Received: 1 August 2024

DOI: 10.1049/elp2.12523

ORIGINAL RESEARCH
- -
Revised: 18 October 2024 Accepted: 17 November 2024

- IET Electric Power Applications

Computationally efficient data‐driven model predictive control

for modular multilevel converters

Muneeb Masood Raja | Haoran Wang | Muhammad Haseeb Arshad |

Gregory J. Kish | Qing Zhao

Department of Electrical and Computer Engineering, Abstract

University of Alberta, Edmonton, Alberta, Canada
The application of model predictive control (MPC) for the control of modular multilevel
converters (MMCs) is widely explored because it offers flexibility in integrating multi-
Correspondence
Qing Zhao.
objective control and delivers superior dynamic response. Nonetheless, the increase in
Email: [email protected] computational complexity due to the rise in the number of submodules (SMs) is one of
the major drawbacks of this technique. This paper presents a finite control set model
predictive control (FCS‐MPC) that significantly reduces the computational complexity by
employing sparse identification of non‐linear systems (SINDy) to obtain a simplified
linear model for the MMC. The SINDy model reduces the complexity of performing the
prediction step by integrating input terms into the dynamics of load current and circu-
lating current. This simplifies the implementation compared to the conventional FCS‐
MPC approaches by eliminating the need to evaluate the voltage dynamics. The
computational burden is further reduced while maintaining 2N þ 1 voltage levels at the�
output by restricting the number of combinations for the inserted SMs to only N3 þ 1
instead of ðN þ 1Þ2 . A detailed comparison between the proposed technique and the
existing strategies demonstrates that the proposed technique offers a more computa-
tionally efficient solution for implementing FCS‐MPC on MMCs, while improving the
circulating current suppression due to more accurate predictions. Simulation and
experimental results are presented to validate the performance of the proposed approach.

KEYWORDS
computational complexity, data‐driven control, modular multilevel convertors, predictive control

1 | INTRODUCTION The MMCs are voltage source converters (VSCs) that

comprise of several submodules (SMs) connected in series with
Significant attention has been gained recently by modular each SM having a capacitor inside it. The control objective
multilevel converters (MMCs) due to their applications in high usually involves balancing the voltages across these capacitors to
voltage direct current (HVDC) [1, 2], motor drives [3], energy guarantee output voltage quality, as well as suppressing the
storage systems [4], power electronic transformers [5], and circulating current to ensure low power losses and
static synchronous compensator (STATCOM). The merits of maintain output current control. Several control strategies and
MMC include modularity, scalability, low switching losses, and switching frequency techniques have been proposed for MMCs,
flexible redundant operation [6]. Considering the vast range of including pulse width modulation (PWM) [7], nearest level
applications and the technical challenges to achieve reliable control [8], and selective harmonic elimination [9]. The con-
operation, it is necessary to develop efficient control strategies ventional PWM techniques developed for MMCs require addi-
for MMCs. tional control loops due to which the control complexity

-
This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial‐NoDerivs License, which permits use and distribution in any medium, provided the
original work is properly cited, the use is non‐commercial and no modifications or adaptations are made.
© 2024 The Author(s). IET Electric Power Applications published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.

IET Electr. Power Appl. 2024;1–16. wileyonlinelibrary.com/journal/elp2 1

 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2
- RAJA ET AL.

increases [10–12]. Also, the conventional control structure or three by using the nearest level based on the previous
provides limited dynamic response [13] and requires a modula- voltage level. This study was conducted with N þ 1 voltage
tion step to implement the control. It is becoming increasingly levels at the output and the control complexity is increased as a
popular to remove this modulation step [14]. look‐up table must be used carefully to select the control op-
On the other hand, finite control set model predictive con- tion based on N. Similarly, several other studies [26–28] have
trol (FCS‐MPC) eliminates the intermediate modulation step also been conducted in which the computational burden is
and achieves an improved transient response [15]. It predicts the lowered by using N þ 1 voltage levels. However, the output
future behaviour of the system by using the system model and THD performance and the circulating current suppression are
performs multiobjective optimisation while handling a non‐ diminished relative to the 2N þ 1 voltage level setup. To
linear constraint problem. For the case of MMCs, the predic- address this concern, a preselection algorithm was imple-
tive controller offers a simple control strategy, which can regulate mented to compute the switching possibilities by Gutierrez
the output AC current, suppress the circulating current, and and Kwak [29]. A modulated model predictive control
balance capacitor voltages [16, 17]. The FCS‐MPC generates the (MMPC) is proposed by the authors in refs. [30, 31] which uses
best possible insertion index, where the insertion index is the a two‐step duty cycle calculation instead of switching signals in
number of SMs to be inserted in the upper and lower arms. the FCS‐MPC. Most of the proposed techniques involving
Therefore, the need to employ the modulation block is elimi- reduced computational complexity assume that the SM
nated as the control inputs are directly provided to the system capacitor voltages are perfectly balanced and kept at a fixed
[18]. Irrespective of these features, the FCS‐MPC faces the reference. However, this is generally not the case due to which
dilemma of added computational burden as the number of SMs the accuracy of the prediction model and the control perfor-
rises. Besides that, due to the MPC prediction step relying heavily mance is affected. Furthermore, for each control option, the
on the accuracy of the system model, any assumptions and ap- arm voltages are evaluated by using the summation of all the
proximations considered while deriving the system dynamics can SM capacitor voltages. As the number of SMs increases, more
reduce the effectiveness of the FCS‐MPC due to reduced model terms are added to the summation causing an increase in the
accuracy. In such cases, model identification may be used to computational complexity due to an increased number of basic
improve the accuracy of the control technique [19]. operations to be performed, such as additions and
Numerous efforts have been made to implement a FCS‐ multiplications.
MPC for MMCs. The FCS‐MPC is proposed by the authors in Recently, a controller imitation for the FCS‐MPC has been
ref. [20] to control the output current, the circulating current, and proposed by the authors in refs. [32, 33] by using neural net-
works (NNs) to reduce the computational complexity of a
the capacitor voltages. It evaluates C 2NN control options during
standard FCS‐MPC. They proposed an emulation by training a
each period, where N is the number of SMs in each arm. Based
NN model using the data from a conventional FCS‐MPC.
on this approach, increasing the number of SMs in each arm
These techniques are independent of the complexity and
leads to a rapid increase in the number of control sets limiting the
execution time of the emulated FCS‐MPC. However, the
application of this technique due to the huge computational
emulation shows inferior performance as compared to the
burden. In ref. [21], a direct FCS‐MPC is employed that finds the
conventional FCS‐MPC. In ref. [34], the authors implemented
optimal switching after computing 22N control sets for each a model free adaptive control (MFAC)‐based FCS‐MPC
sample time. Despite the straightforward implementation, as the framework. An advantage of this technique is that the imple-
number of SMs increases, the number of control options in-
mentation relies only on input/output measurement data and
creases exponentially making it non‐viable to implement this
does not depend on the explicit model. However, it has certain
control technique on a large‐scale MMC system, such as with disadvantages, including complex tuning processes, higher
N ¼ 200 [22]. computational cost, and a slower response to rapid transients.
To resolve the issue of computational complexity associ- Several techniques have been used to obtain data‐driven
ated with the FCS‐MPC, several studies have been conducted. models such as autoregressive models [35, 36], NN, and
Moon et al. [23] presented a FCS‐MPC strategy with reduced reinforcement learning (RL) models [37–39]. Due to the
computational complexity by using three different cost func- availability of high processing power, NN‐based models are
tions for the control of AC output current, circulating current, gaining a lot of attention. Nevertheless, training these models
and SM capacitor voltage balancing. The SM capacitor voltage requires huge amounts of data and the obtained models lack
balancing control is decoupled from the cost function in the interpretability. Another main concern of NN‐based models is
indirect FCS‐MPC approach proposed by Vatani et al. [24] in overfitting. Furthermore, in the event of a sudden change in
which the voltage balancing is performed in a separate block. the system, the control scheme should quickly accommodate
This FCS‐MPC technique provides 2N þ 1 output voltage the changes, allowing a brief period for model discovery. The
levels while having ðN þ 1Þ2 control possibilities. Although the sparse identification of non‐linear dynamics (SINDy) frame-
indirect FCS‐MPC reduces the control options substantially as work [40] was introduced to minimise these problems. SINDy
compared to the direct FCS‐MPC, the switching combinations identifies parsimonious system models by solving a sparsity‐
are still high for MMC containing a large number of SMs. As promoting optimisation problem that requires very limited
an extension to the indirect FCS‐MPC, Gong et al. [25] pro- data. The resulting models are interpretable and do not face the
posed a fast MPC that reduces the control options to only two problem of overfitting.
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RAJA ET AL.
- 3

