Received 30 June 2024; revised 26 July 2024; accepted 27 July 2024.

Date of publication 30 July 2024;

date of current version 8 August 2024. The review of this article was arranged by Associate Editor Edris Pouresmaeil.
Digital Object Identifier 10.1109/OJIES.2024.3435862

Enhanced Inverse Model Predictive Control for

EV Chargers: Solution for Rectifier-Side
ALI SHARIDA 1,2 (Student Member, IEEE), ABDULLAH BERKAY BAYINDIR1,2 (Student Member, IEEE),
SERTAC BAYHAN 3,4 (Senior Member, IEEE), AND HAITHAM ABU-RUB 2,3 (Fellow, IEEE)
1
Texas A&M University, Doha, Qatar
2
Texas A&M University, College Station, TX 77840 USA
3
Qatar Environment and Energy Research Institute, Hamad Bin Khalifa University, Ar-Rayyan, Qatar
4
Department of Electrical-Electronic Engineering, Technology Faculty, Gazi University, Ankara 06560, Türkiye
CORRESPONDING AUTHOR: ALI SHARIDA (e-mail: [email protected])
This work was supported by Qatar National Research Fund (a member of Qatar Foundation) under Grant NPRP12C-33905- SP-213 and in part by Qatar Research,
Development and Innovation under Grant ARG01-0428-230023.

ABSTRACT Inverse model predictive control (IMPC) is a control technique that was recently proposed for
power electronic converters. IMPC inherits the advantages of model predictive control (MPC) in terms of
ability to handle complex and nonlinear systems and achieving multiple control objectives, while adhering to
various constraints. Unlike MPC, IMPC offers a significantly reduced computational burden by omitting the
iterative computations of the cost functions and states predictions. Nevertheless, both IMPC and MPC rely
significantly on the dynamic model of the power converter. This makes them susceptible to uncertainties and
disturbances. This article presents a novel technique to enhance the reliability and robustness of the IMPC
for electric vehicle chargers by treating the converter’s dynamic model as a black box. Then, an adaptive
estimation strategy employing a recursive least square algorithm is proposed for online dynamic model
estimation, which is then used by the IMPC for optimal switching states prediction. The key benefit of the
proposed technique is the utilization of an accurate and real-time estimated dynamic model, which facilitates
a reliable states prediction by the IMPC. The effectiveness of the proposed technique is demonstrated through
extensive simulations and experimental validation for a three-phase three-level T-type rectifier.

INDEX TERMS Adaptive control, bidirectional power flow, electric vehicle (EV) chargers, grid-to-vehicle
(G2V), inverse model predictive control (IMPC), multilevel converters, T-type rectifiers, vehicle-to-grid
(V2G).

NOMENCLATURE
Symbol Description N Number of available switching states.
Vabc Vector of grid voltages. kαβ Clarke transformation matrix.
Iabc Vector of grid currents. Vc Difference between dc side capacitors’ voltages.
uabc Vector of poles’ voltages. Wα , Wβ Weighting factors for α and β states.
Lg Input filter inductance. uOαβ Optimal poles voltages vector represented in the αβ
Rg Filter’s internal resistance. frame.
Vdc DC side voltage. uOabc Optimal poles voltages vector represented in the abc
C1 , C2 DC side capacitors. frame.
Sij Switching state of the ith switch in the jth leg. SOabc Computed optimal switching vector.
Snj Normalized pole voltage of the jth leg. S̄Oabc Digitalized optimal switching vector.
Iαβ Vector of grid currents represented in the αβ frame.
Vαβ Vector of grid voltages represented in the αβ frame.
Snabc Matrix contains all possible switching states for the I. INTRODUCTION
tree legs. Globally, electric vehicles (EVs) are gaining an increasing
Sabc MPC-computed optimal switching vector. popularity worldwide in transportation sector due to many
© 2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 5, 2024 795
SHARIDA ET AL.: ENHANCED INVERSE MODEL PREDICTIVE CONTROL FOR EV CHARGERS: SOLUTION FOR RECTIFIER-SIDE

