This article has been accepted for publication in IEEE Transactions on Industrial Electronics.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Predictor-Based Data-Driven Model-Free

Adaptive Predictive Control of Power Converters
Using Machine Learning
Xing Liu, Senior Member, IEEE, Lin Qiu, Senior Member, IEEE, Youtong Fang, Senior Member, IEEE,
and Jose Rodrı́guez, Life Fellow, IEEE

Abstract—In this article, a novel robust data-driven I. I NTRODUCTION

model-free predictive control framework based on the I/O
data of the controlled plants, which is performed by in-
corporating the neural predictor-based model-free adap-
tive control and finite control-set model predictive control,
R ECENTLY, there has been significant research efforts
aimed at the development of systematic design solution
for finite control-set model predictive control (FCS-MPC) of
is first proposed. The salient feature of the suggested power converters [1]–[3]. Many remarkable achievements and
framework is that the uncertainties, such as unmodeled representative results have been obtained due to its merit in
dynamics and external disturbances, can be explicitly ad-
dressed in controlled systems. From a practical stand- the straightforward design of the optimization criterion and
point, however, the potential of this proposal is limited by control objectives, thus allowing one to consolidate explicitly
a significantly increased online computational complexity, nonlinearities and to take system constraints into account
which makes it difficult to implement. To circumvent this in the problem formulation. Despite the progress mentioned
limitation, a supervised imitation learning technique using above, these technological advances bring two fundamental
data labeled is developed to imitate the known suggested
controller, which the majority of the online computational challenges.
burden can be transformed into offline computing by utiliz-
ing a trained artificial neural network subject to robustness A. Related Work
characteristics. In particular, this development motivates
a much simpler robust predictive control solution, which The first one is the system uncertainties for specific pa-
is convenient to implement in applications. Thus, by this rameters and operating point [4], [5]. According to well-
proposal, the online implementation of much more com- established knowledge, FCS-MPC can be regarded as model-
plex predictive control strategies is made possible, and it
explores a new possibility for future development of the based control strategy. In this strategy, the system model
complex control methodology. Finally, extensive simulative plays an important role. However, model plant mismatch is
and experimental investigations for modular multilevel con- inevitable because of various uncertainties in the practical
verter validate the interest and viability of the proposed control environment. To be specific, uncertainties occur not
design methodology. only due to the sensitivity to environmental operating con-
Index Terms—Finite control-set model predictive control, ditions but also due to the presence of external unknown
predictor-based neural network, model-free adaptive con- disturbances in the control process [6]. In this regard, the
trol, machine learning, modular multilevel converter. control performance obtained by the traditional FCS-MPC
method depends greatly on the accuracy of system model,
which often causes the lack of robustness. Meanwhile, system
parameter variations may trigger the instability issues of the
system if not properly accounted for [7]. To alleviate the
This work was supported in part by the National Natural Sci- issue, as an alternative, online identification techniques are
ence Foundation of China under Grant 51807177, Grant 51827810, normally employed to define the system behavior using model
and in part by China Postdoctoral Science Foundation under Grant reference adaptive system (MRAS) technique [8] and recursive
2020M681855, and in part by the National Key Research and De-
velopment Program of China under Grant 2019YFB1504603, and in least squares (RLS) estimation [9] algorithm, thus allowing
part by Natural Science Foundation of Zhejiang Province under Grant the improvement of robust control performance. In spite of
LY21E070004 and Grant LY22E070003. This work of Jose Rodriguez the aforementioned developments, the performance largely
was supported by ANID through projects under Grant FB0008, Grant
1210208, and Grant 1221293. (Corresponding authors: Lin Qiu.) depends on the accuracy of the online identification strategy
X. Liu and Y. Fang are with the College of Electrical Engineering, Zhe- in the industrial control environment.
jiang University, Hangzhou 310027, China (e-mail: xingldl@zju.edu.cn; More recently, an interesting trend is to explore data-based
youtong@zju.edu.cn).
L. Qiu is with the College of Electrical Engineering, Zhejiang Univer- control approach that could potentially offer good robustness
sity, Hangzhou 310027, China. He is also with the Zhejiang University- characteristics without deploying the online identification tech-
University of Illinois at Urbana-Champaign Institute, Hangzhou 310027, nique, instead of an explicit model [10]–[12]. By virtue of
China (e-mail: qiu lin@zju.edu.cn).
J. Rodriguez is with Faculty of Engineering, Universidad San Sebas- its high robustness, a model-free adaptive control (MFAC)
tian Santiago 8370146, Chile (e-mail: jose.rodriguezp@uss.cl). strategy, which is based on the input-output data of a real

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

plant, is firstly proposed [13]. Thus, the parameter sensitivity variations due to the effect of the parametric uncertainties and
issue can be alleviated, and the robustness of the system can environmental disturbances, and then, the performance on the
be raised. Meanwhile, to take the system uncertainties into nonlinear uncertain system could be limited, which may limit
account, the idea in [13] is further developed by incorporating the real-time control applications.
a discrete-time extended state observer (DESO) into the MFAC
framework [14], where the DESO technique is employed to B. Contributions
estimate the unknown residual nonlinear time-varying term Inspired by the latest advances in the field of machine
and to compensate the influence of uncertainties and distur- learning and robust predictive control, to simultaneously taken
bances, thus enhancing the robustness of the controlled plant. the aforementioned issues into account, in this work we seek
Next, the authors in [15] proposed a novel robust FCS-MPC to develop a systematic design procedure for power converter
solution. It is realized by combining a predictor-based neural systems in a more general form. Specifically, a novel robust
network (NN) design with FCS-MPC for modular multilevel data-driven model-free adaptive predictive control framework
converter (MMC), enabling a smooth and fast identification is presented, and we show that the framework can compensate
of system dynamics. After that, the idea in [15] is further the loss of accuracy in system model, and thus is able to
explored by embedding the event-triggered mechanism into explicitly deal with the uncertainties of the controlled plant.
predictor-based NN predictive control [16], and later extended But, it inevitably undergoes a prohibitive online computational
to neural predictor-based dynamic surface predictive control cost while applying the proposed methodology to the control
[17] and data-driven neural predictors-based robust MPC [18]. of power converters in the experimental platforms and validate
Alternatively, many remarkable accomplishments have also experimentally. Fortunately, to facilitate an efficient online
been obtained with the help of deep reinforcement learning- implementation, as an alternative, a surrogate controller is
based PWM control [19], reinforcement learning-based opti- developed to imitate the known suggested methodology with
mal control [20], Lyapunov-based FCS-MPC [21], and fuzzy the help of the supervised imitation learning technique using
approximation-based FCS-MPC [22]. It is shown that recent data labeled. In view of this, the majority of the online compu-
results on this field can be successfully generalized to cope tational burden in this proposal can be transformed into offline
with uncertainties, such as unknown physical parameters, computing by utilizing a trained ANN subject to robustness
unmodeled dynamics, and environmental disturbances, but characteristics. The novelty of the proposed solution relies on
inevitably suffer from another problem of explosion of control the fact that any prior knowledge and explicit information
complexity, which is the second challenge discussed in the of system model parameters and weighting factors are not
following. required, thereby resulting in considerable enhancement of
To solve this, most research and applications have been robustness, and meanwhile, the computational demand may
focused on strongly simplified formulations. Among them, the be significantly decreased by utilizing the machine learning
development of artificial intelligence (AI) has motivated the technique instead of a complex algorithm. Finally, extensive
exploration of more sophisticated strategies to control power simulative and experimental investigations for MMC control
converters [23]. Recently, a combination of machine learning system validate the interest and viability of the proposed
with conventional FCS-MPC scheme has been singled out as design methodology.
an interesting control solution [24]. As such, an artificial neural We contribute three main points to the relevant literature.
network (ANN), trained offline by the data collected from the • This is the first time, to the best of the author’s knowl-
designed original controller, is deployed, which serves as a edge, as one potential tool of representing the system un-
surrogate controller that can provide low control complexity. certainties, a predictor-based NN is introduced into data-
Meanwhile, a similar idea is presented to imitate the FCS- driven model-free adaptive predictive control solution to
MPC controller with multistep prediction horizons [25]. Next, identify the unknown nonlinear system dynamics and un-
in order to alleviate the issue of fussy selection of weighting certainties, allowing the improvement of robustness in the
factors in MPC cost function, the authors in [26] proposed presence of parametric uncertainties and environmental
a novel weighting factors design solution with the use of an disturbances.
ANN. Based on this idea, an AI technique is further developed • Another important contribution of this article is the de-
for real-time online determination of the weighting factors that velopment of an efficient way to deal with the complex
regulates the average switching frequency (SF), independently algorithm by deploying the machine learning technique
of the operating point [27]. In this sense, the weighting factors subject to robustness characteristics, this allows for a
are automatically updated in real-time, thus allowing the reduction in the required online computation and a more
system to quickly adapt to any reference step changes. After abstract level of investigation, which opens a very attrac-
that, an ANN trained to approximate the relationship between tive area for future research.
circulating current references and capacitor voltage ripples is • The third one is that the proposal is quite general, since
studied under unbalanced grid faults [28]. Although the afore- a prominent merit of the suggested controller design
mentioned solutions by deploying the presented AI technique is that it can be naturally extended toward modern
can effectively reduce the computation burden of real-time power converters and drives systems without any prior
control and allow a satisfactory system behavior grade, the limitations. In view of this merit, other researchers can
uncertainties of system has not been paid sufficient attention. incorporate this solution into their particular application
In particular, the trained surrogate controller may suffer from outperforming previous published control solutions.