Motivated by the recent developments described above, this lower arm. An arm inductance La connects the upper and lower
paper presents a data‐driven FCS‐MPC that uses the SINDy arms. There are N SMs in each arm connected in series with each
algorithm to obtain a linear dynamic model of the MMC. The other. The half bridge topology is the most commonly used
prediction of the arm voltages is not required during the pre- topology which is why it is adopted in this research. An RL load
diction step of the MPC. This is guaranteed by integrating the having a resistance R and an inductance L is connected between
insertion index in the output current and circulating current the two arms of each leg. The arm currents are denoted by ijx ,
dynamic equations, hence ensuring a computationally efficient where the subscript x 2 fa; b; cg are the three phases and the
evaluation of the cost function for each control set compared to subscript j 2 fu; lg represents the upper and lower arms. Since
the existing techniques available in the literature. Additionally, all the phases of an MMC are identical, all the relationships are
since the prediction model does not involve computing the arm considered for a single phase and the subscripts are dropped for
voltages, the computational burden remains constant for any simplicity.
number of SMs. To further minimise the computational
complexity of the proposed scheme, the control sets are limited
�
to only N3 þ 1 instead of the ðN þ 1Þ2 switching states in case 2.2 | Mathematical model
of indirect FCS‐MPC while still maintaining 2N þ 1 output
voltage levels. The effect of changing the weights associated with The mathematical model of MMCs is derived using the
the insertion index terms in the cost function is also studied to Kirchhoff current law (KCL) and the Kirchhoff voltage law
improve the THD performance of the output current. Finally, (KVL). Applying KVL to each arm of a phase of MMC as
the circulating current suppression is improved by using the shown in Figure 1, the following dynamic equations are
proposed method compared to the conventional FCS‐MPC obtained.
approaches. The proposed scheme is implemented in both
simulation and experiment, in a real‐time fashion. The generated Udc diu ðtÞ dio ðtÞ
results depict the effectiveness of the proposed technique. The − uu ðtÞ − La ¼ Rio ðtÞ þ L ð1Þ
2 dt dt
rest of the paper is organised as follows. In Section 2, MMC
topology and the mathematical model are discussed. Section 3 Udc di ðtÞ dio ðtÞ
− ul ðtÞ − La l ¼ −Rio ðtÞ − L ð2Þ
describes the proposed SINDy framework for the MMC. In 2 dt dt
Section 4, the proposed FCS‐MPC formulation is described. The
proposed scheme is validated through the demonstration of where io is the output current, uu and ul denote the voltage in
simulation results in Section 5 and experimental results in Sec- the upper and lower arms, respectively, and iu and il are the
tion 6. Finally, the conclusion is drawn in Section 7. upper and lower arm currents, respectively. We can compute
the relationship between upper and lower arm currents and the
circulating current icir and output current io using the below
2 | SYSTEM MODELING OF MMC expression.

2.1 | MMC topology io ðtÞ ¼ iu ðtÞ − il ðtÞ ð3Þ

Figure 1 shows a three‐phase MMC consisting of three legs 1

icir ðtÞ ¼ ðiu ðtÞ þ il ðtÞÞ ð4Þ
connected in parallel with each phase having an upper and a 2

FIGURE 1 Circuit diagram of a MMC.

 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4
- RAJA ET AL.

