. [29] TDK. NTC thermistors for temperature measurement miniature sensors with
[25] Sree RP, Chidambaram M. Control of unstable bioreactor with dominant unstable bendable wires. 2013, Datasheet - EPCOS (TDK).
zero. Chem Biochem Eng Q 2003;17:139–46.
[26] Bettega J, Richiedei D. Trajectory tracking in an underactuated, non-minimum
phase two-link multibody system through model predictive control with embed-
ded reference dynamics. Mech Mach Theory 2023;180:105165. http://dx.doi.
org/10.1016/j.mechmachtheory.2022.105165.

14