advantages such as their fossil-fuel-free nature, low mainte- suddenly changing. Consequently, look-up table controllers
nance needs, and efficient motors. In addition, EVs offer faster are often integrated with additional algorithms to address ex-
acceleration compared to conventional vehicles [1]. Recently, ternal disturbances and model uncertainties.
EVs have been employed to support the grid by engaging in Multiple advanced control techniques were proposed for
vehicle-to-grid (V2G) and grid-to-vehicle (G2V) functional- bidirectional rectifiers in general and EV chargers in spe-
ities. These functionalities enable the EVs to store surplus cific to avoid the limitations of the aforementioned classical
energy in their batteries, while acting as distributed energy controllers. These methods include sliding mode controllers
storage systems. This stored energy can be utilized to support [10], passivity-based control [11] , Lyapunov-based control
the grid during peak demand periods. This offers enhancing [12], and fuzzy logic controller [13]. Sliding mode control
stability, increasing inertia, reducing peak demands, and low- is simple, robust, and contains minimized number of tuning
ering power requirements of the utility grid. However, V2G parameters. However, it is exposed to chattering which may
and G2V applications necessitate bidirectional ac–dc convert- deteriorate the tracking accuracy and generates unmodeled
ers and robust controllers to manage the power flow and its high frequency dynamics. Passivity-based control depends on
direction. Therefore, numerous control techniques were pro- the mathematical model to compensate or augment one or
posed to control the bidirectional ac–dc converters [2], [3]. more of system’s states, which represents the key limitation.
In [4], the double integral sliding mode control was in- Lyapunov-based mechanisms guarantee the global-stability of
troduced as a method to regulate the grid currents within a the control system, but cannot guarantee the steady-state er-
bidirectional EV charger. However, achieving a pure integra- ror, and cannot fulfill predefined control requirements. Fuzzy
tor in practical applications is challenging due to the presence logic controllers are simple, but they often encounter chal-
of high-frequency noise in sensor measurements. This noise lenges due to the absence of a systematic approach for
can impact the response of the controller and lead to the designing rules or memberships.
chattering problem. Symmetrical decoupling voltage compen- Recently, adaptive and robust control techniques have be-
sation principle was also utilized in the control of bidirectional come widely spread in fast EV charging applications due to
rectifiers [5]. Although achieving perfect state decoupling is their ability to handle disturbances and abnormal conditions of
theoretically feasible, practical implementation is highly chal- the grid. In [14], an adaptive control strategy was introduced
lenging due to model uncertainty, which adversely impacts to regulate the bidirectional ac–dc converter and mitigate
the controller’s transient response and increases steady-state current harmonics. The proposed controller comprises three
error. In [6], a cascaded PI current-voltage control with states primary algorithms: Fryze–Buchholz–Depenbrock for har-
decoupling algorithm was implemented, employing three PI monic current detection, droop control, and the recursive
controllers to regulate currents and voltages. However, this least squares (RLS) algorithm for load change estimation.
method also requires state decoupling. Direct power control However, estimating only load change is not sufficient, and
(DPC) techniques were extensively utilized in bidirectional the integration of multiple algorithms to control a single
power converters due to their abilities to regulate output power converter introduces higher complexity and computational
effectively [7]. Similar to PI, the key limitation of DPC meth- burden. Moreover, most of the ac–dc converters in fast charg-
ods is their reliance on the dynamic model to achieve state ing applications do not have a direct interaction with the EV,
decoupling. A multi-mode control strategy for a bidirectional which increases the complexity of dc-link impedance esti-
converter was presented in [8], which considers four dis- mation. Observer-based robust control techniques were also
tinct modes: normal operation, balanced current, unbalanced employed to enhance the robustness of EV chargers’ control
voltage compensation, and ripple-free dc voltage. However, systems as in papers [15] and [16]. In paper [15], a coupling-
all modes are governed by a proportional controller with a effect observer was proposed to eliminate the coupling effects
fixed proportional gain value. The proportional controller is between parallel rectifiers in a scalable EV charging system.
associated with several limitations, such as steady-state error However, the proposed observer did not include the uncertain-
and high overshoot. These limitations make proportional con- ties of the internal dynamics of each rectifier. On the contrary,
trollers incompatible with bidirectional converters. Multiple Sharida et al. [16] utilized Kalman filter to estimate the in-
look-up table-based control methods were proposed in the ternal uncertainties to enhance the robustness of the control
literature for bidirectional converters and hybrid microgrids system. Although Kalman filter is one of the most robust
[9]. However, look-up tables are not suitable for fast-dynamic observers, the initialization process plays a crucial role on its
systems that may be exposed to external disturbances and stability and performance especially at the beginning of the
uncertainties. Therefore, look-up table-based controllers are operation.
usually combined with additional adaptive algorithms to deal Model predictive control (MPC) is one of the most common
with external disturbances and model’s uncertainties, which controllers in power electronics applications in general due
increases the complexity and computational load of the con- to its significant advantages [17], [18]. Uniquely, MPC offer
troller. Several papers proposed look-up table-based control optimal control over a horizon, while adhering to multiple
methods for bidirectional converters and hybrid microgrids objectives and constraints. However, MPC has the highest
[9]. However, look-up tables are incompatible with fast- computational burden among all controllers and completely
dynamics systems where inputs and outputs are widely and rely on the dynamic model to perform states prediction and

796 VOLUME 5, 2024

FIGURE 1. Circuit diagrams of the considered EV charger.

cost function optimization. Although many approaches were TABLE 1. Three-Level T-Type Rectifier Switching States
proposed to reduce the computational time such as [19], [20],
and [21], many iterative cycles of predictions and cost func-
tion optimization are still required at each control cycle. In
terms of enhancing the robustness of the MPC, numerous
methods were proposed to compensate the disturbances and
uncertainties through disturbance observer [22]. However, in-
tegrating an observer will further increase the computational
load and exacerbates the challenge of reducing the computa- 4) Offers a sensorless estimation and compensation solu-
tional burden. tion for the equivalent DC link impedance.
One of the most promising solutions for MPC challenges is The rest of this article is organized as follows. System
the inverse model predictive control (IMPC) [2]. IMPC starts description is presented in Section II. Black box-based mod-
by predefining the optimal cost function values and computes elling is presented in Section III. The design of the proposed
the switching state that achieves this cost function. How- robust IMPC is discussed in Section IV. Experimental and
ever, IMPC does not require iterative prediction and, thus, simulation results for ac–dc side utilizing a T-type as a case
requires a minimal computational time. The IMPC depends study are shown in Section V. Then, Section VI concludes
on the dynamic model to transform the predefined optimal this article.
cost function into the available switching states vectors similar
to traditional MPC. The disturbances and uncertainties in this II. SYSTEM DESCRIPTION
case lead to wrong transformation and wrong switching vector Fast EV charger consists of two cascaded converters, ac–dc
selection. Inspired from IMPC proposed in paper [2], this converter and dc–dc converter, as shown in Fig. 1. The ac–dc
article proposes a new technique for enhancing the robustness converter is responsible to interface the charging system with
of IMPC utilized for EV charging applications. The proposed the ac grid, while the dc converter aims to regulate the power
method utilizes the RLS as an adaptive estimation technique injected or extracted from the EV. This study focuses solely on
to estimate the parameters of the entire dynamic model. the ac–dc side of the EV charger to avoid paper overlength.
The estimated parameters are then used to update the dy- T-type rectifier has multiple advantages over its counter-
namic model utilized by the IMPC. The proposed technique parts, such as supports bidirectional power flow, supports
can significantly enhance the performance of the EV charger, switched fault tolerant operation, low power losses, reduced
fulfill charging requirements, and solves the problem of the components count, simple control, reduced input filter, and
DC link impedance identification. The contributions of this higher efficiency. These advantages make the T-type converter
article can be summarized as follows. one of the most suitable solutions for EV chargers [23]. The
1) A robust control technique is proposed for the ac–dc poles voltages of this rectifier are regulated through twelve
side of the EV chargers is proposed. power electronic switches distributed among three legs. The
2) Proposed a control method that guarantees optimal re- poles’ voltages have finite set of values, containing three-
sponse of the charger, while adhering to constraints, possibilities for each one, namely positive (P), zero (0), and
uncertainties, and abnormal grid conditions. negative (N) as shown in Table 1. The symbol Sij represents
3) The proposed control technique is suitable for wide the gate of the ith switch in the jth leg, i ∈ [1 2 3 4],
range of charging power and bidirectional power flow. j ∈ [a b c], uj represent the jth leg pole voltage, and Snj