From Equations (1–4), the continuous time dynamic model Um

of the output current and circulating current are given below: 2 3
u1 ðt1 Þ u2 ðt1 Þ … up ðt1 Þ
6 u1 ðt1 þ Ts Þ u ð t þ Ts Þ … up ðt1 þ Ts Þ 7
¼6 2 1 7
dio ðtÞ 1 4 ⋮ ⋮ ⋱ ⋮ 5
¼ ðu ðtÞ − uu ðtÞ − 2Rio ðtÞÞ ð5Þ
dt 2L þ La l u1 ðt1 þ mTs Þ u2 ðt1 þ mTs Þ … up ðt1 þ mTs Þ
ð10Þ
dicir ðtÞ 1
¼ ðU − ul ðtÞ − uu ðtÞÞ ð6Þ The next step involves evaluating a library of candidate
dt 2La dc
functions denoted by ΘðXm ; Um Þ. This library contains the
terms associated with the states and inputs. Polynomial and
Similarly, the dynamic model of the capacitor voltage across trigonometric terms are the most commonly used terms,
each SM is shown as follows: however, any other term or function can be used as a part of
the library. Generally, a simple choice is considered for the
duci Sij ij
¼ ð7Þ library of candidate functions such as polynomials, and more
dt C SM complex terms are added based on the requirement.
where C SM is the capacitance of each SM, the subscript After choosing a suitable library, a sparse regression
i ¼ 1; 2; 3; ::; N represents the SM number in each arm, ij is problem is solved to find the active non‐linearities by
the jth arm current, and Sij ¼ f1; 0g is the switching signal of computing the sparse coefficient vectors Ξ. The SINDy with
the corresponding SM. When Sij ¼ 1, the SM is connected, control generates the following dynamic model of the system:
whereas Sij ¼ 0 under the condition when the SM is bypassed.
̇
X ¼ ΞΘT ðXm ; Um Þ ð11Þ
̇
3 | SINDY MODEL WITH CONTROL where X is the time derivative of Xm and it can be evaluated
numerically or analytically. The columns of Ξ contain the terms
The SINDy algorithm generates the dynamic model of a system that are active in the system model. The optimisation problem
using noisy measurements by solving a sparsity‐promoting requires minimising the below cost function to get Ξ:
optimisation problem [40]. To incorporate the input terms in
the dynamics, a generalised form of SINDy introduced in ref. � �
1 ̇
[41] is presented to discover a reduced order linear mathematical ξi ¼ arg min kX i − ^ξi ΘT ðXm ; Um Þk22 þ λk^ξi k1 ð12Þ
2
model of MMCs.
Consider a non‐linear system represented by the following ̇ ̇
where λ is the hyperparameter, X i is the ith row of X , and ξi is
dynamic equation:
the ith row of Ξ. The sparsity is introduced due to the presence
of the 1‐norm term in the cost function.
d
xðtÞ ¼ f ðxðtÞ; uðtÞÞ ð8Þ
dt
3.1 | Proposed SINDy framework
where system state and input vectors are denoted by x and u, A dynamic model of MMC is developed using the SINDy
respectively, and f ðx; uÞ contains the dynamic equations of the approach, the details of which are explained below.
system having order n. In order to obtain the dynamic model
of a system using SINDy, the first step is the data collection in
which the system states x and inputs u are sampled at each time Algorithm 1. Model Identification of the MMC
step t1 ; t1 þ Ts ; t1 þ 2Ts ; …; t1 þ mTs and stacked in the form using SINDy.
of matrices as shown below:
1: Acquire system parameters io , icir , uu , ul ,
nu , and nl to generate Xm and Um
2: Evaluate the time derivative of Xm to get
Xm
2 3 X_ ¼ dX
dt
m

x1 ðt1 Þ x2 ðt1 Þ … xn ðt1 Þ 3: Select a suitable ΘðXm ; Um Þ and λ.

6 x1 ðt1 þ Ts Þ x2 ðt1 þ Ts Þ … xn ðt1 þ Ts Þ 7
¼6
4
7
5
4: Use sequential thresholding least square
⋮ ⋮ ⋱ ⋮ (STLSQ) [40] to solve the SINDy problem
x1 ðt1 þ mTs Þ x2 ðt1 þ mTs Þ … xn ðt1 þ mTs Þ 5: Acquire and save the acquired model
ð9Þ
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RAJA ET AL.
- 5

From the mathematical model of the MMC described in An advantage of using the SINDy approach is that accurate
Equations (5–6), it can be observed that the io and icir dy- system models can be obtained with less data compared to
namics do not contain the control inputs nu and nl , where nu neural network‐based algorithms, which in turn requires fewer
and nl are the SMs operated in both arms of the MMC. For computational resources and less time for model training. In
simplicity, the control inputs will be referred to as the insertion this research the SINDy model was trained using Google
indices. Since io and icir are the systems states to be controlled, Colab with a TPU runtime, leveraging a cloud TPU that offers
the computational burden and the control complexity can be 180 teraflops of computational power, alongside an Intel Xeon
reduced if the insertion indexes are included in the dynamic CPU @ 2.30 GHz and 13 GB of RAM. Using these compu-
equations of io and icir . Motivated by this, the SINDy algo- tational resources, the training time for the SINDy model was
rithm is used to obtain a dynamic model for the output and under 60 s.
circulating currents with the insertion indexes included in the
dynamic equations. Therefore, there is no need to evaluate uu
and ul based on different insertion indexes to calculate the 4 | FCS‐MPC DESIGN
predictions of io and icir required for the FCS‐MPC imple-
mentation. The following SINDy‐based dynamic model of In this section, the proposed FCS‐MPC scheme for MMC is
MMC for io and icir is obtained: presented. The overall block diagram of the presented scheme
is shown in Figure 2. The main control targets are output
8 current, circulating current, and SM capacitor voltage
> dicir � balancing. The control of output current and circulating cur-
>
< ¼ Θ xT ; u T ξ 1
dt rent is implemented by selecting the optimal insertion indexes
ð13Þ
> di
: o ¼ Θ xT ; uT �ξ
> that minimise a specific cost function. Similar to the indirect
dt 2 FCS‐MPC, the capacitor voltage balancing is not included in
the implementation of the controller to further minimise the
computational burden.
where x ¼ ½io icir uu ul �T is the system state vector and This section presents a FCS‐MPC with reduced compu-
u ¼ ½ nu nl �T is vector containing the control input terms. tational burden. It involves obtaining the MMC model using
Based on different candidate functions and hyperparameters, SINDy that substantially reduces the computational
the obtained models for io and icir only contain the first‐order complexity required to evaluate the prediction step involved in
terms involving inputs and states. It is possible to obtain FCS‐MPC. It will also be shown during model verification that
multiple models based on different choices of candidate SINDy‐based model provides a very accurate representation of
functions. However, the SINDy model used in this research the system, thereby providing precise predictions of the output
offers a computationally efficient solution for adding input and circulating currents.
terms to the io and icir dynamics without affecting perfor- When the conventional indirect FCS‐MPC is designed for
mance. Furthermore, because SINDy uses limited data for MMCs, in the mathematical modelling, it is assumed that the
training, it has a shorter training time compared to other capacitors are balanced. Furthermore, this approach does not
machine learning based techniques. account for factors such as switching losses. These inaccuracies
can lead to prediction errors, potentially degrading the control
performance. On the other hand, the SINDy‐based FCS‐MPC
3.2 | Model development uses a data‐driven model that is able to capture the capacitor
voltage ripples caused by capacitor voltage imbalance in the
To obtain a dynamic model using SINDy, two steps are per- system model, therefore, it provides a better prediction model.
formed with the PySINDy library [42], as stated below. This improves the performance of the FCS‐MPC, which relies
on the accuracy of the prediction model while reducing
1. Acquiring data: The system is run on open‐loop for only computational complexity. A trade‐off of using a SINDy‐based
10s with a sampling time of 0:0001s to collect the training
data. The system was sufficiently excited to ensure that the
obtained model captures the system dynamics. In case when
obtaining open loop data is not feasible, the system can be
operated in closed loop by implementing any controller
such as a PI or a conventional FCS‐MPC, that provides
sufficient excitation to the system.
2. Model generation: In the second step, X and U matrices are
constructed by using the data from the first step. Next, the
finite differencė method is used to evaluate the time de-
rivative matrix X . The model is then trained under various
ΘðXm ; Um Þ and λ. Finally, the obtained model is tested
using the test data to ensure accuracy. FIGURE 2 Proposed FCS‐MPC for MMCs.
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6
- RAJA ET AL.

