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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Model Predictive Control Using Artificial Neural

Network for Power Converters
Daming Wang, Z. John Shen, Fellow, IEEE, Xin Yin, Sai Tang, Xifei Liu, Chao Zhang, Jun Wang, Senior
Member, IEEE, Jose Rodriguez, Fellow, IEEE, Margarita Norambuena

Abstract—There has been an increasing interest in using MPC, with flexible system models, constraint conditions and
model predictive control (MPC) for power electronic cost functions, can be applied to systems with complex
applications. However, the exponential increase in constraints, nonlinearities, and multivariable objectives. These
computational complexity and demand of computing features have made MPC an attractive solution for power
resources hinders the practical adoption of this highly
promising control technique. In this paper, a new MPC
converter applications in recent years [1]-[6].
approach using an artificial neural network (termed ANN- However, the application of MPC in power converters is
MPC) is proposed to overcome these barriers. A power hindered by two major barriers: the exponential increase in
converter with a virtual MPC controller is first designed and computational complexity and the requirement of accurate
operated under a circuit simulation or power hardware-in- system models. First, the number of possible switching states
the-loop (PHIL) simulation environment. An artificial neural increases exponentially in multi-level and multi-phase power
network (ANN) is then trained offline with the input and converters, making the MPC algorithm calculation extremely
output data of the virtual MPC controller. Next, an actual complex and resource demanding process, and presenting a true
FPGA-based MPC controller is designed using the trained challenge even for the modern microcontrollers with ever-
ANN instead of relying on heavy-duty mathematical
computation to control the actual operation of the power
increasing computing power [1][2]. For this reason, the use of
converter in real time. The ANN-MPC approach can MPC is usually limited to low-order systems [5][6]. For more
significantly reduce the computing need and allow the use complex multi-phase and/or multi-level power converters, the
of more accurate high-order system models due to the use of MPC is limited to low switching frequencies due to the
simple mathematical expression of ANN. Furthermore, the computation limitation (e.g. 2.5-10kHz) [7]. This,
ANN-MPC approach can retain the robustness for system unfortunately, goes against the general trend of higher
parameter uncertainties by flexibly setting the input frequency and higher power density power converter designs in
elements. The basic concept, ANN structure, off-line the recent years [8]. A number of studies have been carried out
training method, and online operation of ANN-MPC are in an attempt to reduce the computational requirements of MPC
described in detail. The computing resource requirement of
the ANN-MPC and conventional MPC are analyzed and
control. Reference [9] reduced the number of iterations of the
compared. The ANN-MPC concept is validated by both MPC controller for a 3-phase 4-level flying capacitor (FC)
simulation and experimental results on two kW-class flying converter by decoupling the switching states in each phase.
capacitor multilevel converters. It is demonstrated that the Reference [10] improved the selection process of the candidate
FPGA-based ANN-MPC controller can significantly reduce regions to reduce the MPC computation time. Reference [11]
the FPGA resource requirement (e.g., 2.11 times fewer slice used a lookup table to reduce the number of candidate voltage
LUTs and 2.06 times fewer DSPs) while offering a control vectors to make the FCS-MPC more computationally efficient
performance same as the conventional MPC. for matrix converter motor drive. Reference [12] proposed a
Index Term—Model predictive control, Artificial neural MPC controller which only evaluates 3 of the 24 regions of the
network, Power converters, Computational burden.
hexagon in each sampling period to reduce the computational
requirement. While these prior works offer some improvement
I. INTRODUCTION
in these specific applications, they do not offer a general
Model predictive control (MPC) can evaluate the system solution to the fundamental computational challenge in
states and calculate the optimal solution in real time, resulting adopting MPC in complex higher-order and high-frequency
in excellent control performance and robustness. Furthermore, power converters.
Manuscript received October 19, 2020; revised January 07, 2021, and March
In recent years, artificial neural network (ANN) has emerged
16, 2021; accepted April 07, 2021. This work was supported in part by the Chinese as an excellent solution for input-output mapping problems [13],
National Natural Science Foundation under Grant No. 51977068. (Corresponding and was investigated to improve the control performance of
authors: Jun Wang; Daming Wang.)
Daming Wang, Xin Yin, Sai Tang, Xifei Liu, Chao Zhang, and Jun Wang are MPC in power converters [14][15]. Reference [14] used ANN
with the College of Electrical and Information Engineering, Hunan University, to estimate the capacitor voltages in a 3-phase modular
Changsha 410082, China (e-mail: [email protected]; [email protected];
[email protected]; [email protected]; [email protected]; multilevel converter (MMC). Reference [15] used ANN to
[email protected]). calculate weight factors for the MPC controller in a two-level
Z. J. Shen is with the Department of Electrical and Computer Engineering,
Illinois Institute of Technology, Chicago, IL 60616, USA (e-mail:
VSC. Reference [16], [17] proposed neural network imitation
[email protected]). method to control resonant power converters and three-phase
Jose Rodriguez is with the Faculty of Engineering, Universidad Andres Bello, inverter with simulation verification, respectively. References
Santiago 8370146, Chile (e-mail: [email protected]).
Margarita Norambuena is with Universidad Tecnica Federico Santa Maria, [18]-[20] used ANN to simplify the calculation of multistep
department of electrical engineering, Valparaiso, Chile, 2390123 (e-mail: prediction calculation for FCS-MPC [18][19], and CCS-MPC
[email protected]).

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Constraint prediction process, they can be evaluated together with the