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

v −v
C. Outline of the Article where vgx = glx 2 gux denotes the possible output voltage of
For the purpose of clear methodology, this article is struc- MMC, Rz = Rg + 0.5Rarm , and Lz = Lg + 0.5Larm .
tured in the following section. Section II briefly discusses The discrete-time model of (2) with respect to the ac-side
the background about the standard FCS-MPC approach of current can be obtained by forward Euler approximation
MMC. Section III describes the proposed predictor-based Ts Rz Ts
data-driven model-free adaptive predictive control architecture ig (k + 1) = (1 − )ig (k) + vg (k) (3)
Lz Lz
for MMC. Furthermore, section IV provides a performance
where ig (k + 1) denotes the predicted grid-side current, and
evaluation and compares five different control solutions, while
Ts denotes the sampling period.
conclusions and future work are outlined in section V.
For clarify, in this literature, the presentable output voltage-
level is addressed as finite control set (FCS). Thus, they can be
II. FCS-MPC M ETHOD OF MMC evaluated by deploying a predefined cost function that reacts
The circuit configuration of a three-phase MMC topology the performance. It yields
is depicted in Fig. 1, where each phase of the MMC consists 1
of 2 arms, e.g., an upper arm (denoted by subscript, “u”) and vg = Udc Ng (4)
2N
a lower arm (denoted by subscript, “l”). Each arm includes
where
N series connected half-bridge submodule (SM), one arm
filter inductor (denoted by Larm ), and one arm filter resistor Ng = {−N, −(N − 2), ..., 0, ..., N − 2, N }
(denoted by Rarm ). Each SM can provide two voltage levels,
and where Ng denotes the presentable output voltage-level. N
e.g., zero or Ucgxi , Ucgxi (x = {a, b, c}, i = {1, 2, ..., 2N })
denotes the SM number per arm.
denotes SM capacitor voltage. They depend on the state of
two complementary power switches Sgx1 and Sgx2 . Furthermore, assuming that the number of arms in the SM
is N , and then, the SM capacitor nominal voltage is UNdc , the
converter output voltage-level reference Ng∗ can be expressed
as (5), depending on the number of SM on-state per phase.
 ∗
ngal − n∗gau ∗
= Nga




 2
 n∗ − n∗

gbl gbu ∗
= Ngb (5)

 2
∗ ∗
n − ngcu


 gcl ∗

 = Ngc
2
and  ∗ ∗
ngal + ngau = N

∗ ∗
ngbl + ngbu = N (6)
 ∗
 ∗
ngcl + ngcu = N
where n∗gau and n∗gal denote the SM on-state numbers of up
and low-arms in phase-a. n∗gbu and n∗gbl denote the SM on-
Fig. 1. Common topology of the three-phase MMC.
state numbers of up and low-arms in phase-b. n∗gcu and n∗gcl
denote the SM on-state numbers of up and low-arms in phase-
First of all, according to Kirchhoff’s law, the upper and
c. The on-state numbers of SM in upper and lower arms can
lower arm voltage of phase-x with respect to the dc-link mid-
be obtained by the two aforementioned equations (5) and (6).
point can be deduced as [29]
Next, the continuous-time model of SM capacitor voltage
U di di

 dc = vgux + Larm ux + iux Rarm + igx Rg + Lg gx
 is given as
2 dt dt Z t
U di di 1
−
 dc
= −vglx − Larm
lx
− ilx Rarm + igx Rg + Lg
gx Ucgxi (t) = Ucgxi (0) + icgxi (τ )dτ (7)
2 dt dt Cg 0+
(1)
where Udc denotes the dc-link voltage, vgux and vglx denote where Ucgxi (0) represents the initial value of SM capacitor
the upper and lower arm voltages of phase-x, respectively. igx voltage, Cg represents the SM capacitor, and icgxi represents
denotes the ac-side current of phase-x, igx = iux − ilx . iux the current flowing through the SM capacitor (i = 1, · · ·, 2N ).
and ilx denote the upper and lower arm currents of phase- The discrete form of (7) can be described as
x, respectively. Lg denotes the output filter inductor, and Rg Ts
Ucgxi (k + 1) = Ucgxi (k) + icgxi (k). (8)
denotes the output resistor. Cg
Based on (1), the ac-side current dynamic of phase-x can In cost function design of the conventional FCS-MPC ap-
be illustrated as proach, two different control objectives are addressed in terms
digx Rz 1 of ac-side current tracking and SM capacitor voltage tracking,
= − igx + vgx (2)
dt Lz Lz and then, all these terms are integrated into a complete cost

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

function. It yields Cost Function Design Sorting Technique Plant Model

fc = λ1 fc1 + λ2 fc2 (9)

Sytem Control Law Neural Predictor
fc1 = i∗g (k + 1) − ig (k + 1)