approach is the requirement to acquire data and train the ic ðl þ 1Þ ¼ ic ðlÞ

model before it can be used to implement the FCS‐MPC. In
comparison, for the conventional approach, the mathematical þ Ts ðAio ðlÞ þ Bicir ðlÞ þ Cuu ðlÞ þ Dul ðlÞ þ Enu ðlÞ þ Fnl ðlÞÞ
model represented by Equations (5)‐(6) can be obtained using ð18Þ
KCL and KVL methods and employed directly without going
through the model acquisition step. io ðl þ 1Þ ¼ io ðlÞ
In order to achieve the capacitor voltage balancing, the þ Ts ðGio ðlÞ þ Hicir ðlÞ þ Iuu ðlÞ þ Jul ðlÞ þ Knu ðlÞ þ Lnl ðlÞÞ
technique applied by Gong et al. [43] has been implemented.
ð19Þ
Based on the control output generated by FCS‐MPC in terms of
the insertion indexes, the balancing scheme chooses the SMs in
the upper and lower arms to be inserted or bypassed depending where A; B; C; D; E; F; G; H; I; J; K, and L are the constant
on the direction of the arm current ij . coefficients with system states and inputs. From Equa-
tions (14)‐(17), it can be observed that while using the math-
ematical model to implement a FCS‐MPC on MMCs, the
4.1 | Prediction models of FCS‐MPC for computational burden due to the evaluation of the output
MMCs current and circulating current prediction is more than the
equivalent SINDy‐based prediction model given by Equa-
The forward Euler method is employed to acquire the one‐step tions (18)‐(19). Furthermore, the number of evaluations grows
prediction of output current io ðl þ 1Þ and circulating current linearly as the number of SMs increases in the case of the
icir ðl þ 1Þ. The obtained discrete‐time model based on the mathematical model, whereas they remain constant irrespective
mathematical model represented by Equations (5)‐(6) is shown of the number of SMs. It is due to the fact that in order to
below. evaluate the arm voltages uj for a specific insertion index nj ,
the sum of all the capacitor voltages must be computed, and
1 when N is large which is typically a case with MMCs, the
io ðl þ 1Þ ¼ io ðlÞ þ ðu ðlÞ − uu ðlÞ − 2Rio ðlÞÞTs
2L þ La l
computational burden of implementing FCS‐MPC increases
when the mathematical model is used. On the contrary, for the
ð14Þ SINDy‐based prediction model of output and circulating
currents, the insertion index nj is a part of the discrete‐time
1
icir ðl þ 1Þ ¼ icir ðlÞ þ ðU − ul ðlÞ − uu ðlÞÞTs ð15Þ model and the sum of capacitor voltages is not evaluated.
2La dc
Remark 4.1 As a comparison, each evaluation of io ðk þ 1Þ and
where Ts is the sample time. In the literature [24–31], the upper
icir ðk þ 1Þ require 2N þ 17 basic operations to be performed
and lower arm voltages are calculated under the assumption
when using the mathematical model. However, only 26 basic
that all the capacitor voltages are well‐balanced. Therefore, the
operations are performed when using the SINDy approach,
arm voltages uj for a insertion index nj is given below:
where the basic operations include addition and multiplication
and each basic operation requires 1 machine cycle to process.
N
nj X
uj ðlÞ ¼ uci ðlÞ ð16Þ
N i¼1
4.2 | Cost function
Using Equation (7), the predicted value of the sum of SM
capacitor voltage is obtained as shown below: The FCS‐MPC optimisation problem of MMCs involves
minimising a cost function under all the control options and
N N constraints. The optimisation problem is outlined as follows:
X X nj ij Ts
uci ðl þ 1Þ ¼ uci ðlÞ þ ð17Þ
i¼1 i¼1
C SM min Jðxðl þ 1Þ; UðlÞÞ
s:t: xlþ1 ¼ f ðxðlÞ; uðlÞÞ
In a similar manner, the discrete‐time model of the output � � ð20Þ
current and circulating current based on the SINDy‐based UðlÞ ¼ u1 ðlÞ; u2 ðlÞ; …; u � ðlÞ
N
model as described by Equation (11) is computed as follows: 3 þ1
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RAJA ET AL.
- 7

T A B L E 1 Basic operations performed

MPC algorithms Machine cycles per prediction step
by prediction step per control set for different
FCS‐MPC methods. Direct FCS‐MPC [21] 10N þ 17

Indirect FCS‐MPC [24–29, 45] and MMPC [30, 31, 46] 2N þ 17

Proposed 26

TABLE 2 Total basic operations performed during evaluation of prediction step based on N.

Number of SMs (N) 4 20 50 100 200

No ðCo Þ Indirect FCS‐MPC [24] 297 (17) 6857 (401) 42,617 (2501) 170,217 (10,001) 680,417 (40,001)

Preselection method [29] 93 (5) 176 (8) 321 (13) 472 (16) 1199 (47)

Simplified indirect FCS‐MPC [45] 65 (3) 131 (5) 242 (8) 444 (14) 814 (24)

Proposed 48 (3) 108 (9) 188 (17) 348 (33) 688 (67)

TABLE 3 Simulation parameters of the MMC system. The following cost function is considered for this research:
Parameters Values
Arm inductance La 15mH J ¼ λc ji∗cir − icir ðl þ 1Þj þ λo ji∗o − io ðl þ 1Þj
ð21Þ
Load inductance L 50mH þ λu nu ðlÞ þ λl nl ðlÞ
Load resistance R 20Ω
where λc , λo , λu , λl are the weighting factors of the circulating
SM capacitance CSM 4700μF
current, output current, insertion index of the upper arm, and
DC link voltage Udc 1500V insertion index of the lower arm, respectively, i∗cir is the
SMs per arm N 20 reference of the circulating current, and i∗o is the output current
reference. A sinusoidal reference is provided for the output
Sample time Ts 100μs
current, whereas the circulating current reference applied by
Output frequency f 50Hz Vatani et al. [44] is employed to ensure good circulating current
suppression. The cost function defined in Equation (21) is
evaluated under all the control options defined in U and the
insertion index that yields the smallest value of the cost
function is selected by the control scheme.

Remark 4.2 The weighting terms associated with the insertion

indexes help to reduce the output current THD and improve
the system's efficiency. However, the weights should be care-
fully selected, as increasing them could negatively affect steady
state performance.

4.3 | Strategy for selecting the insertion

index

The proposed FCS‐MPC control strategy considers N3 þ 1

switching combinations for the insertion indexes nj to compute
the cost function. The symmetric structure of MMCs is used to
develop an efficient selection strategy that could reduce the
switching combinations considered by the MPC. Instead of
F I G U R E 3 Model validation of the SINDy model. From top down, using the optimal ðN þ 1Þ2 combinations used in the case of
waveforms are: (i) circulating current and (ii) output current. an indirect FCS‐MPC, the switching combinations that
generate nu þ nl ¼ fN − 1; N; N þ 1g are picked. To further
where J represents the cost function, f is the system model in reduce the combinations, only N6 nearest switching combina-
discrete‐time domain, x is the system state vector, and U tions to the previously selected insertion index are used.
contains all the possible input switching combinations. Therefore, the switching combinations are reduced to N3 þ 1.
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8
- RAJA ET AL.