conditions sampled value x(k) at the beginning of every regular sample
xr(k+1) period t=kT, based on a prediction function p, to predict the
Optimization
s(k)/d(k)
Load possible system transitions xpi(k+1) = p{x(k), si}, for i = 1, … ,
n. Afterwards, in the optimization process, the optimal control
Converter action s(k) is determined through a cost function g which is
(n switching states)
p
usually dependent on the desired reference value and the
x (k+1) Prediction x(k)
n (system model) predictions gi=g{xr(k+1), xpi(k+1)}. The evaluation of the cost
function with the n predictions will lead to n different costs. In
FCS-MPC/CCS-MPC the FCS-MPC method, one switching state can be the optimal
Fig. 1. Generic control diagram of MPC composed of the prediction and selection for two or more sample periods. This leads to a
optimization processes. variable switching frequency, which may cause resonances,
[20] in relatively low-order particular topologies i.e. two-level vibrations, or acoustic noise in some applications [2].
VSI, MMC, and DC–DC converters. The use of ANN in more To address the above issuses of FCS-MPC, the CCS-MPC
complex converters especially in multilevel and multiphase has been developed. A CCS-MPC that combines the merits of
systems needs to be futher investigted. the deadbeat MPC and FCS-MPC techniques has been applied
This paper introduces a general ANN-MPC approach to for torque control of permanent magnet synchronous motor
address the computational challenge in adopting MPC in highly (PMSM) [22] and of IM [23]. The CCS-MPC is based on
complex power converters. A power converter with a virtual voltage calculation via minimization of the cost function for
MPC controller is first designed and operated under a circuit reference tracking [24]. For CCS-MPC, the optimal control
simulation or power hardware-in-the-loop (PHIL) simulation action d(k) is composed of duty cycle of each power switch.
environment. An ANN is then trained offline with the input and Thus, every power switch has various states in one sample
output data of the virtual MPC controller. Next, an actual period and the switching frequency of each switch can be of
FPGA-based MPC controller is designed using the trained fixed values.
ANN instead of relying on heavy-duty mathematical
B. Major Challenges of MPC
computation to control the actual operation of the power
converter in real time. The trained neural network is capable of Despite its distinct advantages, MPC also has several serious
producing good approximation and offering comparable control disadvantages.
performance of the conventional MPC controller. What’s more, Generally, for a power converter, the number of possible
the structure of ANN-MPC in this paper is very intuitive and switching states can be calculated as 2(x-1)y, where x is the
simple, making it suitable for controlling multicell power number of independent power switches (pairs) of each leg in the
converters. The authors recently reported the ANN-MPC converter and y is the number of phases (or legs) in the
concept in a single-phase 5-level flying capacitor converter in converter [1]. As aforementioned, MPC implementation can be
[21]. In this paper, the general principle of ANN-MPC are more divided into the prediction process and optimization process.
extensively discussed as a common solution for highly complex Hence, for FCS-MPC, the controller needs to predict 2(x-1)y sets
multi-level and/or multi-phase or high-order power converters. of system states in the prediction process. Similarly, in the
Furthermore, the ANN-MPC can be used to approximate both optimization process, the controller needs to calculate 2(x-1)y
FCS-MPC and CCS-MPC. This paper theoretically analyzes different costs. Thus, the computational burden of FCS-MPC
the basic ANN-MPC concept, ANN structure, off-line training can be expressed as:
method, and online operation of ANN-MPC as a general N FC S -M P C  K ( n p  n c )  2 y ( x  1) ( n p  n c ) (1)
method for various power converters. The concept of ANN- Where NFCS-MPC is the number of calculations in one sample
MPC is validated by the case studies in this paper. period. For the purpose of analysis, four arithmetic and basic
operations (e.g., addition, multiplication, exponential function
II. CONVENTIONAL MODEL PREDICTIVE operation) are all considered as one calculation. K is the number
CONTROL of possible switching states. np is the number of calculations of
MPC can be divided into two categories: finite control set prediction function p, while nc is the number of calculations of
MPC (FCS-MPC) and continuous control set MPC (CCS- cost function g. Both np and nc depend on the complexity of the
MPC). Fundamental principles, control performances and controlled system. It is obvious that the computational burden
major challenges of these two MPC methods are discussed in increases rapidly when the number of levels or phases of the
this section. power converter increase. With the increasing number of power
switches, the coefficient K increases exponentially, while p and
A. FCS-MPC and CCS-MPC g will be more complex, resulting in increased np and nc.
The general principle of MPC is shown in Fig. 1. The With CCS-MPC, the controller can achieve higher control
controlled system is a power converter with n sets of possible precision but demands more heavy computation compared with
switching states. The MPC is composed of two parts: prediction FCS-MPC [24]:
and optimization. N CCS-MPC  N FCS-MPC (2)
With the FCS-MPC method, each power switch maintains a Furthermore, it can be concluded from Fig. 1 that the quality
certain state (on or off) throughout a sampling period. The of system model, that is, the accuracy of prediction function p,
control actions si are finite in number i = 1 , . . ., n. In the directly determines the accuracy of the predicted system states

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

H=[H1 H2 … Hm]T Independent Dependent

I=[I1 I2 … Il]T O=[O1 O2 … On]T
ANN
H1 O1
Variable Variable
I1
I2 O2 Constraint
H2
conditions
H3 s(k)/d(k)
Il On
Hm xr(k+1)
Output
Input
Hidden Layer
W[1], b[1], g(ꞏ)[1] Layer W[2], b[2], g(ꞏ)[2] x(k)
(a)