 " 2N #
X X
∗ (10) Parameter
f = Ucgx − Ucgxi (k + 1) Estimation Law
 c2

 Neural Update Law
x=a,b,c i=1

where Ucgxi (k+1) denotes the predicted SM capacitor voltage, NN

∗ ∗ ∗
and Ucgx = Udc /N is the capacitor voltage reference. Udc
∗
denotes the dc-link voltage reference. ig (k + 1) denotes the Fig. 2. Schematic representation of the proposed predictive control
predicted ac-side current reference. λ1 and λ2 denote the architecture.
weighting factors for the two terms, respectively.
It should be noted that the determination procedure of these
weighting factors becomes tedious due to the multivariable implies that a generic controller can be designed and developed
nonlinear relationship between the converter outputs and their independently for power converters. In the design procedure
corresponding references [1], [5], [7]. More precisely, the of the suggested controller, it does not require the information
cost function includes several control objectives (variables) and knowledge of detailed model of the controlled plant but
simultaneously. The relative importance of one objective over only input-output measurement data. Hence, this proposal has
the other can be set through the weighting factors. Control strong robustness which is the key requirement of power
variables possess different physical natures, and these vari- converter control systems. Alternatively, in our work, the key
ances result in coupling effects. However, the optimization of feature of this proposal is that the undesired effects of the
the weighting factors can be a very challenging task for cost uncertainties and the complex multi-objective optimization
functions with multiple objectives that are in conflict or can procedure can be explicitly dealt with, without involving accu-
not be unified. As a consequence, the numerical procedure for rate modeling information and weighting factor combinations,
the weighting factor selection is still an open research topic while guaranteeing adaptability to unknown nonlinear system
[30]. Due to the lack of theoretical design procedure, some dynamics, model variations and environmental disturbances.
guidelines based on an empirical method to obtain appropriate Finally, the schematic representation of the proposed predictive
weighting factors are reported in [31]. Other solutions include control architecture for MMC is shown in Fig. 2.
the optimization of cost function based on NN [32], fuzzy
logic control-based auto-tuning strategy for online selection
A. The Local Compact Form Dynamic Linearization
of the weighting factors [33], and multi-objective optimization
(Local-CFDL) Model
based on a ranking strategy [34].
Furthermore, in real-world applications, uncertainty dynam- First of all, in this subsection, we take a discrete-time
ics and disturbances widely exist in power converter systems, nonlinear system into account
and it is difficult to obtain the knowledge of the accurate
model. Specifically, due to the model mismatch, the predicted y(k + 1) = f (y T (k), uT (k), dT (k)) (11)
future behavior of the controlled variables becomes inaccu- where y(k) = [y(k), · · ·, y(k − ny )]T ∈ <ny +1 , u(k) =
rate. In that sense, this implies that the selected switching [u(k), · · ·, u(k − nu )]T ∈ <nu +1 , and d(k) = [d(k), · ·
configuration with the conventional scheme is not the optimal ·, d(k − nd )]T ∈ <nd +1 . y(k) and u(k) denote the output
one for the next sampling period in the presence of parameter and input of controlled system, respectively. ny , nu , and nd
mismatch, which deteriorates the performance of the control denote the unknown system orders. d(k) denotes the bounded
system or even cause instability. Thus, obtaining the system disturbance. f (·) denotes an unknown nonlinear function.
model knowledge that is accurate enough is a challenge in such Following reference [13], on the basis of two assumptions
complicated power converter systems. As a consequence, we and theorem 1, one has
point out that it is more important and meaningful to deploy
the large amount of the process data produced by the system so ∆y(k + 1) = f (y T (k), uT (k), dT (k)) (12)
as to fulfill the control demand of power converters. Motivated T T
− f (y (k), [u(k − 1), u(k − 2), · · ·, u(k − nu )], d (k))
by aforementioned issues, a way to get around these issues is + f (y T (k), [u(k − 1), u(k − 2), · · ·, u(k − nu )], dT (k))
to explore a novel predictive control methodology, and it will
∂f
be demonstrated in detail as follows. − f (y T (k − 1), uT (k − 1), dT (k − 1)) = ∆u(k)
∂u(k)
∂f ∂f
III. T HE P ROPOSED P REDICTOR -B ASED DATA -D RIVEN + ∆y(k) + · · · + ∆y(k − ny )
M ODEL -F REE A DAPTIVE P REDICTIVE C ONTROL ∂y(k) ∂y(k − ny )
∂f ∂f
In this section, in order to estimate the lumped disturbances, + ∆u(k − 1) + · · · + ∆u(k − nu )
including the model uncertainties and unmodeled dynamics, a ∂u(k − 1) ∂u(k − nu )
novel predictor-based data-driven model-free adaptive predic- ∂f ∂f
+ ∆d(k) + · · · + ∆d(k − nd )
tive control methodology for MMC is first proposed, which ∂d(k) ∂d(k − nd )

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Sorting Technique Plant Model Sorting Technique Plant Model

Supervised Predictor-Based Data-Driven Model-Free

Adaptive Predictive Control Architecture Measurement Measurement
Learning

Hidden Layer Hidden Layer

Trained ANN Offline Trained ANN Online

Output Layer Input Layer Output Layer Input Layer

Output Node Neuron Input Node Output Node Neuron Input Node

Input Node Input Node

Output Node Output Node

Fig. 3. An overview of the proposed data-driven model-free adaptive predictive control architecture using machine learning technique.

where ∆y(k) = y(k) − y(k − 1), ∆u(k) = u(k) − u(k − 1), the control process. In this sense, the performance on a
and ∆d(k) = d(k) − d(k − 1). nonlinear uncertain system in (11) would be very limited. As
Next, we define a consequence, the control architecture should be designed to
 be robust with respect to system uncertainties.
 Φ(k) = ∂f /∂u(k)

ξ(k) = f (y T (k), [u(k − 1), u(k − 2), · · ·, u(k − nu )], Remark 2: In this literature, compared with the standard

dT (k)) − f (y T (k − 1), uT (k − 1), dT (k − 1)).

FCS-MPC approach, a remarkable feature of this proposal
(13) is that the effect of parameter uncertainty and unmodeled
Then, system (12) can be transformed into the following dynamics is considered as lumped disturbance first, and then
form: the function approximation technique combined with the sug-
∆y(k + 1) = Φ(k)∆u(k) + ξ(k) (14) gested control framework is applied to suppress its influence.
To be specific, in the following design procedure of the
where ξ(k) denotes the unknown mismatched perturbation suggested controller, we will employ the NNs to approximate
accounting for unknown modeling uncertainties and input the nonlinear uncertainties term ξ(k), while guaranteeing
disturbances. adaptability to different operating conditions.
Consider an objective function as
J(u(k)) = |y ∗ (k + 1) − y(k + 1)|2 + λ|u(k) − u(k − 1)|2 C. Predictor-Based Neural Network Estimator Design
(15) Based on (2), the ultra-local mathematical model of MMC
where λ denotes the weighting factor. yields
ẋ = f (x) + bu (18)
B. Data Driven-Based Model-Free Adaptive Controller T T
Design where x = igx = [iga , igb , igc ] , u = vgx = [vga , vgb , vgc ] ,
f (x) = − R 1
Lz igx , and b = Lz .
z

Based on (15), the control law can be described by the Note that, for a given continuous nonlinear function f (x),
following form it can be approximated by a NN with an activation function
ρΦ(k) (y ∗ (k + 1) − y(k) − ξ(k)) β(x) [15]. Then, one has
u(k) = u(k − 1) + (16)
λ + |Φ(k)|2 f (x) = Ξ∗ T β(x) + ε (19)
where ρ denotes a step factor.
Due to the fact that Φ(k) in (16) is unknown, its estimation where Ξ∗ represents the optimal weight, satisfying kΞ∗ kF ≤
law yields [14] Ξ∗ with Ξ̄∗ ∈ < being positive constants. ε represents the
minimum approximation error. ||ε|| ≤ ε̄ with ε̄ ∈ < being
Φ̂(k) = Φ̂(k − 1) (17) positive constants.
Through the NN approximation, the formula (18) can be
 