VOLUME 5, 2024 797

SHARIDA ET AL.: ENHANCED INVERSE MODEL PREDICTIVE CONTROL FOR EV CHARGERS: SOLUTION FOR RECTIFIER-SIDE

is the normalized pole voltage (the ratio of pole the voltage to the output voltage. Therefore, (9) can be rewritten as follows:
the dc-link voltage). √
C1 dvdc vdc 3vd id
+ = . (10)
A. AC SIDE DYNAMIC MODELING 2 dt RL vdc
In general, the ac side dynamics of the converter can be de- To generate the explicit dynamic model of the dc link volt-
scribed as follows [16]: age, (10) can be reformulated as follows:
diabc √
uabc = −Lg − Rgiabc + vabc (1) dvdc 2 3vd id 2vdc
dt = − . (11)
dt C1 vdc C1 RL
where vabc = [va vb vc ]T is the vector of instantaneous grid
voltages, Lg and Rg are the inductance and the internal re- Finally, the discretized formula of (11) can be represented
sistance of the input filter, iabc = [ia ib ic ]T is the vector of as follows:
√  
instantaneous grid currents, and uabc = [ua ub uc ]T is the 2 3Ts vd (k) id (k) 2Ts
vector of poles’ voltages. vdc (k + 1) = + 1− vdc (k) .
C1 vdc (k) C1 RL
For simplicity, (1) can be transformed into the αβ frame, (12)
which reduces the number of controlled states to two as fol-
lows: III. BLACK BOX-BASED MODELING
diα A. UNCERTAINTY AND DISTURBANCE MODELING
uα = − Rgiα − Lg + vα (2)
dt In fact, (5) and (12) are exposed to disturbances and un-
diβ certainties. The uncertainties are mainly generated in the ac
uβ = − Rgiβ − Lg + vβ . (3) side from the passive parameters Rg and Lg . On the other
dt
hand, dc side uncertainties are generated by C1 and RL . While
This model can then be discretized using Newton–Euler
the disturbances are mainly generated by losses and unmod-
approximation as follows:
eled dynamics such as coupling effects between the dynamic
dx(t ) x(k + 1) − x(k) states. The disturbances and uncertainties lead to increase the
≈ (4)
dt Ts prediction error proportional to the number of uncertainties
where Ts is the sampling time. and disturbances. This in turn, leads to drift the control signal
Therefore, the discretized model can be expressed as fol- from the optimal one. To face this challenge, it is worth to
lows: define a new dynamic model that include these uncertainties
  and disturbances, while assuming that their values, any other
RgTs Ts
iα (k + 1) = 1 − iα (k) + [vα (k) − uα (k)] term multiplied by these values, or added to these values are
Lg Lg unknown. Therefore, (5) and (12) can be written as follows:
  
RgTs Ts    
iβ (k + 1) = 1 − iβ (k) + vβ (k) − uβ (k) (5) R̂g + Rg Ts
Lg Lg iα (k + 1) = 1 − iα (k) . . .
L̂g + Lg
B. DC SIDE DYNAMIC MODEL Ts
The dc side dynamics can be obtained based on the power + [vα (k) − uα (k)] + cα
L̂g + Lg
balance formula between the ac and dc sides as follows [24]: 
√  
iovdc = 3vd id (6) R̂g + Rg Ts
iβ (k + 1) = 1 − iβ (k) . . .
L̂g + Lg
where vd and id are the d-terms of the grid and current voltage
represented in the synchronous dq-frame, vdc is the dc-link Ts  
+ vβ (k) − uβ (k) + cβ (13)
voltage, and Io is given by the following: L̂g + Lg
√
io = ic1 + idc . (7) 2 3Ts vd (k) id (k)
vdc (k + 1) = + ...
and ic1 is the current through the upper capacitor, and given Ĉ1 + C1 vdc (k)
by the following: 
2Ts
dvc1 1−    vdc (k) + cdc
ic1 = C1 . (8) Ĉ1 + C1 R̂L + RL
dt
(14)
Substitute (7) and (8) into (6)
√ where R̂g and Rg are the expected value of Rg and its expec-
dvc1 vdc 3vd id
C1 + = . (9) tation error, L̂g and Lg are the expected value of Lg and its
dt RL Vdc expectation error, R̂L and RL are the expected value of RL
Assuming that during balanced conditions, the voltages of and its expectation error, Ĉ1 and C1 are the expected value
the upper and lower capacitors are balanced and equal half of of C1 and its expectation error, cα and cβ are the disturbance