I1 l
H1  g [1] ( w1[1]n I n  b1 )
H1 m
O1  g [2] ( w1[n2] H n  b1 )
O1 I O
n 1 n 1 Fig. 3. The MPC is essentially treated as a mapping system from the system
(b)
states, references, and constraint conditions to the control signals in essence. It
Fig. 2. A typical two-layer neural network with different number of neurons,
can be replaced by an ANN with the same control performance.
weight matrixes, biases, and active functions of hidden and output layers.
and further affects the selection of the control actions in the prediction and cost functions during the on-line real-time
optimization process. For a complex multivariable system, a control process.
high-order mathematical model may be needed to precisely As shown in Fig. 3, ANN-MPC use the same input as MPC
describe its operation process. However, such models are (i.e., the sampled states, references, and constraint conditions).
limited by the practical computing power of the controller. Therefore, the dimension of I is equal to the number of MPC
controller’s input elements. Practically, the ANN-MPC
controller first normalizes every input element into an
III. CONCEPT of ANN-MPC
appropriate range e.g., [−1, 1] through (5), helping to achieve a
This section introduces ANN, and explains how and why it better distribution of data. Ii stands for each element of the
should be used in MPC. The ANN-MPC is then introduced and controller’s input.
discussed, including the general principle, network structure, I  [ x x r s.t.]T
off-line training method, and advantages.
 I i  min( I i ) 
A. Artificial Neural Network Based MPC I i*  2   1 (5)
 max( I i )  min( I i ) 
According to Section II, MPC can be essentially abstracted
into a function. The MPC controller’s input i.e., sampled system I * =[ I1* I 2*  I i*  I l* ]T
sates (x), references (xr), and constraint conditions (s.t.), can be Note that, in this paper, the output layer of ANN-MPC is
treated as independent variables. The controller’s output, such designed to directly offer the control signal for each
as control signals (s or d), can be treated as dependent variables. independent power switch (pair). Its number of neurons is equal
The whole control process can be treated as a complex to the number of power switches (pairs). In this way, the
nonlinear mapping: structure of ANN is very intuitive, and the training set can be
s ( k ) or d ( k )  f ( x ( k ), x r ( k +1), s .t . ) (3) easily achieved. Moreover, with the increase of power switches,
ANN has emerged as an excellent solution for input-output the size of output layer in ANN-MPC increase linearly.
mapping problems [13]. In this paper, a very important feature Therefore, the ANN-MPC is suitable for controlling multicell
of ANN—its hierarchical mathematical expression is power converters. Currently, there is no way to accurately
emphasized. In this way, very complex nonlinear mapping can obtain the number of neurons in hidden layer. In this work, the
be described by simple mathematical expressions. The basic suggested number of nodes in the hidden layer is as few as
concept of ANN-MPC proposed in this paper is to set up an possible while ensuring the fitting accuracy. This helps to
ANN to replace the calculation process of the conventional MPC reduce the computational complexity.
in controlling the actual operation of the power converter in real Equation (6) is the calculation process of ANN-MPC derived
time. The heavy-duty computation in the conventional MPC is from (4). It is obvious that the dimensions of I* and s depend
shifted to the off-line training process. on the number of controller’s input elements and power
1) Structure of ANN switches, respectively:
A small-scale network is promising to be adopted in order to H = g [1] ( W [1] I * + b [1] )
(6)
ensure the real-time response of controller. According to s or d  g [ 2 ] ( W [ 2 ] H  b [ 2 ] )
existing researches, two-layer ANN is successfully used in MPC The zero-centered tanh function as shown in Eq. (7) is
application [25]. Fig. 2 shows a classical two-layer neural suggested as g[1].
network with the following mathematical expressions: n
1  e 2x
H = g [1] ( W [1] I + b [1] ) Y Y  (-1, 1) (7)
(4) 1  e  2x
O  g [2]
(W [2]
Hb [2]
)
A very important point is the selection of the output layer’s
where I is the input vector of the network, H and O are the activation function g[2]. Generally, for ANNs, the commonly
output vectors of hidden layer and output layer, respectively. used g[2] is linear function in function approximation
W[1] and W[2] are the weight matrixes of the hidden and output application [25], and Softmax function in multi-classification
layer, respectively, while b[1] and b[2] are the bias vectors. g[1] application [18][19]. However, the number of neurons in output
and g[2] are the active functions of each layer. Thus, the whole layer using Softmax function increase rapidly with the
controller only needs to calculate (4) instead of the 2(x-1)y exponential increase in possible switching states. As mentioned

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Simulation or PHIL experiment TABLE I

Dimension of different vectors or matrixes in a two-layer ANN
Vector/Matrix Dimensionality (row x column)
[x] (sr) I lx1
Load
MPC Controller H, b[1] mx1
[xr, s.t.]
O, b[2] nx1
W[1] mxl
① Generating training data W[2] nxm
+
[x, xr, s.t.] s - Error ANN-MPC has unique advantages such as simple
Ʃ mathematical expressions and excellent approximation
capacity, providing a general solution to the computational
② Forward propagation ③ Calculating error complexity challenge of using MPC in power electronic
applications. Assume that the number of input elements and
 
Data Flow ,
w b neurons in the hidden and output layers are l (input), m (hidden
Training Flow ④ Back propagation layer), and n (output layer), respectively, then the dimension of
each vector/matrix can be estimated as listed in Table I. The
Fig. 4. Offline training process of the ANN-MPC with N x i training samples number of calculations of an ANN with dimensions listed in
under circuit simulation or PHIL simulation environment. Table I can be expressed as:
above, in ANN-MPC, each neuron of output layer offers control N ANN-MPC  2lm  2mn   1 m   2 n (9)
signal to its corresponding power switch. For MPC, the output Where ε1 and ε2 are number of calculations of active function
signals of controller is a set of 0-1 sequences (FCS-MPC) or g[1] and g[2], respectively (In this paper, ε1 and ε2 are equal to 4
duty cycles (CCS-MPC). The former is essentially a multilabel and 3, respectively). This computational burden is significantly
classification problem (i.e., mutually inclusive binary reduced in comparison to Eq. (1) or (2). More importantly,
classification in each output), and the latter one is a function unlike the the exponential increase in computational burden in
approximation problem. For an ANN based binary classifier, the conventional MPC, the computational burden of the ANN-
sigmoid function as shown in Eq. (8) is a suitable g[2] [13]. MPC approach increase only moderately with increasing
Moreover, for a CCS-MPC controller, the value of output duty system complexity. Therefore, ANN-MPC is suitable for
cycle is between 0 and 1, which fits well with the range of controlling complex systems with control performance similar
sigmoid function. Therefore, sigmoid function is suggested as to the conventional MPC but much reduced computational
the general active function g[2] in ANN-MPC. burden.
n
1 Moreover, to overcome the difficulty in modeling complex
Y x
Y  (0, 1) (8)
1 e multivariable systems, advanced modeling techniques can be
2) Training process of ANN adopted to build the system models [25]. These models can be
The ANN is iteratively trained offline as shown in Fig. 4. A accuracy enough but will also increase the computational
power converter with a conventional MPC controller is first complexity of controller. In this paper, these advanced
designed and operated under circuit simulation or power modeling methods can be used to set up the training set and then
hardware-in-the-loop (PHIL) simulation environment. The train the ANN-MPC controller offline. Thus, the ANN-MPC
input ([x, xr, s.t.]) and output (s or d) data of this “virtual” can control a complex system online with much reduced
controller are collected as the input and target of training set for computational complexity.
ANN-MPC, respectively. 2) Compensate changes in the model parameter values
Remark 1: The range of training set should be wide enough The performance of the MPC controller depends on the
to cover operation area from no load to full load of the converter; quality of the prediction model. However, the system
Remark 2: The distribution of training set should be uniform, parameters vary unpredictably in most applications. Therefore,
preventing the ANN from overfitting or underfitting. adaptation techniques need to be considered in order to
Generally, the number of collected samples is quite large (e.g. compensate potential shifts of the model parameters. Observers
tens/hundreds of thousands or larger) to ensure adequate data are often used to adjust the model parameters of the
coverage. A neural network is then trained offline with the conventional MPC controller [26]-[28]. For ANN-MPC,
recorded samples. Backpropagation (BP) algorithm is one of however, there is no modifiable system model. A solution is to
the most commonly used training method for neural networks set the corrected system parameters as a kind of input elements
with excellent fitting precision [13], and is adopted to train the for the ANN, i.e. I= [x, xr, system parameters] T. The corrected
ANN-MPC in this work. During the training process, the ANN system parameters can be obtained using observers. Thus, the
estimates the control signals (s or d) using the recorded [x, xr, ANN-MPC after incorporating the observer can compensate
s.t.] and compared with the recorded control signals (sr or dr) changes in the model parameter values and ensure sufficient
of the “virtual” MPC. The error is then used to update the control robustness for system parameter uncertainties.
weight matrixes and biases of ANN via BP algorithm. This Similarly, the ANN-MPC controller becomes compatible with
process repeats a number of iterations until the ANN outputs other extended MPC control techniques by flexibly setting the
the same control signals as the conventional MPC. input elements.
B. Advnatages of ANN-MPC
1) Solution for High-order Complex Systems