η∆u(k − 1) ∆y(k) − Φ̂(k − 1)∆u(k − 1) − ξ(k − 1)
+ reconstructed as
µ + ||∆u(k − 1)||2
ẋ = Ξ∗ T β(x) + bu + ε. (20)
where µ is a weighting factor, and η > 0 is a positive
constants. To design an update law for Ξ̂, consider a state predictor as
Remark 1: Based on the above observations, it is
x̂˙ = Ξ̂T β(x) + bu − k1 x̃ (21)
noteworthy that, in practical perspective, one of the well-
known challenges that affects the performance of power where k1 ∈ < is positive constants. x̃ = x̂ − x. x̂ denotes the
converter control systems is represented by parametric observed system state. Ξ̂ is the estimate of Ξ∗ that is updated
uncertainties. The uncertainties occur not only due to the as follows:
sensitivity to environmental operating conditions but also ˙
due to the presence of external unknown disturbances in Ξ̂ = −Γ1 [β(x)x̃T + kw1 Ξ̂] (22)

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

where Γ1 ∈ < and kw1 ∈ < are positive constants, respec-

tively.
By applying the approximate Euler formula, the discrete-
time model of (19), (21), and (22) can be described by
T

x̂(k + 1) = x̂(k) + Ts [Ξ̂ (k)β(x(k)) + bu(k) − k1 x̃(k)]

Fig. 4. General topology of the 10-neuron hidden layer feed-forward
Ξ̂(k + 1) = (1 − Γ1 Ts kw1 )Ξ̂(k) − Γ1 Ts β(x(k))x̃T (k) ANN.

ˆ
f (k + 1) = Ξ̂T (k + 1)β(x(k + 1)).
(23)
Finally, following the detailed analysis in previous sections, To this end, a supervised imitation learning technique using
the proposed neural predictor-based data-driven model-free data labeled is developed to imitate the known suggested
adaptive control methodology is constructed by incorporating controller [25]. In this design, an imitator, trained offline by
control law (16), parameter estimation law (17), and neural using the data collected from the proposal, is presented to
predictor (23) together. We have that control the MMC with high accuracy subject to robustness
   characteristics. As such, this feature motivates a much simpler


 ρΦ̂(k) y ∗ (k + 1) − y(k) − fˆ(k) robust predictive control solution, which is convenient to

 u(k) = u(k − 1) + implement in applications. Thus, by this proposal, the online
λ + |Φ̂(k)|2




 implementation of much more complex predictive control
Φ̂(k) = Φ̂(k − 1)




 strategies is made possible, and as it will be elaborated in
the following subsection. Alternatively, in order to achieve
  
η∆u(k − 1) ∆y(k) − Φ̂(k − 1)∆u(k − 1) − fˆ(k − 1)

+ the SM capacitor voltage balancing, a conventional sorting
µ + ||∆u(k − 1)||2




 approach is employed in our work. Due to the page limit,
x̂(k + 1) = x̂(k) + Ts [Ξ̂T (k)β(x(k)) + bu(k) − k1 x̃(k)] comprehensive explanation on the capacitor voltage balancing




 control for MMC is presented in [35].
Ξ̂(k + 1) = (1 − Γ1 Ts kw1 )Ξ̂(k) − Γ1 Ts β(x(k))x̃T (k)





ˆ
f (k + 1) = Ξ̂T (k + 1)β(x(k + 1))

105
Error Histogram with 20 Bins
(24) 4.5 Training
where Φ̂(k) = Φ̂(1) if ||Φ̂(k)|| ≤ τ or ||∆u(k − 1)||2 ≤ τ 4
Validation
Test

or sign(Φ̂(k)) 6= sign(Φ̂(1)). τ is a small positive constant. 3.5

Zero Error

Φ̂(1) denotes the initial value of Φ̂(k). 3

Instances

2.5

D. The Design of Cost Function 1.5

Another interesting investigation aspect, apart from the 0.5

challenge associated with the unknown uncertainties, is the 0

0.07541

0.2262

0.3771

0.5279

0.6787

0.8295

0.9803

1.131

1.282

1.433
-0.07541
-0.9803

-0.8295

-0.6787

-0.5279

-0.3771

-0.2262
-1.433

-1.282

-1.131

cumbersome tuning of weighting factors. To overcome this, a

possible path to this long-standing research issue is to design Errors = Targets - Outputs

a new cost function for FCS-MPC without weighting factors. Fig. 5. The NN training error histogram.
Specifically, it consists of transforming the cost function to an
equivalent optimization issue where the variables involved are Best Validation Performance is 0.037896 at epoch 999
101
an equivalent output voltage reference and the possible output Train
Validation
voltage vectors. Motivated by the above observations, it can Test
Best
Mean Squared Error (mse)

be expressed as
100
fcnew = |u∗ (k + 1) − u(k + 1)| . (25)
This suggested design offers simplicity and great flexibility
for a designer to incorporate several control objectives by 10-1

simply changing the optimization criteria without appreciably

affecting performance. It is worthwhile to remark here that,
although, the proposal can be obtained and implemented easily 10-2
0 100 200 300 400 500 600 700 800 900 1000
only by using input-output data through the data-driven control 1000 Epochs

theory, a challenge presented by this proposal is that the Fig. 6. The NN training performance.
prohibitive computational cost will be increased due to the
complicated solving process of algorithm, which limits the
practical implementation to the control of MMC for real-time E. Machine Learning Technology
execution. Hence, it is to be expected that a computationally Machine learning technology plays a key role in modern
efficient optimization operation mechanism should be estab- power electronics control system. In this technology, ANN
lished in overall control system framework. is an extremely flexible computational model that can be

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Fig. 9. Experimental platform of three-phase MMC system. (a) NI

PXI testing platform for the MMC system; (b) PC; (c) three-phase
MMC system; (d) an SM overview.

TABLE I: SYSTEM AND CONTROL PARAMETERS

Fig. 7. The NN training regression. Parameter Symbol Simulation Experiment

Output resistance Rg 0.01Ω 0.01Ω

Gradient = 0.00078861, at epoch 1000 Output inductance Lg 10mH 10mH
-8
Arm resistance Rarm 0.2Ω 0.2Ω
gradient

10

Arm filter inductance Larm 10mH 10mH

SM capacitor Cg 2200µF 2200µF
10-9 DC-link reference voltage Udc 450V/500V 450V/500V
Mu = 1e-06, at epoch 1000 Sampling/control period Ts 100µs 100µs
10-8
Number of SMs per arm N 4 4
mu

10-9
Validation Checks = 1, at epoch 1000
10-7 pattern recognition network. This algorithm typically requires
val fail