798 VOLUME 5, 2024

in α and β axes, respectively, and cdc is the disturbance of the To estimate the unknown parameters, it is worth converting
dc side dynamics. the dynamic model (16) into the standard form of the estima-
An important remark for (14) is that it contains RL and RL . tor model as follows:
However, in fast EV charger applications, the output of the ⎡ ⎤ ⎡ ⎤
rectifier is connected directly to a resistive load with a capac- iα (k + 1) aα bα cα 01×3 01×3
⎣ iβ (k + 1) ⎦ = ⎣01×3 aβ bβ cβ 01×3 ⎦
itive filter. Instead, it is directly connected to another dc–dc
converter. This makes the prediction of the load resistance vdc (k + 1) 01×3 01×3 adc bdc cdc
     
very difficult until an additional dc current sensor is used. y θ
In fact, taking a deep look on the dynamic model, it can be ⎡ ⎤
iα (k)
noticed that all terms are function of the passive parameters,
⎢vα (k) − uα (k)⎥
converting the entire dynamic model into a black-box model. ⎢ ⎥
⎢ 1 ⎥
To simplify the presentation of the dynamic model, assume ⎢ ⎥
⎢ iβ (k) ⎥
that ⎢ ⎥
⎢ v
×⎢ β (k) − uβ (k) ⎥ (17)
⎥
⎢ 1 ⎥
  ⎢ v (k)i (k) ⎥
R̂g + Rg Ts ⎢ d d ⎥
Ts ⎢ vdc (k) ⎥
aα = 1 − , bα = ⎣ v (k) ⎦
L̂g + Lg L̂g + Lg dc
  1
R̂g + Rg Ts Ts   
aβ = 1 − , bβ = ϕ
L̂g + Lg L̂g + Lg
√ where y is the output vector, θ represents the unknown param-
2 3Ts 2Ts
adc = , bdc = 1 −    eters matrix, and ϕ represents the measurements vector.
Ĉ1 + C1 Ĉ1 + C1 R̂L + RL To estimate the unknown matrix, RLS algorithm can be
(15) used as follows:
 −1
K (k) = P (k − 1) ϕ (k) I + ϕ(k)T P (k − 1) ϕ (k) (18)
where all of these parameters are considered unknowns.
After substituting (15) in (13) and (14), the simplified dy- ε (k) = Y (k) − ϕ(k)T θ̂ (k − 1) (19)
namic model can be written as follows:
θ̂ (k) = θ̂ (k − 1) + K (k) ε (k) (20)
 −1
P (k) = I9 − K (k) ϕ(k)T P (k − 1) (21)
iα (k + 1) = aα iα (k) + bα [vα (k) − uα (k)] + cα
  where K(k) is the estimator gain matrix, P(k) is the error
iβ (k + 1) = aβ iβ (k) + bβ vβ (k) − uβ (k) + cβ
covariance matrix, ϕ(k) is the measurements vector, θ̂(k) is
vd (k) id (k) the estimated parameters matrix, ε(k) is the estimation error
vdc (k + 1) = adc + bdc vdc (k) + cdc . (16)
vdc (k) vector at the instant k, I9 is 9 × 9 identity matrix, P(k-1) is the
previous value of the error covariance matrix, and θ̂(k − 1) is
the previous value of the estimated parameters matrix.
If (15) is estimated correctly, (16) can also be modeled From RLS, it can be noticed that two matrices require
accurately even without the need of extra dc current sensors, initialization, θ̂ 0 and P0 . A good approach to initialize θ̂ is
as the available sensors can be utilized to estimate all required to refer to its definition shown in (17), which consists of
parameters. expected, well-known, and unknown parameters. Assuming
that the unknown values are zeros, and substitute the expected
and well known values leads to initialize the unknown matrix
B. BLACK-BOX ESTIMATION with the best possible approximation as follows:
The simplified black-box dynamic model can be separated ⎡ ⎤
R̂ T Ts
into three terms, 1) outputs, 2) measurements, and 3) un- 1 − L̂g s 0 01×3 01×3
⎢ g L̂g ⎥
knowns. The outputs represent the values of Iα (k+1), iβ (k+1) ⎢ R̂ T ⎥
θ̂ 0 = ⎢ 03×1 1 − L̂g s Ts
0 01×3 ⎥.
and vdc (k+1). The measurements include sensor-based mea- ⎣ g √ L̂ g ⎦
2 3Ts 2Ts
surements like iα (k), iβ (k), vα (k), and vβ (k), vd (k), id (k) or 01×3 01×3 Ĉ1
1 − Ĉ1 R̂L
0
pseudo measurements like uα (k) and uβ (k). The poles volt- (22)
ages uα (k) and uβ (k) are considered pseudo measured as they As the estimation error has a peak value at the beginning,
are obtained mathematically based on the applied switching the error covariance matrix must be initialized with a rela-
vector shown in Table 1, where there are no sensors are used tively high value, i.e.
to measure them. The unknown parameters are all parameters
shown in (15). P0 > 10ε0 . (23)