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Cell4 Cell3 Cell2 Cell1 Converter where ir is the reference value of the load current. vrci is the
reference value of the Ci voltage which is in descending order
+ S4 S3 S2 S1
from dc bus to load. vci(k+1) is the predictive value of the Ci
0.5Vdc iL
- + + + voltage. Weight coefficient wci is used to weigh the proportion
vc3 vc2 vc1
C3- - C1- of each status variable.

Z=R+jwL
C2 +
vo MATLAB/Simulink is used to compare different MPC
-
methods. The system specifications are summarized in Table II.
A FCS-MPC controller and an ANN-MPC controller are set up
iL, vC1, vC2, vC3
MPC sr1, sr2, sr3, sr4 respectively to control the load current and FC voltages of the
irL
Controller FCS-MPC 5L FC converter. The ANN-MPC controller is trained by the
simulation data of the FCS-MPC controller and the parameters
Target of the ANN is listed in Table III. In this work, the operating
Input process of the single-phase 5L converter is simulated in two
Sampling frequency: 100kHz Data Collection
fundamental periods (0.04s). The initial value of each FC
voltage is [0V, 0V, 0V] (A set), and [150V, 250V, 350V] (B
Fig. 5. Schematic of single-phase 5L FC converter with FCS-MPC controller set), respectively. The reference value of each FC voltage is
as supervisor for ANN-MPC. [100V, 200V, 300V]. The reference load currents are pre-set
sinusoidal waveforms with amplitudes of 6/ 8/ 10/ 12/ 14/ 16/
IV. SIMULATION STUDY 18/ 20A, respectively. The three FCs will be charged to their
In this section, the proposed ANN-MPC is validated by corresponding reference values in the first fundamental period,
simulation study of two multilevel and multiphase systems: a and then remain stable in the next fundamental period. By this
single-phase 5-level (5L) flying capacitor (FC) converter and a means, the range of the training set can cover all the possible
three-phase 3-level (3L) FC converter. Control performance operating area of load current and FC voltages during both
and computational burden are compared between the transient process and stable processes. To make the distribution
conventional FCS-MPC and the proposed ANN-MPC. of training set be uniform, the simulated load current, FC
voltages and the corresponding control signals are sampled and
A. Case Study 1: Single-phase 5-level FC Converter recorded at a fixed frequency of 100kHz. In total, the training
A single-phase 5L FC converter is shown in Fig. 5 with each data are made up of 64000 samples, 85% of which are used to
modular cell composed of one capacitor Ci and two switches Si set up the training set and the others are adopted to set up the
and 𝑆 operating in a complementary manner. To develop test set.
mathematical models for the FC converter, we have Fig. 6 shows the load current and FC voltages waveforms of
V the 5L FC converter under a step-up change in the amplitude of
vo (k )  s4 (Vdc  vc3 (k ))    s1vc1 (k )  dc ir amplitude changes from 8A to 16A
2 20 400
iL
v ( k )
i p (k  1)  i(k )eTR / L  o (1  eTR / L ) (10) ir vC3
R 10 300

Voltage (V)
iL, ir (A)

T vC2
vcp (k  1)  vci (k )  (i(k )  i p (k  1))(si 1  si ) 0 200
i
2Ci
where vci(k) is the voltage drop on Ci, i(k) is the load current, vC1
-10 100
vo(k) is the output voltage at the beginning of sample period
t=kT, ip(k+1) is the predictive load current, vpci(k+1) is the
-20 0
predictive voltage drop on Ci. Switching function 𝑠 , 𝑠 ∈ 0,1 . 0.00 0.02 0.04 0.06 0.08
If si=0, the top switch of i cell is off, and vice versa. Time (s)
In this paper, the cost function is given by tracking error of (a)
the load current and the voltage drop on each flying capacitor ir amplitude changes from 8A to 16A
20 400
g   ir  i(k 1)   wci  vcri  vci (k 1)
2 2
iL
(11)
i 1
ir vC3
10 300
TABLE II
Voltage (V)
iL, ir (A)

5L-FCMC Parameters vC2

Parameters Value 0 200
Vdc 400V
Flying capacitor 400μF vC1
-10 100
Sample frequency 100kHz

TABLE III -20 0

0.00 0.02 0.04 0.06 0.08
ANN Parameters
Time (s)
Structure Specification (b)
Nodes of input/hidden/output layer 5/17/4 Fig. 6. 1P5L FCMC dynamic response of (a) FCS-MPC and (b) ANN-MPC
Active function of hidden layer Tanh funciton with RL load of 4mH/10Ω and a step-up change in the amplitude of reference
Active function of output layer Sigmoid function load current from 8A to 16A.

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Phase-C ix amplitude changes from 8A to 16A

Phase-B

ia, ib, ic (A)

10 ia
Phase-A
0 ib
S2a ic
ic
-10
ib Filter Load
S1a
Vdc C1a ia

Voltage (V)
210 vaC
S1a 200 vbC
vcC
190
S2a
0.00 0.02 0.04 0.06 0.08
Fig. 7. Schematic of the 3-phase 3L FC converter Time (s)
reference load current from 8A to 16A for the FCS-MPC and (a)
ix amplitude changes from 8A to 16A
ANN-MPC, respectively. The load is a 4mH inductor in series
with a 10Ω resistor. It is observed that the load current and FC 10

ia, ib, ic (A)

voltages can effectively track the reference values under steady ia
0 ib
state or dynamic conditions for both the FCS-MPC and ANN-
ic
MPC methods. The dynamic response under the step-up change -10
in the amplitude of load current reference from 8A to 16A at t
= 0.35s is shown in Fig. 6. The load current jumps rapidly and 210