10
-8 more memory but less time. Hence, it is much faster than
standard training algorithms. Furthermore, the performance
10-9
1000 Epochs of the training can be observed in Figs. 5-8. In Fig. 5, NN
Fig. 8. The NN training state. training error histogram is given, and NN training performance
is shown in Fig. 6. As depicted, the mean squared error at
epoch 920 converges to its stable, and it can be seen that the
best validation performance is 0.037896 at epoch 999. Next,
optimized to learn input-to-output mappings based on his-
NN training regression and NN training state are depicted in
torical data [23], [24], [36], [37]. Motivated by the previous
Figs. 7-8, where epoch denotes the training time. Mu denotes
observations, in this subsection, a supervised imitation learning
the momentum constant or momentum parameter which is
approach to approximate the system behavior as a black box
included in weight update expression to avoid the issue of
is introduced into the proposed design so as to improve
local minimum. Its range is between 0 and 1. Validation
the computationally efficient, which is depicted in Fig. 3.
checks denotes the generalization capability check of NNs.
Specifically, a feedforward ANN architecture is addressed as a
From the results obtained, these trained performance metrics
computational approach which allows description of complex
show that the surrogate controller is capable of imitating an
system behavior in a relatively simple model, and is shown
approximation of the system behavior of the proposal on a
in Fig. 4, where the ANN consists of 6 input neurons, 10
very high level. Due to the fact that its execution time is
hidden neurons, and 6 output neurons. To be specific, 6 input
independent of the complexity, and consequently, it can be
neurons are the measured three-phase ac-side currents (iga , igb ,
inferred that the online implementation of much more complex
and igc ) and their references (i∗ga , i∗gb , and i∗gc ), respectively.
predictive control strategies can be achieved by deploying
Meanwhile, 6 output neurons are the SM on-state numbers of
the machine learning technology, while yielding satisfactory
up and low-arms in abc three phase n∗gau , n∗gal , n∗gbu , n∗gbl ,
performance.
n∗gcu and n∗gcl , respectively. Thus, the ANN is trained by the
input-output data to obtain the mapping relationship.
Next, in order to train this network, a neural pattern IV. P ERFORMANCE E VALUATION
recognition app is deployed in Matlab/Simulink to serve this In this section, we verify the validity of the proposed
purpose. In that sense, a total of 70% randomly divided solution on an efficient predictive control regulated MMC
input data is employed, and 15% for validation and 15% case study that is described in detail in Section III. In our
for testing. Meanwhile, a Levenberg-Marquardt training al- work, for the implementation of this test, the real MMC
gorithm is selected to update the weights and biases in the prototype using Silicon Carbide Metal-Oxide-Semiconductor

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

40 30 10 150

0 0 0 125

THD=2.67%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.61%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.71%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.72%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.55%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]

Fig. 10. The performance evaluation with five different control approaches in steady-state. (a) Level-based FCS-MPC method. (b) ESO-based
FCS-MPC method. (c) Deadbeat predictive control method. (d) Fast MPC method. (e) Proposed FCS-MPC solution.

Field-Effect Transistors (SiC-MOSFETs) C3M0065100K with 13, where the THD values of the output currents are 4.60%,
the circuit parameters listed in TABLE I is controlled by 3.30%, 4.64%, 5.68%, and 3.18% under the measurement
employing a NI PXI platform, as shown in Fig. 9. For an noise disturbance condition, respectively. It is noteworthy that
effective implementation of the algorithm, a compensation the performance of level-based FCS-MPC, deadbeat predictive
technique that eliminates the delay introduced by the control control, and fast MPC methods is degraded seriously by noise.
algorithm is deployed [38]. Due to the fact that the current variations are highly sensitive
to measurement noise, while noise suppression is not taken
A. Simulation Result Analysis into account in these control approaches. This will be further
demonstrated in a real-world application in the presence of
To verify the proposed methodology, simulation results measurement noise. To alleviate this, in our work, a predictor-
obtained with an MMC subjected to five different control so- based NN is embedded into data-driven model-free adaptive
lutions are presented in this subsection. For a fair comparison, predictive control solution to identify the system dynamics and
the system parameters are selected as same. As depicted in uncertainties, allowing the improvement of robustness in the
Figs. 10(I)-13(I), it is shown that the control performance presence of parametric uncertainties and measurement noise.
of three-phase output currents with the five different con- As a result, by deploying this proposal, it can be observed
trol approaches under steady-state, transient-state, parameter that the performance degradation caused by the measurement
mismatch, and measurement noise disturbance conditions can noise can be alleviated, and the desired control performance
be obtained. As demonstrated, the total harmonic distortion can be achieved, as shown in Fig. 13(e). Consequently, it can
(THD) values of the output currents in system parameter be concluded that the proposed control methodology is capable
mismatch are 6.50%, 4.80%, 6.53%, 8.30%, and 3.89%, of accomplishing the control objective of sinusoidal grid-
respectively. Next, Figs. 10(II)-13(II) present the control per- side currents under measurement noise conditions. Finally, the
formance of the upper and lower arm currents in different op- harmonic spectrum with five different control strategies under
eration conditions. Furthermore, in Figs. 10(III)-13(III), it can the parameter match and mismatch (Lg = 7mH) conditions
be observed that the performance of three-phase circulating can be observed in Figs. 14-15. It should be noted that this
currents (ica , icb , and icc ) are given. Finally, it can be shown proposal can provide a good current quality, while keeping
in Figs. 10(IV)-13(IV) that the SMs capacitor voltages can be the low computational demand due to the introduction of the
controlled to track their corresponding references. machine learning.
On the other hand, the preceding design assumes that the
measured states are free of noise; however, this may not be
true in practice. In order to show the measurement noise prob- B. Robustness Performance Analysis
lem, we investigate the performance with measurement noise To further illustrate the robust performance of the proposed
conditions, and the noise is assumed to be Gaussian white methodology, a thorough analysis of the effect of parameter
in this literature. The investigated results are provided in Fig. mismatch in term of output filter inductor is evaluated. To

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

40 30 10 150

0 0 125
0

THD=2.67%
-40 -30 -10 100
4.95 5 5.05 4.95 5 5.05 4.95 5 5.05 4.95 5 5.05
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.61%
-40 -30 -10 100
4.95 5 5.05 4.95 5 5.05 4.95 5 5.05 4.95 5 5.05
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.71%
-40 -30 -10 100
4.95 5 5.05 4.95 5 5.05 4.95 5 5.05 4.95 5 5.05
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.72%
-40 -30 -10 100
4.95 5 5.05 4.95 5 5.05 4.95 5 5.05 4.95 5 5.05
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=2.55%
-40 -30 -10 100
4.95 5 5.05 4.95 5 5.05 4.95 5 5.05 4.95 5 5.05
Time [s] Time [s] Time [s] Time [s]

Fig. 11. The performance evaluation with five different control approaches in transient-state. (a) Level-based FCS-MPC method. (b) ESO-
based FCS-MPC method. (c) Deadbeat predictive control method. (d) Fast MPC method. (e) Proposed FCS-MPC solution.

40 30 10 150

0 0 0 125

THD=6.50%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=4.80%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=6.53%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=8.30%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=3.89%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]

Fig. 12. The performance evaluation with five different control approaches under the parameter mismatch (Lg = 7mH). (a) Level-based
FCS-MPC method. (b) ESO-based FCS-MPC method. (c) Deadbeat predictive control method. (d) Fast MPC method. (e) Proposed FCS-MPC
solution.

be specific, the performance analysis (THD, average SF, and robustness against the parametric variation. Next, under model
tracking error (TE)) between five different control methods is parameter mismatch condition, the average SF with the ESO-
investigated, corroborating the effectiveness of the proposal, as based FCS-MPC scheme is lower than that of the level-based
depicted in Figs. 16-18. As illustrated in these figures, com- FCS-MPC method and deadbeat predictive control as well as
pared with the state-of-the-art FCS-MPC techniques, the pro- the proposal, but it undergoes the higher TE (from Lg = 10mH
posed solution can mitigate performance degradation caused to Lg = 13mH). Furthermore, although acceptable SF in the
by the model parameter sensitivity issue, allowing for strong presence of parametric uncertainties can be obtained by the

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

40 30 10 150

0 0 0 125

THD=4.60%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=3.30%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=4.64%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=5.68%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]
40 30 10 150

0 0 0 125

THD=3.18%
-40 -30 -10 100
2 2.02 2.04 2 2.02 2.04 2 2.02 2.04 2 2.02 2.04
Time [s] Time [s] Time [s] Time [s]

Fig. 13. The performance evaluation with five different control approaches under the measurement noise disturbance. (a) Level-based FCS-
MPC method. (b) ESO-based FCS-MPC method. (c) Deadbeat predictive control method. (d) Fast MPC method. (e) Proposed FCS-MPC
solution.