The choice of the best nearest levels is based on the output proposed control options are suitable for MMCs with any
performance in which these nearest levels are reduced until a number of SMs. In case when increased controllability in terms
noticeable change in the output was observed. Since the of the circulating current is required, the domain of nu þ nl can
insertion indexes depend upon N, the proposed technique will be increased which will increase the switching combinations.
work with any number of SMs. This approach was tested for For this research, since there are 20 SMs in each arm ðN ¼ 20Þ,
MMCs having different submodules, and it was found that the the total number of nearest levels chosen is 4.

F I G U R E 4 Steady state performance based on simulation results. From top down, waveforms are: (i) output current, (ii) circulating current, (iii) arm
currents, (iv) arm voltages, and (v) SM capacitor voltages.
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RAJA ET AL.
- 9

TABLE 4 Simulation performance comparison.

N N
Performance Conventional Conventional Conventional 3 þ1 Proposed 3 þ1
criteria ðN þ 1Þ2 options 3N þ 1 options options options
io THD (%) 1.2 1.2 1.2 1.2

io reference tracking Excellent Excellent Excellent Excellent

icir RMS (A) 1.7 1.8 1.8 1.5

Switching frequency (Hz) 102 102 102 100

F I G U R E 5 Dynamics performance based on simulation results. From top down, waveforms are: (i) output current, (ii) circulating current, (iii) arm currents,
(iv) arm voltages, and (v) SM capacitor voltages.
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10
- RAJA ET AL.

F I G U R E 6 Steady state performance based on simulation results. From top down, waveforms are: (i) output current, (ii) circulating current, (iii) arm
currents, (iv) arm voltages, (v) Insertion indexes, and (vi) SM capacitor voltages.

TABLE 5 Experimental parameters of the MMC system.

method for preselecting the control options or the order of the
Parameters Values operations performed.
Arm inductance La 5mH The overall complexity of the proposed scheme is also
computed and compared when N is increased while taking into
Load inductance L 10mH
account the number of control options. Table 2 shows the
Load resistance R 10:6Ω comparison of the total number of basic operations No per-
SM capacitance CSM 5000μF
formed during the evaluation of the prediction step for the entire
control set. The terms inside braces represent the number of
DC link voltage Udc 210V control options denoted by C o . It can be seen that the proposed
SMs per arm N 8 method has an advantage in terms of the computational burden
Sample time Ts 500μs
as it has the least No for any N. The control options chosen in our
research are N3 þ 1, and they depend on the number of SMs
Output frequency f 50Hz
which makes it convenient to implement it on MMCs with a
greater number of SMs while still keeping the number of control
4.4 | Comparison options low. It can be seen that with the proposed method, we
get more flexibility in terms of the number of control options
To show the superiority of the proposed SINDy‐based FCS‐ while still keeping the overall computational complexity lower as
MPC, a detailed comparison is drawn against the existing tech- compared to other methods. Furthermore, the performance of
niques that tried to tackle the issue of reducing the FCS‐MPC the proposed SINDy‐based FCS‐MPC is also compared with the
computational complexity applied in MMCs. Table 1 shows conventional indirect FCS‐MPC having ðN þ 1Þ2 and 3N þ 1
the number of arithmetic operations performed during the control options to verify the effectiveness of the proposed
prediction step of the FCS‐MPC for various techniques and the approach.
proposed scheme. It can be seen that the complexity remains the The preselection method employed in ref. [45] uses circu-
same for each evaluation of the prediction step of the proposed lating current icir , circulating current reference i∗cir , and sum of
method, whereas in the case of the conventional approach, it insertion indices sj ¼ nu þ nl to determine the control options
grows linearly when the number of SMs increases. It is also worth for evaluating the FCS‐MPC. The algorithm reduces the number
noting that the term 2N þ 17 represents the minimum number of control options to reduce the computational complexity,
of machine cycles per control set taken among [24–29, 45] and however in order to choose the appropriate set of control op-
[30, 31, 46] as the total operations could vary based on the tions, additional comparison and arithmetic operations are
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RAJA ET AL.
- 11

5.1 | Validation of SINDy model for MMC

The SINDy model of the MMC described in (11) is validated and

compared with the actual values. Figure 3 shows the results of io
and icir produced by the SINDy model, as well as the model
defined in Equations (5)‐(6), and compares these results with the
measured output and circulating currents. The accuracy of the
SINDy model is also verified under sudden load changes in which
the amplitude and phase angle of the output current are changed
simultaneously. It can be deduced from the observed states that
both models provide an accurate estimate of io , whereas the
SINDy model generated a better estimate of the circulating
FIGURE 7 Proto‐type of the single phase MMC. current compared to the mathematical model. A better estimate
means an increase in the accuracy of the predictions generated by
the FCS‐MPC algorithm which will result in better circulating
performed. The computational complexity of this method is current suppression.
computed for different number of SMs while considering the
minimum possible comparison and arithmetic operations per-
formed for selecting the control sets. Furthermore, compared to 5.2 | Steady state response
the indirect method, it generates a larger circulating current,
whereas the proposed method results in superior circulating The performance of the proposed FCS‐MPC scheme under
current suppression compared to the indirect FCS‐MPC steady state conditions is investigated. Figure 4 illustrates the
approach. performance of the SINDy‐based FCS‐MPC at steady state
The controller imitation for the FCS‐MPC proposed by the with N3 þ 1 control options and its comparison with conven-
authors in refs. [32, 33] using ANNs reduces the computational tional indirect FCS‐MPC for ðN þ 1Þ2 , 3N þ 1, and N3 þ 1
complexity of a conventional FCS‐MPC. However, the emula- control options. In the conventional approach, the mathe-
tion shows inferior performance compared to the conventional matical model represented by Equations (5)‐(7) is used to
and the proposed FCS‐MPC. Furthermore, in ref. [34], the au- implement the indirect FCS‐MPC, while the proposed FCS‐
thors proposed an MFAC‐based FCS‐MPC framework that re- MPC employs the model presented in Equation (11). The
lies only on I/O measurement data without requiring an explicit output current reference is set to 30 A to evaluate and compare
model. Although this approach offers the advantage of model the steady state performance of the proposed method with the
independence, it has several drawbacks, such as a complex tuning conventional approach.
process, higher computational costs, and slower responses to It can be observed from Figures 4a‐c that the proposed
rapid transients. preselection method for the insertion indexes generates a
similar steady state response as compared to the case when
Remark 4.3 The implementation of FCS‐MPC for MMCs ðN þ 1Þ2 and 3N þ 1 control options are used. In Figure 4d,
based on the SINDy model can be combined with existing
the performance of the FCS‐MPC based on SINDy with N3 þ 1
methods that use different switching combinations to
control options under steady state conditions is evaluated. The
reduce computational complexity. It is capable of producing
proposed method yields better circulating current suppression
the same results with reduced computational burden due to the
compared to the conventional approach and provides good
simplicity of the prediction step.
output current tracking while reducing the computational
burden. This is possible because, in the case of the conven-
tional FCS‐MPC, the arm voltages are evaluated under the
5 | SIMULATION RESULTS
assumption that the SM capacitor voltages are perfectly
balanced which can introduce inaccuracies in the prediction
The simulation results of the proposed FCS‐MPC are generated
model.
in Simulink/MATLAB, whereas the model training for gener-
ating the SINDy data‐driven model is carried out in Python using Table 4 shows the performance comparison of the pro-
the PySINDy library. The simulation parameters used for this posed method with conventional approaches. It can be
research are listed in Table 3. The design of initial values of La and observed that the SINDy‐based FCS‐MPC with N3 þ 1 control
C SM is based on the method detailed in ref. [47]. These param- options provides similar output current performance and
eters are further tuned in an iterative process to improve the better circulating current performance, while yielding lower
results. switching frequency, resulting in improved system efficiency.
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
12
- RAJA ET AL.