Voltage (V)
vaC
tracks the reference value. 200 vbC
vcC
B. Case Study 2: 3-phase 3-level FC Converter 190
A 3-phase 3L FC converter is selected as a multi-phase multi-
0.00 0.02 0.04 0.06 0.08
level case study. 3-phase converters are one of the most Time (s)
widespread converter topologies with many well established (b)
control strategies. The circuit topology of the 3-phase 3L FC Fig. 8. 3P3L FCMC dynamic response of (a) FCS-MPC and (b) ANN-MPC
with RL load of 4mH/10Ω and a step-up change in the amplitude of reference
converter is shown in Fig. 7. The Six, 𝑆 and Cix form a modular load current from 8A to 16A.
cell. For a 3-phase system, to decouple the controlled objects,
Accordingly, the cost function is adjusted to:
the coordinates of system state variables are usually
transformed. In this case, the load currents are transformed from g  wi (i (k  1)  ir ) 2  wi (i (k  1)  ir ) 2
the static 3-phase coordinates to the static α-β coordinates, and  (v1a (k  1)  v1ra ) 2  (v1b (k  1)  v1rb ) 2 (13)
the prediction functions of 3-phase 3L FC converter can be
 (v1c (k  1)  v ) r 2
expressed as 1c

v xN (k )  S 2 x (Vdc  v1x (k ))  s1x v1x (k ) MATLAB-Simulink simulation is performed to compare

1 different MPC methods. The system specifications are
v nN (k )   vaN (k )  vbN (k )  vcN (k )  summarized in Table IV. The ANN-MPC controller for the 3-
3
phase 3L FC converter is set up as shown in Table V.
i xp (k  1)  ix (k )e TR / L  (1  e TR / L )(vxN (k )  vnN ( k )) / R
(12) Fig. 8 shows the 3-phase load current and 3-phase FC
T voltages of the 3L FC converter with FCS-MPC and ANN-
v1 x (k  1)  v1x (k ) 
p
(i p (k  1)  ix (k ))( S 2 x  S1x )
2C1x x MPC controller, respectively. The load currents and FC

i p (k  1) i p (k  1)   f clark i p (k  1) i p ( k  1) i p (k  1)
    voltages can track the reference values very well under the
steady state and transient conditions. The transient behavior that
a b c

where x∈{a, b, c}, v1x(k) is the voltage drop on C1x, ix(k) is the a step-up change in the amplitude of load current reference from
load current, vxN(k) is the output voltage of each phase at the 8A to 16A at t = 0.4s is shown in Fig. 8. The load current jumps
beginning of sample period t=kT, ixp(k+1) is the predictive load rapidly and tracks the reference value.
current, vp1x(k+1) is the predictive voltage drop on C1x. iαp(k+1),
iβp(k+1) are the predictive load current in α-β coordinates. C. Computational Complexity Analysis
The computational burden of the ANN-MPC and FCS-MPC
TABLE IV
3P3L FCMC Parameters are analyzed by (1) and (9). Table VI compares the
Parameters Vaule computational burden for the two converter case studies for
Vdc 400V FCS-MPC and ANN-MPC, respectively. For the 1P5L FCMC,
Flying capacitor 200μF the number of available switching states is 16. The number of
Sampling frequency 100kHz calculations of prediction function is calculated to be 37, and
TABLE V TABLE VI
ANN Parameters Comparsion of Computational Burden
Structure Specification # of calculation in one sampling cycle
Control Method
Nodes of input/hidden/output layer 8/42/6 1-phase 5-level 3-phase 3-level
Active function of hidden layer Tanh funciton FCS-MPC 816 4992
Active function of output layer Sigmoid function ANN-MPC 386 1362

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

The FPGA control board is connected

at the back of the platform irL

Actual FPGA Controller

(a)

Oscilloscope irL
RT-LAB Host PC
Platform

(b)
Fig. 9. Control hardware-in-the-loop (CHIL) simulation setup using RT-Lab to
Fig. 11. Experimental 1P5L FCMC dynamic response of (a) FCS-MPC and (b)
simulate the power stage and a real FPGA MPC controller.
ANN-MPC with RL load of 4mH/10Ω and a step-up change in the amplitude
that of cost function is 14. The number of calculation in one of reference load current from 8A to 16A.
period is 816 and 386 for FCS-MPC and ANN-MPC, is used to simulate the power converter while an FPGA-based
respectively, indicating a 211% reduction by ANN-MPC. controller (Artix-7 100T AX7102) is used as the MPC
Similarly, for the 3P3L FCMC, the number of available controller to compare different methods. The 5L-FCMC
switching states is 64. The number of calculation in one period parameters and ANN parameters are the same as that listed in
is 4992 and 1362 for FCS-MPC and ANN-MPC, respectively, Table II & III, respectively.
indicating a 367% reduction by ANN-MPC. Furthermore, Fig. 10 shows the load current and FC voltages waveforms
ANN-MPC mainly involves simple addition and multiplication of the 5L FC converter under a step-up change in the frequency
calculations while the types of calculations in FCS-MPC of reference load current from 50Hz to 100Hz with the FCS-
depend on the complexity of the system model and constraint MPC and ANN-MPC, respectively. The load is a 4mH inductor
conditions. in series with a 10Ω resistor. For both the FCS-MPC and ANN-
MPC methods, the load current and FC voltages can effectively
V. EXPERIMENT RESULTS track the reference values under steady state or dynamic
Experiment is performed to further validate the proposed conditions. During steady state, under ANN-MPC, the THD of
ANN-MPC method by control hardware-in-the-loop (CHIL) load current with amplitude of 16A is 0.71% while the ripple
simulation platform shown in Fig. 9. A RT-Lab HIL platform rate of flying capacitor voltage vC1, vC2, and vC3 are 4.45%,
2.43%, and 1.61%, respectively. And, under FCS-MPC, the
THD of load current with amplitude of 16A is 0.38% while the
ripple rate of flying capacitor voltage vC1, vC2, and vC3 are 3.92%,
2.08%, and 1.38%, respectively. The dynamic response under

iL=16.9A

iL=16.2A

(a)

(a)

iL=16.2A

(b) (b)
Fig. 10. Experimental 1P5L FCMC dynamic response of (a) FCS-MPC and (b) Fig. 12. Experimental dynamic response of the 5L FC converter under ANN-
ANN-MPC with RL load of 4mH/10Ω and a step-up change in the frequency MPC (a) without and (b) with load resistance as input element under a step
of reference load current from 50Hz to 100Hz. change of the load resistor from 10Ω to 2Ω.