VOLUME 5, 2024 799

SHARIDA ET AL.: ENHANCED INVERSE MODEL PREDICTIVE CONTROL FOR EV CHARGERS: SOLUTION FOR RECTIFIER-SIDE

However, due to the adaptive nature of generating the gain which leads to the compute the optimal pole voltages in the
matrix K(k), θ̂ 0 and P0 can be set to any value different than abc frame (uOabc )
(22) and (23). Moreover, they will be updated automatically −1
by the estimation algorithm after the first estimation cycle. uOabc = kαβ uOαβ (32)
However, this will affect the speed, accuracy, and oscillation where kαβ is the Clarke transformation, and given by the
of the estimation error at the beginning of the algorithm. following:
Therefore, it is recommended to adopt (22) and (23) for the ⎡ ⎤
initialization. 1 − 21 − 21
2⎢ √ √ ⎥
Finally, after estimating θ 0 , the entire dynamic model be- kαβ = ⎣ 0 23 − 23 ⎦ . (33)
comes completely known, and converges to the real model. 3
1 1 1
Therefore, the dynamic model in (16) becomes as follows: 2 2 2

îα (k + 1) = âα iα (k) + b̂α [vα (k) − uα (k)] + ĉα Then, uOabc is be converted to its corresponding normalized
  optimal poles’ state as follows:
îβ (k + 1) = âβ iβ (k) + b̂β vβ (k) − uβ (k) + ĉβ
SOabc = uOabc /vdc . (34)
vd (k) id (k)
v̂dc (k + 1) = âdc + b̂dc vdc (k) + ĉdc . (24) As g = 0 is not guaranteed to be feasible, SOabc may not
vdc (k)
be feasible too. Therefore, SOabc must be discretized to obtain
Finally, (24) will be used by the IMPC to predict the states the nearest feasible switching vector (S̄Oabc )
at the instant k+1 as discussed in the next section.  
SOabc
S̄Oabc = min |Sn | × round . (35)
IV. ADAPTIVE AND ROBUST IMPC DESIGN min |Sn |
A. IMPC CURRENT CONTROLLER
In fact, (35) guarantees that S̄Oabc is always feasible, be-
The general formula of grid current tracking cost function is
longs to Sn , and represents the closest poles’ voltages vector
given by the following:
  to SOabc . Finally, the switching state of each switch can be
   
g = Wα iα∗ − îα (k + 1) + Wβ iβ∗ − îβ (k + 1) . (25) obtained as shown in Algorithm 1 based on the computed
S̄Oabc .
To achieve current tracking objective, the cost function
must approach zero (i.e., g ≈ 0). Therefore B. IMPC VOLTAGE CONTROLLER
   
 Starting from the dc side term of (24), the dc-link voltage
0 ≈ iα∗ − îα (k + 1) + iβ∗ − îβ (k + 1) . (26)
dynamics are given by the following:
This case occurs when vd (k) id (k)
vdc (k + 1) = adc + bdc vdc (k) + cdc . (36)
iα∗ ≈ îα (k + 1) (27) vdc (k)

iβ∗ ≈ îβ (k + 1) . (28) In (36), the grid voltage is constant and cannot be changed
by the controller, id is controllable, and vdc is the target. There-
As the rectifier has additional two objectives, dc-link fore, only id (k) is controllable. However, id (k) has an infinite
voltage regulation and dc side capacitor balancing, the equi- set of values and should not be normalized or discretized as in
librium conditions described in (27) and (28) can be updated the current IMPC. Therefore, the reference current of the grid
as follows: can be computed as a function of the reference dc-link voltage
iα∗ (k + 1) ≈ îα (k + 1) + Wα1 vdc + Wα2 vc (29) based on (36) as follows:
vdc (k)  ∗ 
iβ∗ (k + 1) ≈ îβ (k + 1) + Wβ1 vdc + Wβ2 vc . (30) id∗ (k) = vdc − b̂dc vdc (k) − ĉdc . (37)
âdc vd (k)
Substituting these constraints in (24), leads to compute the
Finally, assuming that iq∗ = 0, iαβ
∗ can be generated by trans-
required optimal poles’ voltages (uOαβ ) required to achieve ∗ ∗
forming idq into iabc using the Park transformation, then from
the desired optimal g, ∗ to i∗ using the Clarke transformation.
iabc αβ
b̂αβ uOαβ (k) = âαβ iαβ (k) + b̂α vαβ (k) + ĉαβ . . . The block diagram of the proposed controller is shown in
∗
+ Wαβ1 vdc + Wαβ2 vc − iαβ (k + 1) . Fig. 2. The PLL function is used to estimate the phase angle
of the utility grid which is then used to generate the reference
(31)
current signal based on the reference dc-link voltage or the
Nevertheless, the computed optimal pole voltages might not reference power. The generated current reference signal is
always be feasible with the chosen converter, as each con- then applied to the IMPC along with dc-link voltage and—
verter has a finite-set of poles voltages. This concern is solved grid currents and voltages in the αβ reference frame. The
in the IMPC by computing the nearest feasible switching state. IMPC utilizes these measurements and compute the optimal
To achieve this, (31) should be converted into the abc frame, switching signals and applies them to the switches.

800 VOLUME 5, 2024

 TABLE 2. System’s Parameters

FIGURE 2. Block diagram of the enhanced IMPC.

Algorithm 1: S̄O j to Sj Transformation.