F I G U R E 8 Steady state performance based on experimental results. From top down, waveforms are: (i) output current, (ii) circulating current, (iii) arm
currents, (iv) arm voltages, and (v) SM capacitor voltages.

5.3 | Dynamic response Figure 5 shows that the proposed method provides a fast dy-
namic response and allows io to respond quickly and accurately
The dynamic response of the MMC is also observed with the even after changing the i∗o aggressively. Similar to the steady state
proposed scheme by changing the operating conditions. In this performance, the circulating current suppression, in this case, is
case, the amplitude and phase angle of the output current also better for the proposed FCS‐MPC compared to the con-
reference are varied simultaneously. The amplitude is changed ventional methods with ðN þ 1Þ2 , 3N þ 1 and N3 þ 1 control
from 30 to 20 A, whereas a 180o phase angle is added at t ¼ 2 s. options. It can be seen that the circulating current is almost pure
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RAJA ET AL.
- 13

F I G U R E 9 Dynamics performance based on experimental results. From top down, waveforms are: (i) output current, (ii) circulating current, (iii) arm
currents, (iv) arm voltages, and (v) SM capacitor voltages.

TABLE 6 Experimental performance comparison.

Performance criteria Conventional ðN þ 1Þ2 options Conventional 3N þ 1 options Conventional N

3 þ 1 options Proposed N
3 þ 1 options
io THD (%) 1.5 1.8 1.7 1.1

io reference tracking Excellent Excellent Excellent Excellent

icir RMS (A) 0.82 0.81 0.81 0.78

Switching frequency (Hz) 262 258 248 248

 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
14
- RAJA ET AL.

F I G U R E 1 0 Steady state performance based on experimental results. From top down, waveforms are: (i) output current, (ii) circulating current, (iii) arm
currents, (iv) arm voltages, (v) Insertion indexes, and (vi) SM capacitor voltages.

DC for SINDy‐based FCS‐MPC after the change in i∗o is Figure 8 shows that the three cases of the conventional approach
introduced, whereas the results of conventional indirect FCS‐ produce almost identical results, whereas the proposed SINDY‐
MPC show oscillations in the circulating current. The capac- FCS‐MPC guarantees better circulating current suppression
itor voltage balancing was seen to diverge because of the large while providing a computationally efficient solution to the
change in amplitude and phase angle of i∗o but it returns back to problem. Furthermore, the output current tracking performance
the steady state in a short span of time. If only the amplitude was and the capacitor voltage balancing for the implemented strategy
changed, this deviation of capacitor voltage balancing would be was observed to be good.
minimal. Furthermore, the SINDy‐based FCS‐MPC showed In order to evaluate the dynamic performance of the SINDy‐
better capacitor voltage balancing. FCS‐MPC with the proposed input preselection, the output
current reference amplitude is changed from 5 to 7 A with a 180o
phase shift. The resulting response of the conventional ap-
5.4 | Effect of changing input weights proaches and the proposed scheme is shown in Figure 9. Similar
to the steady state case, the dynamic performance for the three
Figure 6 shows the effect of changing the weights associated with cases of the conventional FCS‐MPC approach generates very
the insertion indexes, that is, λu and λl . As the input weighs are similar results. However, the proposed FCS‐MPC demonstrated
increased from λu;l ¼ 0:001 to λu;l ¼ 0:002, the FCS‐MPC tends better harmonic performance for the output current. Further-
to select a smaller number for the insertion indexes. This causes more, the circulating current response was more robust to
fewer SMs that are operated, hence increasing the overall system changes in the weighting factors and showed an overall smaller
efficiency. Furthermore, the output THD is reduced from 1.2% higher‐order harmonics in the circulating current.
to 0.8%. The performance comparison of the proposed method
with the conventional approaches is shown in Table 6. A
similar output current tracking performance is observed for all
6 | EXPERIMENTAL RESULTS cases; however, the SINDy‐based FCS‐MPC with N3 þ 1 results
in better output current harmonic performance and circulating
To further validate the proposed FCS‐MPC, experiments are current performance and yields lower switching frequencies,
conducted on a single phase MMC laboratory prototype having 8 thereby improving system efficiency. It is worth noting that in
SMs per arm. The experimental system parameters used in this the conventional FCS‐MPC case with ðN þ 1Þ2 control op-
research are given in Table 5 and the MMC prototype is shown in tions, it is more challenging to select appropriate weighting
Figure 7. The proposed scheme is verified during steady state factors to ensure a circulating current close to pure DC value
conditions, with the reference of output current set to 5 A. and to minimise higher order harmonics component. This is
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RAJA ET AL.
- 15

mainly because we are not limiting the selection of the inser- CO N FL I C T O F I NT E R E S T S TA T E ME N T