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

TABLE VII parameters, and the load current keep tracking the reference
Comparison of control performance between FCS-MPC and ANN-MPC without such an error.
Performance FCS-MPC ANN-MPC
THD 0.38% 0.71%
Response time 0.4ms 0.4ms VI. CONCLUSION
vC1 ripple rate 3.92% 4.45% In this paper, a new artificial neural network based model
vC2 ripple rate 2.08% 2.43%
vC3 ripple rate 1.38% 1.61% predictive control concept (ANN-MPC) is proposed for power
converter. An artificial neural network (ANN) is then trained
the step-up change in the frequency of load current reference offline with the input and output data of the virtual MPC
from 50Hz to 100Hz at t = 0.5s is shown in Fig. 10. The controller. Next, an actual FPGA-based MPC controller is
frequency of load current changes rapidly and tracks the designed using the trained ANN instead of relying on heavy-
reference value. Fig. 11 shows the system states of the 5L FC duty mathematical computation to control the actual operation
converter under a step-up change in the amplitude of reference of the power converter in real time. Thanks to the simple
load current from 8A to 16A with the FCS-MPC and ANN- mathematical expression and excellent approximation
MPC, respectively. It is observed that the load current and FC capability of the artificial neural network, the ANN-MPC offer
voltages can effectively track the reference values under steady a general solution to overcome the computational barriers of
state or dynamic conditions for both the FCS-MPC and ANN- conventional MPC. The main advantages of the ANN-MPC
MPC methods. The dynamic response under the step-up change include retaining the MPC control performance with
in the amplitude of load current reference from 8A to 16A at t significantly reduced computational complexity, and allowing
= 0.5s is shown in Fig. 11. During dynamic process, with FCS- the use of more accurate high-order system models due to the
MPC, the load current can track the reference value in 0.4ms. simple mathematical expression of ANN. This is particularly
While, under ANN-MPC, the load current can also track the important for multicell power systems as their number of
reference value in 0.4ms. According to the above analysis, the switching states increases exponentially. Furthermore, the
ANN-MPC shows comparable control performance of the ANN-MPC approach using system parameters as input
conventional MPC controller. The control performance of both elements can retain the control robustness of the conventional
FCS-MPC and ANN-MPC are summarized in Table VII. MPC control. The ANN-MPC concept is validated by both
The execution time and the computing resource usage of the simulation and experimental results on a single-phase 5-level
FPGA are compared between FCS-MPC and ANN-MPC in FCMC and a 3-phase 3-level FCMC converter. It is
Table VIII. Both MPC schemes are implemented using the demonstrated that the FPGA-based ANN-MPC controller can
identical FPGA optimization method, that is, the high-level- significantly reduce the FPGA resource requirement (e.g., 2.11
synthesis technique. The basic frequency of the FPGA is times fewer slice LUTs and 2.06 times fewer DSPs) while
200MHz. Note that only the actual computational times are offering a control performance same as the conventional MPC.
considered for the comparison while the waiting times of
controller are not counted. It can be seen from Table VIII that, ACKNOWLEDGMENT
ANN-MPC controller requires 2.11 times fewer slice LUTs, 2.2
times fewer flip flops and 2.06 times fewer DSP48Es than FCS- J. Rodriguez acknowledges the support of ANID through
MPC. In addition to the significantly reduced FPGA resource, projects FB0008, ACT192013 and 1210208.
ANN-MPC offers 15% less computation time than FCS-MPC.
A load step change case is studied to validate the robustness REFERENCES
of ANN-MPC controller for load parameter uncertainties. In [1] P. Cortes, M. P. Kazmierkowski, R. M. Kennel, D. E. Quevedo, and JosÉ
Rodriguez “Predictive Control in Power Electronics and Drives,” IEEE
this case, the estimated value of load resistance is set as the Transactions on Industrial Electronics, vol. 55, no. 12, pp. 4312-4324, Dec.
input element of ANN. Fig. 12 shows the control performance 2008.
of the ANN-MPC (a) without and (b) with load resistance as [2] S. Kouro, P. Cortes, R. Vargas, U. Ammann, J. Rodriguez, “Model
input element under a step change of the load resistance from Predictive Control—A Simple and Powerful Method to Control Power
Converters,” IEEE Transactions on Industrial Electronics, vol. 56, no. 6,
10Ω to 2Ω. The amplitude of reference current is 16A in this pp. 1826-1838, Jun. 2009.
case. As shown in Fig. 12, the 5L converter is well controlled [3] J. Scoltock, T. Geyer and U. K. Madawala, "Model Predictive Direct Power
by both ANN-MPC controllers with initial resistance value. Control for Grid-Connected NPC Converters," in IEEE Transactions on
However, if the load resistance decreases, the load current Industrial Electronics, vol. 62, no. 9, pp. 5319-5328, Sept. 2015.
[4] M. Tomlinson, H. d. T. Mouton, R. Kennel and P. Stolze, "A Fixed
cannot accurately track the reference and will exceed the Switching Frequency Scheme for Finite-Control-Set Model Predictive
reference value by about 1A with ANN-MPC (a). For ANN- Control—Concept and Algorithm," in IEEE Transactions on Industrial
MPC (b) using load resistance as one of its input elements, the Electronics, vol. 63, no. 12, pp. 7662-7670, Dec. 2016.
neural network can compensate the change in the system [5] J. Park, D. J. Kim, and K. Lee, "Predictive Control Algorithm Including
Conduction-Mode Detection for PFC Converter," in IEEE Transactions on
TABLE VIII Industrial Electronics, vol. 63, no. 9, pp. 5900-5911, Sept. 2016.
FPGA Resource Utilization [6] S. Vazquez, A. Marquez, R. Aguilera, D. Quevedo, J. I. Leon and L. G.
Franquelo, "Predictive Optimal Switching Sequence Direct Power Control
FCS-MPC ANN-MPC
for Grid-Connected Power Converters," in IEEE Transactions on Industrial
Execution times (Tclk=5ns) Electronics, vol. 62, no. 4, pp. 2010-2020, April 2015.
Ttotal 1.32μs (264 Tclk) 1.12μs (225 Tclk) [7] F. Defay, A. Llor and M. Fadel, "Direct Control Strategy for a Four-Level
Resources (number of used units) Three-Phase Flying-Capacitor Inverter," in IEEE Transactions on
Slice LUTs 11047 5235 Industrial Electronics, vol. 57, no. 7, pp. 2240-2248, July 2010.
Flip Flops 13359 6150
DSP48E 107 52