The computational time analysis was discussed in details

in [2] where the IMPC approach was proposed for the first FIGURE 3. Controller’s steady state and dynamic response.
time. For three-level converters (i.e., T-type), the computa-
tional time was reduced from 87µs in the conventional MPC
to 4.8 μs in the IMPC. As this article proposes robust- matrix (θ̂) and the corresponding estimated model response
ness enhancement for IMPC, computational time comparison using an oscilloscope. Therefore, this section shows the
is not included in this article. However, it is worth noting response of the estimator through various scenarios including
that parameters estimation and compensation, which are the steady state and transient operations. Simulation parameters
main contributions of this article, require around 3 μs to be and weighting factors are shown in Table 2.
achieved. This makes the computational load for the enhanced Fig. 3 shows the response of controller during normal op-
controller around 8 μs which is still much less than the con- eration including steady state and transient response. It can be
ventional MPC (88 μs), at the same time the proposed method noticed that the grid currents are sinusoidal, in phase with the
offers enhanced robustness compared to both MPC and IMPC. grid voltages, have low THD value compared to the standards
[25], and the power factor is unity. The dc-link voltage is
V. RESULTS well-regulated with no oscillations or ripples, converges to
A. SIMULATION RESULTS the reference value smoothly and rapidly. Similarly, the volt-
The proposed control technique and the considered T-type ages of the dc side capacitors are well-regulated and balanced
rectifier are implemented through MATLAB/ Simulink to val- through all steady state and transient periods.
idate the design of the model parameters estimator. MATLAB/ Fig. 4 show the estimation response of the RLS. In this
Simulink environment is adopted to validate this functionality section, only Iα and the dc-link voltage are presented as Iβ
as it is difficult to show the contents of the estimated unknown exhibits a similar response to Iα . The estimation process of

VOLUME 5, 2024 801

SHARIDA ET AL.: ENHANCED INVERSE MODEL PREDICTIVE CONTROL FOR EV CHARGERS: SOLUTION FOR RECTIFIER-SIDE

FIGURE 5. Experimental setup.

FIGURE 4. Estimator’s response during steady state and transient.

the grid current model is achieved accurately and smoothly,

the estimation error approaches zero during the full simulation
interval as both initialization and actual values start from zero.
However, at the beginning of operation, it can be noticed that
the estimation error of the dc-link voltage is very high. This
happens due to the fact that the controlled rectifier operates
as uncontrolled rectifier before starting the control loop. This
causes to charge the dc side capacitors while the initial value
of estimated model is set to zero. However, the estimated
model converges to the actual one rapidly at the next cycle of
operation. This ensures that the estimator can always converge
to the actual value regardless the set initial values. However,
better initialization helps the estimator to perform smoothly at
the beginning of operation. Fig. 4 also shows the behavior of
the estimated model’s parameters where the parameters seem
to be almost stable during steady state, while cα has some
oscillations as it represents the losses and unmodeled dynam-
ics. At the instant of transient, the control signal changes
sharply, while the grid current changes smoothly. This tran-
sient nonlinear behavior between the control signal and the
corresponding grid currents represents the major challenge in
the controller design. As the dynamic model during transient
becomes different than the model during steady state. This
deteriorates the response temporarily until the transient is
finished. However, as the proposed control approach estimates
the model accurately online, the controller is able to accu-
FIGURE 6. Steady-state response under normal operation. (a) Rectifier
rately predict the behavior of the physical system regardless operation. (b) Inverter operation.
in which period the converter is working in.

B. EXPERIMENTAL RESULTS and a Chroma 63084 programmable electrical load. The pro-
The experimental investigations are conducted through the posed control approach is implemented through the low-cost
configuration illustrated in Fig. 5 to assess the effectiveness microcontroller STM32H745.
of the proposed control technique. The setup comprises a Fig. 6 shows the steady-state response in bidirectional
California Instrument (MX30) regenerative grid emulator, an power flow. As a rectifier, the ac–dc converter is responsible
L filter linking the grid emulator to the rectifier, a laboratory- to regulate the dc-link voltage, and generates the required
scale T-type rectifier, a current–voltage measurement board, reference current signal based on the dc-link voltage error.

802 VOLUME 5, 2024

FIGURE 7. DC-link voltage transient response under normal operation. FIGURE 9. Steady-state response under distorted grid voltage operation.

FIGURE 8. Steady-state response under unbalanced grid voltages FIGURE 10. Dynamic response under 50% passive parameters uncertainty.
operation.

during severe distortion of the grid voltages. In this exper-

However, when it works as an inverter, the grid currents will iment, a 5% 5th order and 5% 7th order harmonics are
be regulated based on a constant reference current or power, intentionally added to the grid voltages. However, the grid
while the regulation of the dc-link voltage is assigned to the currents are well regulated, sinusoidal, and the power factor
dc–dc converter. Fig. 6(a) shows that the current is in phase remains unity. Similarly, the dc-link voltage is well-regulated
with the voltage (positive current), while in Fig. 6(b) the cur- and the dc side capacitors voltages are balanced.
rent is shifted 180° from the voltage which mean the direction In the next experiment, the passive parameters are changed
of the current is changed (negative current). physically by 50% without updating their values in the control
Fig. 7 shows the transient analysis for dc-link voltage under algorithm. The results of this experiment are illustrated in
normal operating conditions These experiments show that the Fig. 10, showing that the response is almost not affected. The
grid currents are sinusoidal, in phase with grid voltages, and currents are still well regulated, the power factor is almost
the power factor is almost unity. Moreover, the dc-link voltage unity, the dc-link voltage is well-regulated, the dc transient
is almost constant with no ripples or oscillations, the steady- response is smooth, and the capacitors voltages are completely
state error is almost zero, and the transient process is very balanced.
smooth. In addition, the voltages of the dc side capacitors are Grid voltage sag and swell are among the most significant
balanced even during voltage transient. abnormal grid conditions. Fig. 11 shows the response of the
Fig. 8 shows the response of the controller when the grid proposed controller under 10% voltage sag and swell. It can
voltages are 10% unbalanced, where Va = 100Vrms , Vb = be noticed that in both cases, the dc-link voltage is well-
110Vrms , and Vc = 120Vrms . However, the grid currents are regulated, with almost no ripples, with grid currents slightly
almost balanced thanks to the adaptive model identification, increased in case of voltage sag to compensate the voltage
which estimates the relation between grid currents and volt- drop, and slightly decreased during voltage swell as expected.
ages for each phase independently. Fig. 9 shows the response Moreover, the currents are sinusoidal, the power factor is