5 5 5 5 5
Fundamental (50Hz) = 30.05 Fundamental (50Hz) = 30.98 Fundamental (50Hz) = 30.05 Fundamental (50Hz) = 30.04 Fundamental (50Hz) = 29.86
THD= 2.67%, SF=1231Hz THD= 2.61%, SF=1175Hz THD= 2.71%, SF=1231Hz THD= 2.72%, SF=1218Hz THD= 2.55%, SF=1233Hz
Lg=10mH Lg=10mH Lg=10mH Lg=10mH Lg=10mH

2.5 2.5 2.5 2.5 2.5

0 0 0 0 0
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200

Fig. 14. The harmonic spectrum with five different control approaches under the parameter match. (a) Level-based FCS-MPC method. (b)
ESO-based FCS-MPC method. (c) Deadbeat predictive control method. (d) Fast MPC method. (e) Proposed FCS-MPC solution.

5 5 5 5 5
Fundamental (50Hz) = 30.06 Fundamental (50Hz) = 31.07 Fundamental (50Hz) = 30.07 Fundamental (50Hz) = 31.14 Fundamental (50Hz) = 30
THD= 6.50%, SF=1278Hz THD= 4.80%, SF=1209Hz THD= 6.53%, SF=1278Hz THD= 8.50%, SF=1070Hz THD= 3.89%, SF=1276Hz
Lg=7mH Lg=7mH Lg=7mH Lg=7mH Lg=7mH

2.5 2.5 2.5 2.5 2.5

0 0 0 0 0
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200

Fig. 15. The harmonic spectrum with five different control approaches under the parameter mismatch (Lg = 7mH). (a) Level-based FCS-MPC
method. (b) ESO-based FCS-MPC method. (c) Deadbeat predictive control method. (d) Fast MPC method. (e) Proposed FCS-MPC solution.

TABLE II: THE COMPARISON RESULTS BETWEEN FIVE DIFFERENT CONTROL SCHEMES

Execution Time (µs) THD (%) Tracking Error (%)

Solution
Number of SM Optimization Sorting Sum Match/Mismatch Match/Mismatch
Level-based FCS-MPC [29] N=4 11.748 18.939 30.687 2.67/6.5 2.98/6.09
ESO-based FCS-MPC [39] N=4 55.812 22.386 78.198 2.61/4.8 4.27/5.17
Deadbeat Predictive Control [40] N=4 11.88 19.929 31.809 2.71/6.53 2.98/6.09
Fast MPC [35] N=4 12.972 21.636 34.608 2.72/8.5 3.19/7.63
Proposed FCS-MPC N=4 4.156 19.776 23.932 2.55/3.89 3.38/4.26

fast MPC approach, it suffers from the highest TE, which may ferent control strategies is conducted, and the THD, execution
be unacceptable for power converter application under output time, and state-steady TE of different control methods are also
filter inductor Lg mismatch (from Lg = 7mH to Lg = 9mH). investigated, and as shown in TABLE II. From TABLE II, note
Finally, a thorough performance evaluation between five dif- that, by deploying the machine learning technique, the online

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

1 0 1 4 0 0 1 0
L e v e l- b a s e d F C S -M P C L e v e l- b a s e d F C S -M P C L e v e l- b a s e d F C S -M P C
9
)

E S O -b a s e d F C S -M P C E S O -b a s e d F C S -M P C E S O -b a s e d F C S -M P C

( H z )
c u r r e n t ( %

8 D e a d b e a t P r e d ic tiv e C o n tr o l D e a d b e a t P r e d ic tiv e C o n tr o l 8 D e a d b e a t P r e d ic tiv e C o n tr o l

1 3 0 0 F a st M P C F a st M P C

)
F a s t M P C
7

fr e q u e n c y
P ro p o s e d F C S -M P C P ro p o s e d F C S -M P C

e r r o r ( %
P ro p o s e d F C S -M P C
6 6
1 2 0 0
o f g r id - s id e

T r a c k in g
4 4

S w itc h in g
3
1 1 0 0
2 2
T H D

1
0 1 0 0 0 0
7 8 9 1 0 1 1 1 2 1 3 7 8 9 1 0 1 1 1 2 1 3 7 8 9 1 0 1 1 1 2 1 3
T h e o u tp u t filte r in d u c to r ( m H ) T h e o u tp u t filte r in d u c to r ( m H ) T h e o u tp u t filte r in d u c to r ( m H )

Fig. 16. Evolution of the THD with five Fig. 17. Evolution of the average SF with Fig. 18. Evolution of the TE with five
different control solutions under output five different control solutions under out- different control solutions under output
filter inductor Lg mismatch. put filter inductor Lg mismatch. filter inductor Lg mismatch.

Fig. 19. Experimental results with the proposed FCS-MPC solution. (a) The three-phase ac-side output currents in steady-state. (b) The upper
and lower arm currents as well as SM capacitor voltages in steady-state. (c) The three-phase ac-side output currents in transient-state. (d)
The upper and lower arm currents as well as SM capacitor voltages in transient-state.

Fig. 20. Experimental results with the proposed FCS-MPC solution under the parameter mismatch (Lg = 7mH). (a) The three-phase ac-side
output currents in steady-state. (b) The upper and lower arm currents as well as SM capacitor voltages in steady-state. (c) The three-phase
ac-side output currents in transient-state. (d) The upper and lower arm currents as well as SM capacitor voltages in transient-state.

operating efficiency of the proposed control methodology is parametric uncertainties. Hence, it is noteworthy that the
improved by up to 64.62%, 92.55%, 65.02%, and 67.96% satisfactory control performance can be achieved by the pro-
during optimization (N=4) in comparison to the other four posed FCS-MPC solution, while guaranteeing adaptability to
control methods at the same sampling and control period, different conditions. Finally, this test proves the capability of
respectively. In the next subsection, the experimental tests are the proposed FCS-MPC solution to enhance the system ro-
elaborated in more detail. bustness, and meanwhile, these results state that this proposal
works as expected.
C. Experimental Result Analysis
In this subsection, the results of experimental evaluation for V. C ONCLUSION AND F UTURE W ORK
this proposal are presented in Figs. 19-20. Fig. 19(a)-19(d) A fundamentally different approach to the implementation
depicts the obtained results in terms of three-phase output of a novel model-free adaptive predictive controller based on
currents and the upper/lower arm currents as well as the SMs AI technology is proposed in this article, which the merits
capacitor voltages in steady-state and transient-state operation of FCS-MPC and machine learning can be consolidated.
under model parameter match condition. As can be seen from More importantly, the proposed methodology not only copes
Fig. 19, this proposal is capable of ensuring an acceptable with the uncertainty dynamics and unmodeled dynamics
level of performance optimization, while the arm currents and as well as disturbances but also guarantees the feasible
SMs capacitor voltages are also well regulated. Furthermore, online computational effort. Compared with the previous
as shown in Fig. 20, when the parameter mismatch occurs, the works on FCS-MPC strategies, the novelty of the proposal
performance of this suggested proposal is slightly degraded, is that it can be realized by introducing a supervised
which shows the good robustness performance against the imitation learning technique point of view into the neural