tion indexes based on previous values. The author and the co‐author do not have a conflict of interest
Finally, the effect of varying input weights is also studied to disclose.
using the experimental setup. It can be observed from Figure 10
that increasing input weights increases the probability of DA TA AVA I L AB ILI T Y S TA T E M EN T
selecting a smaller insertion index and hence improves the Data sharing not applicable to this article as no datasets were
overall system efficiency. generated or analysed during the current study URL https://
www.sciencedirect.com/science/article/pii/
S0142061521004208.
7 | CONCLUSION AND FUTURE WORK
O RCI D
In this research, a novel data‐driven FCS‐MPC is proposed for Muneeb Masood Raja https://orcid.org/0000-0003-0957-
MMCs with a reduced computational burden. A linear model of 8696
output and circulating currents is obtained with the insertion
index terms included in the dynamics to reduce the computa- RE F E R E N C E S
tional complexity of the prediction step. In the proposed tech- 1. Bergna, G., et al.: An energy‐based controller for hvdc modular multi-
nique, N3 þ 1 control options are used while evaluating the FCS‐ level converter in decoupled double synchronous reference frame for
voltage oscillation reduction. IEEE Trans. Ind. Electron. 60(6), 2360–
MPC instead of the conventional ðN þ 1Þ2 control options 2371 (2013). https://doi.org/10.1109/TIE.2012.2225397
which also play a role in reducing the computational complexity. 2. Liu, Z., et al.: Characteristics analysis and improved arm control of
A comparison is also carried out with the existing techniques modular multilevel converter under asymmetric operation conditions.
to prove the superiority of the proposed scheme in terms of Int. J. Electr. Power Energy Syst. 105, 272–282 (2019). https://doi.org/
10.1016/j.ijepes.2018.08.037
reducing the computational burden of implementing the FCS‐
3. Debnath, S., Qin, J., Saeedifard, M.: Control and stability analysis of
MPC on MMCs. Furthermore, the presented technique gener- modular multilevel converter under low‐frequency operation. IEEE
ates better circulating current suppression compared to the in- Trans. Ind. Electron. 62(9), 5329–5339 (2015). https://doi.org/10.1109/
direct FCS‐MPC conventional approaches. Finally, the proposed TIE.2015.2414908
scheme can be combined with any of the pre‐existing techniques 4. Li, N., et al.: Soh balancing control method for the mmc battery energy
storage system. IEEE Trans. Ind. Electron. 65(8), 6581–6591 (2018).
to minimise the computational burden of FCS‐MPC as it in-
https://doi.org/10.1109/TIE.2017.2733462
volves evaluating a computationally efficient prediction step 5. Fan, B., et al.: Hierarchical system design and control of an mmc‐based
during the implementation of the MPC. power‐electronic transformer. IEEE Trans. Ind. Inf. 13(1), 238–247
A trial‐and‐error method was used to find the best weighting (2017). https://doi.org/10.1109/TII.2016.2522190
factors for the FCS‐MPC. In the future, gradient‐based opti- 6. He, R., et al.: A capacitor overvoltage elimination strategy for reduced‐
voltage‐sensor‐based mmc. Int. J. Electr. Power Energy Syst. 132,
misation or evolutionary algorithms may be used to identify the
107181 (2021). https://doi.org/10.1016/j.ijepes.2021.107181
ideal weights. Furthermore, a continuous set MPC will also be 7. Hassanpoor, A., et al.: Evaluation of different carrier‐based pwm
implemented in the future and compared with the proposed methods for modular multilevel converters for hvdc application. In:
method. Finally, the scope of this research can be further IECON 2012 ‐ 38th Annual Conference on, pp. 388–393. IEEE In-
expanded by implementing fault diagnosis and fault‐tolerant dustrial Electronics Society (2012). https://doi.org/10.1109/IECON.
2012.6388789
control techniques for different faults, such as switch open and
8. Tu, Q., Xu, Z.: Impact of sampling frequency on harmonic distortion for
short circuit faults, single‐phase‐to‐ground faults, and DC‐bus modular multilevel converter. IEEE Trans. Power Deliv. 26(1), 298–306
short circuit faults. (2011). https://doi.org/10.1109/TPWRD.2010.2078837
9. Konstantinou, G., Ciobotaru, M., Agelidis, V.: Selective harmonic elim-
A UT HO R CONT RI BU T I O NS ination pulse‐width modulation of modular multilevel converters. IET
Power Electron. 6(1), 96–107 (2013). https://doi.org/10.1049/iet‐pel.
Muneeb Masood Raja: Conceptualisation; Formal analysis;
2012.0228
Investigation; Methodology; Software; Validation; Visualisation; 10. Deng, F., Chen, Z.: Voltage‐balancing method for modular multilevel
Writing ‐ original draft; Writing ‐ review and editing. Haoran converters under phase‐shifted carrier‐based pulsewidth modulation.
Wang: Investigation; Writing ‐ review and editing. Muhammad IEEE Trans. Ind. Electron. 62(7), 4158–4169 (2015). https://doi.org/10.
Haseeb Arshad: Writing ‐ review and editing. Gregory J. 1109/TIE.2014.2388195
11. Bahrani, B., Debnath, S., Saeedifard, M.: Circulating current suppression
Kish9: Project administration; Resources; Supervision; Valida-
of the modular multilevel converter in a double‐frequency rotating
tion; Writing ‐ review and editing. Qing Zhao: Conceptualisa- reference frame. IEEE Trans. Power Electron. 31(1), 783–792 (2016).
tion; Methodology; Project administration; Resources; https://doi.org/10.1109/TPEL.2015.2405062
Supervision; Validation; Visualisation; Writing ‐ review and 12. Li, X., et al.: Performance analysis and optimization of circulating current
editing. control for modular multilevel converter. IEEE Trans. Ind. Electron.
63(2), 716–727 (2016). https://doi.org/10.1109/TIE.2015.2480748
13. Yang, Q., Saeedifard, M., Perez, M.A.: Sliding mode control of the
A C KN OW LE DG EMEN T S modular multilevel converter. IEEE Trans. Ind. Electron. 66(2), 887–897
The authors do not have any funding information to disclose. (2019). https://doi.org/10.1109/TIE.2018.2818657
 17518679, 0, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/elp2.12523 by CAPES, Wiley Online Library on [17/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
16
- RAJA ET AL.