VOLUME 5, 2024 803

SHARIDA ET AL.: ENHANCED INVERSE MODEL PREDICTIVE CONTROL FOR EV CHARGERS: SOLUTION FOR RECTIFIER-SIDE

FIGURE 11. Controller response during (a) voltage sag and (b) voltage
swell.
FIGURE 12. IMPC and Robust IMPC response comparison for ±100%
passive parameters variation.

unity, and the THD of the grid currents is much less than the
international standard limits. TABLE 3. Numerical Results Comparison
These experiments ensure the robustness of the proposed
control technique under severe parameters uncertainty, and
abnormal grid conditions.

C. COMPARISON ANALYSIS
To show the effectiveness of the proposed method, it is
compared with the IMPC published in [2]. The compari-
son is conducted under high-uncertainty and high-parameters
variation conditions, especially at the dc-link side (Ĉ1 + changing them in the control algorithm. The inductors values
C1 , R̂L + RL ). The selection of such conditions for com- and capacitor values were first changed, then at t = 0.2, the
parison is considered practical and reasonable as the highest load was doubled. Fig. 12 shows a comparison between the
uncertainty in EV charging application occurs at the dc-link response of IMPC and the enhanced IMPC under these condi-
side due the wide range of charging power. As an EV charger tions. Both of the controllers were able to restore the dc-link
usually supports 30–100 kW, the virtual load can be severely voltage accurately and smoothly, with almost the same voltage
changed, resulted on a high range of load uncertainty. To this drop at the instant of load change (around 20 V).
end, a high uncertainty was added to the passive parameters However, in the case of IMPC, the ripples of the dc-link
where each parameter was increased by ±100%. For exam- voltage are slightly increased to ±4 V, but it is still acceptable.
ple, 100% change in the dc-link impedance, means that the The major effects of the uncertainty appear clearly in the grid
charging power is doubled. Therefore, the passive parame- currents, the currents become unbalanced by 10.9%. However,
ters where suddenly increased online, which is achieved by the response of the enhanced IMPC is almost not affected,
changing the actual values of the passive components without with the ripples of the dc-link voltage are almost zero, and

804 VOLUME 5, 2024

 Overall, the proposed technique represents a significant
step forward in advancing the response of bidirectional EV
chargers. It is promising in addressing EV charger challenges
in the context of grid and renewable energy integration and
the associated challenges such as uncertainties, fluctuation,
the intermittent nature, and the abnormal grid conditions.

ACKNOWLEDGMENT
The statements made herein are solely the responsibility of the
authors.