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

predictor-based data-driven model-free adaptive predictive [13] Z. S. Hou and S. T. Jin, “A novel data-driven control approach for a class
control architecture, facilitating the successful suppression of discrete-time nonlinear systems,” IEEE Trans. Control Syst. Technol.,
vol. 19, no. 6, pp. 1549-1558, Nov. 2011.
of performance deterioration caused by parameter variations [14] R. H. Chi, Y. Hui, S. H. Zhang, B. Huang, and Z. S. Hou, “Discrete-
and model inaccuracy and the improvement of optimization time extended state observer-based model-free adaptive control via local
operation mechanism. Finally, an MMC control system dynamic linearization,” IEEE Trans. Ind. Electron., vol. 67, no. 10, pp.
8691-8701, Oct. 2020.
is used as a case example to provide a comparison with [15] X. Liu, L. Qiu, W. J. Wu, J. E. Ma, Y. T. Fang, Z. H. Peng, and D.
predictive controllers, the numerical assessment test results Wang, “Predictor-based neural network finite-set predictive control for
have been provided to illustrate the validity of the proposed modular multilevel converter,” IEEE Trans. Ind. Electron., vol. 68, no.
11, pp. 11621-11627, Nov. 2021.
FCS-MPC architecture. [16] X. Liu, L. Qiu, W. J. Wu, J. E. Ma, Y. T. Fang, Z. H. Peng, and D.
Wang, “Event-triggered neural predictor-based FCS-MPC for MMC,”
In our future work, we will further focus on the study IEEE Trans. Ind. Electron., vol. 69, no. 6, pp. 6433-6440, Jun. 2022.
[17] X. Liu, L. Qiu, J. Rodrı́guez, W. J. Wu, J. E. Ma, Z. H. Peng,
of optimization efficiency and robustness cases of the FCS- D. Wang, and Y. T. Fang, “Neural predictor-based dynamic surface
MPC by deploying reinforcement learning. Alternatively, in predictive control for power converters,” IEEE Trans. Ind. Electron.,
this literature, attention is paid more to the analysis of compu- to be published, 2022, DOI: 10.1109/TIE.2022.3146643.
tational efficiency and robust performance of FCS-MPC with [18] X. Liu, L. Qiu, J. Rodrı́guez, W. J. Wu, J. E. Ma, Z. H. Peng, D. Wang,
and Y. T. Fang, “Data-driven neural predictors-based robust MPC for
uncertainties. Stability analysis is not thoroughly researched power converters,” IEEE Trans. Power Electron., vol. 37, no. 10, pp.
in our work; as a consequence, this topic is omitted from the 11650-11661, Oct. 2022.
discussion for the time being. It is expected to provide a more [19] M. Schenke, W. Kirchgässner, and O. Wallscheid, “Controller design for
electrical drives by deep reinforcement learning: A proof of concept,”
rigorous investigation and a milestone reference for further IEEE Trans. Ind. Informat., vol. 18, no. 7, pp. 4650-4658, Jul. 2020.
research and development of FCS-MPC for power converter [20] J. J. Duan, Z. H. Yi, D. Shi, C. Lin, X. Lu, and Z. W. Wang,
systems. “Reinforcement-learning-based optimal control of hybrid energy storage
systems in hybrid AC-DC microgrids,” IEEE Trans. Ind. Informat., vol.
15, no. 9, pp. 5355-5364, Sep. 2019.
R EFERENCES [21] H. Makhamreh, M. Trabelsi, O. Kükrer, and H. Abu-Rub, “A Lyapunov-
based model predictive control design with reduced sensors for a PUC7
[1] J. Rodrı́guez, M. P. Kazmierkowski, J. R. Espinoza, P. Zanchetta, H. rectifier,” IEEE Trans. Ind. Electron., vol. 68, no. 2, pp. 1139-1147, Feb.
Abu-Rub, H. A. Young, and C. A. Rojas, “State of the art of finite 2021.
control set model predictive control in power electronics,” IEEE Trans.
[22] X. Liu, L. Qiu, Y. T. Fang, K. Wang, Y. D. Li, and J. Rodrı́guez, “A
Ind. Informat., vol. 9, no. 2, pp. 1003-1016, May 2013.
fuzzy approximation for FCS-MPC in power converters,” IEEE Trans.
[2] X. Liu, L. Qiu, Y. T. Fang, Z. H. Peng, and D. Wang, “Finite-level-state
Power Electron., vol. 37, no. 8, pp. 9153-9163, Aug. 2022.
model predictive control for sensorless three-phase four-arm modular
[23] S. Zhao, F. Blaabjerg, and H. Wang, “An overview of artificial intelli-
multilevel converter,” IEEE Trans. Power Electron., vol. 35, no. 5, pp.
gence applications for power electronics,” IEEE Trans. Power Electron.,
4462-4466, May 2020.
vol. 36, no. 4, pp. 4633-4658, Apr. 2021.
[3] T. Dragičević, “Model predictive control of power converters for robust
[24] D. M. Wang, Z. J. Shen, X. Yin, S. Tang, X. F. Liu, C. Zhang, J.
and fast operation of AC microgrids,” IEEE Trans. Power Electron., vol.
Wang, J. Rodrı́guez, and M. Norambuena, “Model predictive control
33, no. 7, pp. 6304-6317, Jul. 2018.
using artificial neural network for power converters,” IEEE Trans. Ind.
[4] S. Vazquez, J. I. Leon, L. G. Franquelo, J. Rodrı́guez, H. A. Young, A.
Electron., vol. 69, no. 4, pp. 3689-3699, Apr. 2022.
Marquez, and P. Zanchetta, “Model predictive control: A review of its
applications in power electronics,” IEEE Ind. Electron. Mag., vol. 8, no. [25] M. Novak and T. Dragičević, “Supervised imitation learning of finite-
1, pp. 16-31, Mar. 2014. set model predictive control systems for power electronics,” IEEE Trans.
[5] X. Liu, L. Qiu, W. J. Wu, J. E. Ma, Y. T. Fang, Z. H. Peng, and D. Wang, Ind. Electron., vol. 68, no. 4, pp. 3402-3412, Apr. 2021.
“Neural predictor-based low switching frequency FCS-MPC for MMC [26] T. Dragičević, and M. Novak, “Weighting factor design in model
with online weighting factors tuning,” IEEE Trans. Power Electron., vol. predictive control of power electronic converters: An artificial neural
37, no. 4, pp. 4065-4079, Apr. 2022. network approach,” IEEE Trans. Ind. Electron., vol. 66, no. 11, pp.
[6] H. A. Young, M. A. Perez, and J. Rodriguez, “Analysis of finite-control- 8870-8880, Nov. 2019.
set model predictive current control with model parameter mismatch in a [27] S. Vazquez, D. Marino, E. Zafra, M. D. V. Pena, J. J. R. Andina, L. G.
three-phase inverter,” IEEE Ind. Electron., vol. 63, no. 5, pp. 3100-3107, Franquelo, and M. Manic, “An artificial intelligence approach for real-
May 2016. time tuning of weighting factors in FCS-MPC for power converters,”
[7] S. Vazquez, J. Rodrı́guez, M. Rivera, L. G. Franquelo, and M. No- IEEE Trans. Ind. Electron., vol. 69, no. 12, pp. 11987-11998, Dec. 2022.
rambuena, “Model predictive control for power converters and drives: [28] S. D. Wang, T. Dragičević, Y. Gao, S. K. Chaudhary, and R. Teodor-
advances and trends,” IEEE Trans. Ind. Electron., vol. 64, no. 2, pp. escu, “Machine learning based operating region extension of modular
935-947, Feb. 2017. multilevel converters under unbalanced grid faults,” IEEE Trans. Ind.
[8] M. Mehreganfar, M. H. Saeedinia, S. A. Davari, C. Garcia, and J. Electron., vol. 68, no. 5, pp. 4554-4560, May 2021.
Rodrı́guez, “Sensorless predictive control of AFE rectifier with robust [29] J. Moon, J. Gwon, J. Park, D. Kang, and J. Kim, “Model predictive
adaptive inductance estimation,” IEEE Trans. Ind. Informat., vol. 15, no. control with a reduced number of considered states in a modular
6, pp. 3420-3431, Jun. 2019. multilevel converter for HVDC system,” IEEE Trans. Power Del., vol.
[9] S. Kwak, U. C. Moon, and J. C. Park, “Predictive-control-based direct 30, no. 2, pp. 608-617, Apr. 2015.
power control with an adaptive parameter identification technique for [30] S. Kouro, M. A. Perez, J. Rodriguez, A. M. Llor, and H. A. Young,
improved AFE performance,” IEEE Trans. Power Electron., vol. 29, no. “Model predictive control: MPC’s role in the evolution of power
11, pp. 6178-6187, Nov. 2014. electronics,” IEEE Ind. Electron. Mag., vol. 9, no. 4, pp. 8-21, Dec.
[10] X. Liu, L. Qiu, W. J. Wu, J. E. Ma, Y. T. Fang, Z. H. Peng, D. Wang, 2015.
and J. Rodrı́guez, “Event-triggered ESO-based robust MPC for power [31] P. Cortés, S. Kouro, B. L. Rocca, R. Vargas, J. Rodrı́guez, J. I. Leon, S.
converters,” IEEE Trans. Ind. Electron., to be published, 2022, DOI: Vazquez, and L. G. Franquelo, “Guidelines for weighting factors design
10.1109/TIE.2022.3167135. in model predictive control of power converters and drives,” Proc. IEEE,
[11] Z. S. Hou and S. S. Xiong, “On model-free adaptive control and its 2009, pp. 1-7.
stability analysis,” IEEE Trans. Autom. Control, vol. 64, no. 11, pp. [32] O. Machado, P. Martı́n, F. J. Rodrı́guez, and E. J. Bueno, “A neural
4555-4569, Nov. 2019. network-based dynamic cost function for the implementation of a
[12] S. Wang, T. Dragicevic, G. F. Gontijo, S. K. Chaudhary and R. predictive current controller,” IEEE Trans. Ind. Informat., vol. 13, no.
Teodorescu, “Machine learning emulation of model predictive control 6, pp. 2946-2955, Dec. 2017.
for modular multilevel converters,” IEEE Trans. Ind. Electron., vol. 68, [33] X. Liu, D. Wang, and Z. H. Peng, “Cascade-free fuzzy finite control-
no. 11, pp. 11628-11634, Nov. 2021. set model predictive control for nested neutral point-clamped converters