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

[8] Q. Huang, A. Q. Huang, “Review of GaN Totem-Pole Bridgeless PFC,” Transactions on Power Electronics, vol. 33, no. 10, pp. 9064-9075, Oct.
CPSS Transactions on Power Electronics and Applications, vol. 2, no. 3, 2018.
pp. 187-196, Sept. 2017.
[9] M. Norambuena, C. Garcia, J. Rodriguez, P. Lezana, “Finite Control Set
Model Predictive Control Reduced computational cost applied to a Flying
Capacitor Converter,” IECON 2017, pp. 4903-4907, Oct. 2017. Daming Wang (S'18) received the B.S. and M.S.
[10] Z. Zhang, C. M. Hackl, R. Kennel, “Computationally Efficient DMPC for degrees in electrical engineering from the
Three-level NPC Back-to-Back Converters in Wind Turbine Systems With College of Electrical and Information Engineering,
PMSG,” IEEE Transactions on Power Electronics, vol. 32, no. 10, pp. Hunan University, Changsha, China, in 2013 and
8018-8034, Oct. 2017. 2016, respectively. He is currently pursuing the
[11] M. Siami, D. A. Khaburi, M. Rivera, and J. Rodríguez, “A Ph.D. degree in electrical engineering with the
Computationally Efficient Lookup Table Based FCS-MPC for PMSM College of Electrical and Information Engineering,
Drives Fed by Matrix Converters,” IEEE Transactions on Industrial Hunan University, Changsha, China.
Electronics, vol. 64, no. 10, pp. 7645-7654, Oct. 2017. His current research interest includes
[12] A. Mora, R. C. Dobson, R. P. Aguilera, A. Angulo, F. Donoso, and J. applications of WBG power semiconductor
Rodriguez, “Computationally Efficient Cascaded Optimal Switching devices, model predictive control and advanced
Sequence MPC for Grid-Connected Three-Level NPC Converters,” IEEE process control.
Transactions on Power Electronics, vol. 34, no. 12, pp. 12464-12475, Dec.
2019.
[13] Simon Haykin, Neural Networks and Learning Machines. Upper Saddle Zheng John Shen (Fellow, IEEE) received the
River, NJ: Prentice Hall, 2008. B.S. degree from Tsinghua University, Beijing,
[14] X Liu, L. Qiu, Y. Fang, et al., “Finite-Level-State Model Predictive Control China, in 1987, and the M.S. and Ph.D. degrees
for Sensorless Three-Phase Four-Arm Modular Multilevel Converter,” from Rensselaer Polytechnic Institute, Troy, NY,
IEEE Transactions on Power Electronics, vol. 35, no. 5, pp. 4462-4466, USA, in 1991 and 1994, respectively, all in
Sept. 2019. electrical engineering. Between 1994 and 1999,
[15] T. Dragičević, M. Novak, “Weighting Factor Design in Model Predictive he held a variety of positions including Senior
Control of Power Electronic Converters: An Artificial Neural Network Principal Staff Scientist with Motorola. He was on
Approach,” IEEE Transactions on Industrial Electronics, vol. 66, no. 11, the Faculty of the University of Michigan-
pp. 2124 - 2142, Oct. 2018. Dearborn, from 1999 to 2004, and the University
[16] S. Lucia, D. Navarro, B. Karg, H. Sarnago, O. Lucía, “Deep Learning- of Central Florida, from 2004 to 2012. He also
based Model Predictive Control for Resonant Power Converters,” IEEE served as a BoardMember and Chief Scientist of GWSSemiconductor,
Transactions on Industrial Informatics, Jan. 2020. Tempe, AZ, USA, (now a division of Renesas Electronics) from 2002 to
[17] I. S. Mohamed, S. Rovetta, T. D. Do, T. Dragicević and A. A. Z. Diab, "A 2012, responsible for developing megahertz-frequency lateral power
Neural-Network-Based Model Predictive Control of Three-Phase Inverter MOSFET technology for supercomputer and server applications. In
With an Output LC Filter," in IEEE Access, vol. 7, pp. 124737-124749, 2013, he joined the Illinois Institute of Technology, as Grainger Chair
2019. Professor in electrical and power engineering. His research interests
[18] M. Novak, and T. Dragicevic, “Supervised imitation learning of finite set include power electronics, power semiconductor devices and ICs,
model predictive control systems for power electronics,” IEEE automotive electronics, renewable, and alternative energy systems. He
Transactions on Industrial Electronics, Jan. 2020. has authored or coauthored 300 journal and conference articles, and
[19] S. Wang, T. Dragicevic, Y. Gao, R. Teodorescu, “Neural Network based holds 18 issued and several pending U.S. patents in these areas.
Model Predictive Controllers for Modular Multilevel Converters,” IEEE Dr. Shen is the recipient of the 2012 IEEE Region 3 Outstanding
Transactions on Energy Conversion, Sept. 2020. Engineer Award, the 2003 NSF CAREER Award, and two IEEE
[20] J. Chen, Y. Chen, L. Tong, L. Peng, Y. Kang, “A Backpropagation Neural Transactions Paper Awards. He has been an active volunteer in the
Network-Based Explicit Model Predictive Control for DC–DC Converters IEEE Power Electronics Society (PELS), and has served as VP of
With High Switching Frequency,” IEEE Journal of Emerging and Selected Products, AdCom Member, General Chair or Technical Program Chair
Topics in Power Electronics, vol. 8, no. 3, pp. 2124 - 2142, Sept 2020. of several major IEEE conferences including ECCE and ISPSD. He is a
[21] D. Wang, X. Yin, S. Tang, C. Zhang, Z. J. Shen, J. Wang, and Z. Shuai, fellow of the U.S. National Academy of Inventors.
“A Deep Neural Network Based Predictive Control Strategy for High
Frequency Multilevel Converters,” 2018 IEEE Energy Conversion
Congress and Exposition, pp. 2988-2992, Sept. 2018. Xin Yin (M'16) received the B.S., M.S., and Ph.D.
[22] F. Morel, X. Lin-Shi, J.-M. Retif, B. Allard, and C. Buttay, “A comparative degrees in electrical engineering from the
study of predictive current control schemes for a permanent-magnet College of Electrical and Information Engineering,
synchronous machine drive,” IEEE Trans. Ind. Electron., vol. 56, no. 7, Hunan University, Changsha, China, in 1993,
pp. 2715–2728, Jul. 2009. 2000, and 2011, respectively.
[23] J. Scoltock, T. Geyer, and U. K. Madawala, “A comparison of model He joined Hunan University as an Associate
predictive control schemes for MV induction motor drives,” IEEE Trans. Research in 2006. His current research interests
Ind. Informat., vol. 9, no. 2, pp. 909–919, May 2013. include power electronics applications and circuit
[24] A. A. Ahmed, B. K. Koh and Y. I. Lee, "A Comparison of Finite Control fault diagnosis.
Set and Continuous Control Set Model Predictive Control Schemes for Dr. Yin received the Ministry of Education
Speed Control of Induction Motors," in IEEE Transactions on Industrial Science and Technology Progress Award. He
Informatics, vol. 14, no. 4, pp. 1334-1346, April 2018. has participated in the 863 program, the NSFC, and several projects
[25] X. Yuan, G. Huang, and K. Shi, “Improved Adaptive Path Following supported by the Natural Science Foundation of Hunan Province, China.
Control System for Autonomous Vehicle in Different Velocities,” IEEE
TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, vol.
21, no. 8, pp. 3247 - 3256, Aug. 2020. Sai Tang received the B.S. degree in nuclear
[26] J. A. Solsona, S. G. Jorge, and C. A. Busada, “Nonlinear control of a buck engineering and technology from the College of
converter which feeds a constant power load,” IEEE Trans. Power Nuclear Science Technology, University of South
Electron., vol. 30, no. 12, pp. 7193–7201, Dec. 2015. China, Hengyang, China, in 2013. He received
[27] C. Buccella, C. Cecati, H. Latafat, P. Pepe, and K. Razi, “Observer-based the M.S. degree in electric engineering in the
control of LLC dc/dc resonant converter using extended describing College of Electrical and Information Engineering,
functions,” IEEE Trans. Power Electron., vol. 30, no. 10, pp. 5881–5891, Hunan University, Changsha, China, in 2016. He
Oct. 2015. is currently pursuing the Ph.D. degree in
[28] L. Cheng, P. Acuna, R. P. Aguilera, J. Jiang, S. Wei, J. E. Fletcher, Dylan electrical engineering with the College of
D. C. Lu, “Model Predictive Control for DC–DC Boost Converters with Electrical and Information Engineering, Hunan
Reduced-Prediction Horizon and Constant Switching Frequency,” IEEE University, Changsha, China.