14. Rodriguez, J., et al.: State of the art of finite control set model predictive Trans. Ind. Electron. 66(6), 4368–4378 (2019). https://doi.org/10.1109/
control in power electronics. IEEE Trans. Ind. Inf. 9(2), 1003–1016 TIE.2018.2863181
(2013). https://doi.org/10.1109/TII.2012.2221469 32. Novak, M., Dragicevic, T.: Supervised imitation learning of finite‐set
15. Vazquez, S., et al.: Model predictive control for power converters and model predictive control systems for power electronics. IEEE Trans.
drives: advances and trends. IEEE Trans. Ind. Electron. 64(2), 935–947 Ind. Electron. 68(2), 1717–1723 (2021). https://doi.org/10.1109/TIE.
(2017). https://doi.org/10.1109/TIE.2016.2625238 2020.2969116
16. Böcker, J., et al.: Experimental comparison of model predictive control 33. Wang, S., et al.: Machine learning emulation of model predictive control
and cascaded control of the modular multilevel converter. IEEE Trans. for modular multilevel converters. IEEE Trans. Ind. Electron. 68(11),
Power Electron. 30(1), 422–430 (2015). https://doi.org/10.1109/TPEL. 11628–11634 (2021). https://doi.org/10.1109/TIE.2020.3038064
2014.2309438 34. Wu, W., et al.: Data‐driven finite control‐set model predictive control for
17. Ilves, K., et al.: Predictive sorting algorithm for modular multilevel modular multilevel converter. IEEE Journal of Emerging and Selected
converters minimizing the spread in the submodule capacitor voltages. Topics in Power Electronics 11(1), 523–531 (2023). https://doi.org/10.
IEEE Trans. Power Electron. 30(1), 440–449 (2015). https://doi.org/10. 1109/JESTPE.2022.3207454
1109/TPEL.2014.2308018 35. Akaike, H.: Fitting autoregressive models for prediction. Ann. Inst. Stat.
18. Liu, X., et al.: Event‐triggered neural‐predictor‐based fcs‐mpc for mmc. Math. 21(12), 243–247 (1969). https://doi.org/10.1007/BF02532251
IEEE Trans. Ind. Electron. 69(6), 6433–6440 (2022). https://doi.org/10. 36. Billings, S.A.: Nonlinear system identification: narmax methods in the
1109/TIE.2021.3094447 time, frequency, and spatio‐temporal domains, (2013)
19. Dragičevi�c, T.: Model predictive control of power converters for robust 37. Peng, H., et al.: Nonlinear system modeling and predictive control using
and fast operation of ac microgrids. IEEE Trans. Power Electron. 33(7), the rbf nets‐based quasi‐linear arx model. Control Eng. Pract. 17(1), 59–
6304–6317 (2018). https://doi.org/10.1109/TPEL.2017.2744986 66 (2009). https://doi.org/10.1016/j.conengprac.2008.05.005
20. Qin, J., Saeedifard, M.: Predictive control of a modular multilevel con- 38. Zhang, T., et al.: Learning deep control policies for autonomous aerial
verter for a back‐to‐back hvdc system. IEEE Trans. Power Deliv. 27(3), vehicles with mpc‐guided policy search. https://doi.org/10.48550/
1538–1547 (2012). https://doi.org/10.1109/TPWRD.2012.2191577 ARXIV.1509.06791
21. Böcker, J., et al.: Experimental comparison of model predictive control 39. Chen, R.T.Q., et al.: Neural ordinary differential equations. https://doi.
and cascaded control of the modular multilevel converter. IEEE Trans. org/10.48550/ARXIV.1806.07366
Power Electron. 30(1), 422–430 (2015). https://doi.org/10.1109/TPEL. 40. Brunton, S.L., Proctor, J.L., Kutz, J.N.: Discovering governing equations
2014.2309438 from data by sparse identification of nonlinear dynamical systems. Proc.
22. Peralta, J., et al.: Detailed and averaged models for a 401‐level mmc–hvdc Natl. Acad. Sci. USA 113(15), 3932–3937 (2016). arXiv. https://doi.org/
system. IEEE Trans. Power Deliv. 27(3), 1501–1508 (2012). https://doi. 10.1073/pnas.1517384113
org/10.1109/TPWRD.2012.2188911 41. Kaiser, E., Kutz, J.N., Brunton, S.L.: Sparse identification of nonlinear
23. Moon, J.‐W., et al.: Model predictive control with a reduced number of dynamics for model predictive control in the low‐data limit. Proc. R. Soc.
considered states in a modular multilevel converter for hvdc system. A 474(2219), 20180335 (2018). https://doi.org/10.1098/rspa.2018.0335
IEEE Trans. Power Deliv. 30(2), 608–617 (2015). https://doi.org/10. 42. Kaptanoglu, A., et al.: PySINDy: a comprehensive python package for
1109/TPWRD.2014.2303172 robust sparse system identification. J. Open Source Softw. 7(69), 3994
24. Vatani, M., et al.: Indirect finite control set model predictive control of (2022). https://doi.org/10.21105/joss.03994
modular multilevel converters. IEEE Trans. Smart Grid 6(3), 1520–1529 43. Gong, Z., et al.: Design and experimental evaluation of fast model pre-
(2015). https://doi.org/10.1109/TSG.2014.2377112 dictive control for modular multilevel converters. IEEE Trans. Ind.
25. Gong, Z., et al.: Design and experimental evaluation of fast model pre- Electron. 63(6), 3845–3856 (2016). https://doi.org/10.1109/TIE.2015.
dictive control for modular multilevel converters. IEEE Trans. Ind. 2497254
Electron. 63(6), 3845–3856 (2016). https://doi.org/10.1109/TIE.2015. 44. Vatani, M., et al.: Indirect finite control set model predictive control of
2497254 modular multilevel converters. IEEE Trans. Smart Grid 6(3), 1520–1529
26. Moon, J.‐W., et al.: Model predictive control with a reduced number of (2015). https://doi.org/10.1109/TSG.2014.2377112
considered states in a modular multilevel converter for hvdc system. 45. Nguyen, M.H., Kwak, S.: Simplified indirect model predictive control
IEEE Trans. Power Deliv. 30(2), 608–617 (2015). https://doi.org/10. method for a modular multilevel converter. IEEE Access 6, 62405–
1109/TPWRD.2014.2303172 62418 (2018). https://doi.org/10.1109/ACCESS.2018.2876505
27. Liu, P., et al.: Grouping‐sorting‐optimized model predictive control for 46. Jin, Y., et al.: A novel sliding‐discrete‐control‐set modulated model pre-
modular multilevel converter with reduced computational load. IEEE dictive control for modular multilevel converter. IEEE Access 9, 10316–
Trans. Power Electron. 31(3), 1896–1907 (2016). https://doi.org/10. 10327 (2021). https://doi.org/10.1109/ACCESS.2021.3050340
1109/TPEL.2015.2432767 47. Zygmanowski, M., Grzesik, B., Nalepa, R.: Capacitance and inductance
28. Rashwan, A., et al.: Predictive controller based on switching state selection of the modular multilevel converter. In: 2013 15th European
grouping for a modular multilevel converter with reduced computational Conference on Power Electronics and Applications, pp. 1–10. EPE
time. IEEE Trans. Power Deliv. 32(5), 2189–2198 (2017). https://doi. (2013). https://doi.org/10.1109/EPE.2013.6634446
org/10.1109/TPWRD.2016.2639529
29. Gutierrez, B., Kwak, S.‐S.: Modular multilevel converters (mmcs)
controlled by model predictive control with reduced calculation burden.
IEEE Trans. Power Electron. 33(11), 9176–9187 (2018). https://doi.
org/10.1109/tpel.2018.2789455 How to cite this article: Raja, M.M., et al.:
30. Mahmoudi, H., Aleenejad, M., Ahmadi, R.: Modulated model predictive Computationally efficient data‐driven model predictive
control of modular multilevel converters in vsc‐hvdc systems. IEEE control for modular multilevel converters. IET Electr.
Trans. Power Deliv. 33(5), 2115–2124 (2018). https://doi.org/10.1109/ Power Appl. 1–16 (2024). https://doi.org/10.1049/
TPWRD.2017.2727478
elp2.12523
31. Zhou, D., Yang, S., Tang, Y.: Model‐predictive current control of modular
multilevel converters with phase‐shifted pulsewidth modulation. IEEE