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2022.3208594

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

with low switching frequency,” IEEE Trans. Control Syst. Technol., vol. Lin Qiu (Senior Member, IEEE) received the
27, no. 5, pp. 2237-2244, Sep. 2019. B.S. and Ph. D. degree in Electrical Engineering
[34] C. A. Rojas, J. Rodrı́guez, F. Villarroel, J. R. Espinoza, C. A. Silva, from the Department of Electrical Engineering
and M. Trincado, “Predictive torque and flux control without weighting at Tsinghua University in 2011 and 2017 re-
factors,” IEEE Trans. Ind. Electron., vol. 60, no. 2, pp. 681-690, Feb. spectively. He is currently a research professor
2013. with Zhejiang University-University of Illinois at
[35] Z. Gong, P. Dai, X. B. Yuan, X. J. Wu, and G. S. Guo, “Design and Urbana-Champaign Institute, Zhejiang, China.
experimental evaluation of fast model predictive control for modular His current research interests include trans-
multilevel converters,” IEEE Trans. Ind. Electron., vol. 63, no. 6, pp. portation electrification, grid resilience enhance-
3845-3856, Jun. 2016. ment as well as data-driven system control algo-
[36] S. Lucia, D. Navarro, B. Karg, H. Sarnago, and Ó. Lucı́a, “Deep rithm.
learning-based model predictive control for resonant power converters,”
IEEE Trans. Ind. Informat., vol. 17, no. 1, pp. 409-420, Jan. 2021.
[37] T. Dragičević, P. Wheeler, and F. Blaabjerg, “Artificial intelligence aided Youtong Fang (Senior Member, IEEE) received
automated design for reliability of power electronic systems,” IEEE the B.S. and Ph.D. degrees in electrical engi-
Trans. Power Electron., vol. 34, no. 8, pp. 7161-7171, Aug. 2019. neering from Hebei University of Technology,
[38] P. Cortés, J. Rodrı́guez, C. Silva, and A. Flores, “Delay compensation in Tianjin, China, in 1984 and 2001, respectively.
model predictive current control of a three-phase inverter,” IEEE Trans. He is currently a Professor with the College
Ind. Electron., vol. 59, no. 2, pp. 1323-1325, Feb. 2012. of Electrical Engineering, Zhejiang University,
[39] Y. C. Zhang, J. L. Jin, and L. L. Huang, “Model-free predictive current Hangzhou, China. His research interests include
control of PMSM drives based on extended state observer using ultra- the application, control, and design of electrical
local model,” IEEE Trans. Ind. Electron., vol. 68, no. 2, pp. 993-1003, machines.
Feb. 2021.
[40] J. Y. Wang, Y. Tang, P. F. Lin, X. Liu, and J. Pou, “Deadbeat predictive
current control for modular multilevel converters with enhanced steady-
state performance and stability,” IEEE Trans. Power Electron., vol. 35, José Rodrı́guez (M’81-SM’94-F’10-LF’20) re-
no. 7, pp. 6878-6894, Jul. 2020. ceived the Engineer degree in electrical engi-
neering from the Universidad Tecnica Federico
Santa Maria, in Valparaiso, Chile, in 1977 and
the Dr.-Ing. degree in electrical engineering from
the University of Erlangen, Erlangen, Germany,
Xing Liu (Senior Member, IEEE) received the in 1985. He has been with the Department
Ph.D. degree in Marine Electric Engineering of Electronics Engineering, Universidad Tecnica
from Dalian Maritime University, Dalian, China, Federico Santa Maria, since 1977, where he was
in 2018. From Dec. 2018 to Jan. 2019, he joined full Professor and President. Since 2015 to 2019
the Key Laboratory of Marine Technology and he was the President of Universidad Andres
Control Engineering of the Ministry of Communi- Bello in Santiago, Chile. Since 2022 he is President of Universidad San
cations at Shanghai Maritime University, Shang- Sebastian in Santiago, Chile. He has coauthored two books, several
hai, China. He is currently a Research Fellow book chapters and more than 700 journal and conference papers.
with the College of Electrical Engineering, Zhe- His main research interests include multilevel inverters, new converter
jiang University, Hangzhou, China. He is also topologies, control of power converters, and adjustable-speed drives.
currently a Visiting Scholar in the area of High- He has received a number of best paper awards from journals of the
Power Converters and Renewable Energy Generation with the Depart- IEEE. Dr. Rodriguez is member of the Chilean Academy of Engineering.
ment of Electrical Engineering, Tsinghua University, Beijing, China. His In 2014 he received the National Award of Applied Sciences and
main research interests include finite control-set model predictive con- Technology from the government of Chile. In 2015 he received the
trol, data-driven model-free adaptive control, adaptive optimal control, Eugene Mittelmann Award from the Industrial Electronics Society of the
and applications of artificial intelligence in industrial power electronics IEEE. In years 2014 to 2021 he has been included in the list of Highly
and systems. Cited Researchers published by Web of Science.

© 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: TU Delft Library. Downloaded on December 20,2022 at 13:27:00 UTC from IEEE Xplore. Restrictions apply.