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 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
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721, IEEE
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2021.3076721
Transactions on Industrial Electronics

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

His current research interest includes power semiconductor devices he has been included in the list of Highly Cited Researchers published
and their applications. by Web of Science.

Xifei Liu received the B.S. degree in electrical Margarita Norambuena (S'12-M'14) received
engineering from Taiyuan University of the B.S. and M. Sc. degrees in electric
Technology, Taiyuan, China, in 2018. She is engineering from the Universidad Tecnica
currently working toward the M.S. degree in Federico Santa Maria (UTFSM), Valparaiso,
power electronics at Hunan University, Chile, in 2013. She received the Ph.D. degree
Changsha, China. (summa cum laude) in electronics engineering
Her research interests include matrix from the UTFSM in 2017 and received the
converter, digital control, and model predictive Doktoringenieur (Dr-Ing.) degree (summa cum
control of power converters and drives. laude) from the Technische Universitat Berlin
(TUB) in 2018.
She received a scholarship from Chilean National Research, Science
and Technology Committee (CONICYT) in 2014 to pursue the PhD
degree studies in power electronics at UTFSM and Technische
Universitaet Berlin (TUB). She also received a scholarship from German
Chao Zhang (S'18) received the B.S. degree in Academic Exchange Service (DAAD) in 2015 to pursue the PhD degree
electrical engineering and automation from the studies in TUB.\newline
College of Electrical Engineering, Guizhou She received the IEEE IES Student Best Paper Award in 2019 for her
University, Guiyang, China, in 2017. He is doctoral work. Currently she is an Assistant Professor at Universidad
currently pursuing the Ph.D. degree in electrical Tecnica Federico Santa Maria. Her research interest include multilevel
engineering with the College of Electrical and converters, model predictive control of power converters and drives,
Information Engineering, Hunan University, energy storage systems, renewable energy and electromobility.
Changsha, China.
His current research interest includes power
semiconductor devices and their applications.

Jun Wang (S'06-M'10-SM'15) received the B.S.

degree at Huazhong University of Science and
Technology, Wuhan, China in 2000. He received
the M.S. degree in Institute of Semiconductors,
Chinese Academy of Sciences, Beijing, China in
2003. He received the M.E. degree at University
of South Carolina, Columbia, SC, USA in 2005,
and the Ph.D. degree at North Carolina State
University, Raleigh, NC, USA in 2010
respectively, all in Electrical Engineering. He
then worked as a device design engineer at
Texas Instruments, Inc. in Bethlehem, PA, USA
between 2010 and 2013.
Dr. Wang became a professor in the College of Electrical and
Information Engineering at Hunan University in 2014.He has been an
associate editor of the IEEE Journal of Emerging and Selected Topics
in Power Electronics since 2017. His research interests include power
semiconductor devices and their applications in power electronics
systems.

Jose Rodriguez (M´81-SM´94-F´10-LF´20)

received the Engineer degree in electrical
engineering 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, in 1985. He has been with
the Department of Electronics Engineering,
Universidad Tecnica Federico Santa Maria,
since 1977, where he was full Professor and
President. Since 2015 he was the President and
since 2019 he is full professor at Universidad Andres Bello in Santiago,
Chile.
He has coauthored two books, several book chapters and more than
400 journal and conference papers. His main research interests include
multilevel inverters, new converter topologies, control of power
converters, and adjustable-speed drives. He has received a number of
best paper awards from journals of the IEEE. Dr. Rodriguez is member
of the Chilean Academy of Engineering. In 2014 he received the
National Award of Applied Sciences and Technology from the
government of Chile. In 2015 he received the Eugene Mittelmann Award
from the Industrial Electronics Society of the IEEE. In years 2014 to 2020

0278-0046 (c) 2021 IEEE. Personal©use is permitted,

2021 but republication/redistribution
IEEE. Personal requires IEEE permission.
use is permitted, but republication/redistribution See http://www.ieee.org/publications_standards/publications/rights/index.html
requires IEEE permission.See https://www.ieee.org/publications/rights/index.html forfor moreinformation.
more information.
Authorized licensed use limited to: Seoul National Univ of Science & Tech (SNUT). Downloaded on July 30,2021 at 13:23:52 UTC from IEEE Xplore. Restrictions apply.