REFERENCES
[1] F. Blaabjerg, H. Wang, I. Vernica, B. Liu, and P. Davari, “Reliability
of power electronic systems for EV/HEV applications,” Proc. IEEE,
vol. 109, no. 6, pp. 1060–1076, Jun. 2021.
[2] A. Sharida, S. Bayhan, H. Abu-Rub, and U. Fesli, “Inverse model
predictive control for power converters,” IEEE Access, vol. 11,
pp. 68485–68496, 2024.
[3] K. Premkumar, P. Kandasamy, M. V. Priya, T. Thamizhselvan, and S.
R. Carter, “Three-phase rectifier control techniques: A comprehensive
literature survey,” Int. J. Sci. Technol. Res., vol. 9, no. 1, pp. 3183–3188,
2020.
[4] D. Mishra, B. Singh, and B. K. Panigrahi, “Adaptive current control
for a bidirectional interleaved EV charger with disturbance rejection,”
IEEE Trans. Ind. Appl., vol. 57, no. 4, pp. 4080–4090, Jul./Aug. 2021.
[5] X. Wang et al., “Control and modulation of a single-phase AC/DC
converter with smooth bidirectional mode switching and symmetri-
cal decoupling voltage compensation,” IEEE Trans. Power Electron.,
vol. 37, no. 4, pp. 3836–3853, Apr. 2022.
[6] S. Chaurasiya and B. Singh, “A bidirectional fast EV charger for wide
voltage range using three-level DAB based on current and voltage
FIGURE 13. IMPC and Robust IMPC THD comparison for ±100% passive
stress optimization,” IEEE Trans. Transp. Electrific., vol. 9, no. 1,
parameters variation.
pp. 1330–1340, Mar. 2023.
[7] S. Yan, Y. Yang, S. Hui, and F. Blaabjerg, “A review on direct power
control of pulsewidth modulation converters,” IEEE Trans. Power Elec-
tron., vol. 36, no. 10, pp. 11984–12007, Oct. 2021.
[8] C. Ren, L. Liu, X. Han, B. Zhang, L. Wang, and P. Wang, “Multi-mode
the grid currents are balanced. Moreover, Fig. 13 shows the control for three-phase bidirectional AC/DC converter in hybrid micro-
effect of parameters variation on the THD of the grid currents. grid under unbalanced AC voltage conditions,” in Proc. IEEE Energy
In case of IMPC, the THD is increased to 3.21%, while it is Convers. Congr. Expo., 2019, pp. 2658–2663.
[9] F. Tlili, A. Kadri, and F. Bacha, “Advanced control strategy for bidirec-
0.25% for the enhanced one. Although the THD of the IMPC tional three phase AC/DC converter,” Elect. Power Syst. Res., vol. 179,
is less than 5%, it is worth noting that the even components 2020, Art. no. 106078.
(0th and 2th) are significantly increased due the significant [10] A. M. Mohammed, S. N. H. Alalwan, A. Taşcıkaraoğlu, and J. P.
Catalão, “Sliding mode-based control of an electric vehicle fast charg-
increment and unbalance between the positive, negative, and ing station in a DC microgrid,” Sustain. Energy, Grids Netw., vol. 32,
zero sequences of the currents. Finally, a detailed numerical 2022, Art. no. 100820.
results comparison is presented in Table 3. [11] K. Shipra, R. Maurya, and S. N. Sharma, “Brayton-moser passivity
based controller for electric vehicle battery charger,” CPSS Trans.
Power Electron. Appl., vol. 6, no. 1, pp. 40–51, 2021.
[12] H. Makhamreh, M. Trabelsi, O. Kükrer, and H. Abu-Rub, “A lyapunov-
VI. CONCLUSION based model predictive control design with reduced sensors for a puc7
This article proposed a new robust control technique for fast rectifier,” IEEE Trans. Ind. Electron., vol. 68, no. 2, pp. 1139–1147,
EV chargers inspired by IMPC. The proposed controller, com- Feb. 2021.
[13] B. Chelladurai, C. K. Sundarabalan, S. N. Santhanam, and J. M. Guer-
prising two layers, the first layer effectively estimates model rero, “Interval type-2 fuzzy logic controlled shunt converter coupled
parameters, enabling compensation for various dynamic fac- novel high-quality charging scheme for electric vehicles,” IEEE Trans.
tors encountered in real-world scenarios. Meanwhile, the Ind. Informat., vol. 17, no. 9, pp. 6084–6093, Sep. 2021.
[14] G. Wang, X. Wang, and J. Lv, “An improved harmonic suppression con-
second layer, utilized the IMPC mechanism for regulating trol strategy for the hybrid microgrid bidirectional AC/DC converter,”
both dc-link voltage and grid currents. Through rigorous test- IEEE Access, vol. 8, pp. 220422–220436, 2020.
ing and analysis, the proposed controller exhibits adaptive, [15] A. Sharida, S. Bayhan, and H. Abu-Rub, “Enhancing scalability of fast
electric vehicle charging stations: Solutions for AC-DC side integration
robust, fast, accurate, and smooth response characteristics. and regulation,” IEEE Open J. Ind. Electron. Soc., vol. 4, pp. 720–731,
Notably, the proposed control technique showcased excep- 2023.
tional performance in mitigating severe uncertainties and [16] A. Sharida, S. Bayhan, and H. Abu-Rub, “Adaptive control strategy
for three-phase three-level T-type rectifier based on online disturbance
disturbances, including sudden load changes, unbalanced grid estimation and compensation,” IEEE Access, vol. 11, pp. 40967–40977,
voltages, and distorted input voltage. 2023.

VOLUME 5, 2024 805

SHARIDA ET AL.: ENHANCED INVERSE MODEL PREDICTIVE CONTROL FOR EV CHARGERS: SOLUTION FOR RECTIFIER-SIDE

[17] H. Wang and H. Zhang, “Study on an improve finite-control-set-model SERTAC BAYHAN (Senior Member, IEEE) re-
predictive control (FCS-MPC) strategy for a T-type rectifier with direct ceived the M.S. and Ph.D. degrees in electrical en-
power control strategy,” IEEJ Trans. Elect. Electron. Eng., vol. 18, no. 3, gineering from Gazi University, Ankara, Türkiye,
pp. 442–450, 2023. in 2008 and 2012, respectively.
[18] L. Wang, A. Dubey, A. H. Gebremedhin, A. K. Srivastava, and N. His research interests include power electronics
Schulz, “MPC-based decentralized voltage control in power distribu- and their applications in next-generation power and
tion systems with EV and PV coordination,” IEEE Trans. Smart Grid, energy systems, including renewable energy inte-
vol. 13, no. 4, pp. 2908–2919, Jul. 2022. gration, electrified transportation, and demand-side
[19] I. Harbi, M. Abdelrahem, M. Ahmed, and R. Kennel, “Reduced- management. He has acquired $13M in research
complexity model predictive control with online parameter assessment funding and published more than 170 papers in
for a grid-connected single-phase multilevel inverter,” Sustainability, mostly prestigious IEEE journals and conferences.
vol. 12, no. 19, 2020, Art. no. 7997. He has coauthored three books and six book chapters.
[20] I. González-Prieto, M. J. Duran, A. Gonzalez-Prieto, and J. J. Aciego, Dr. Bayhan currently serves as an Associate Editor for Transactions on
“A simple multi-step solution for model predictive control in mul- Industrial Electronics, IEEE JOURNAL OF EMERGING AND SELECTED TOPICS
tiphase electric drives,” IEEE Trans. Ind. Electron., vol. 71, no. 2, IN INDUSTRIAL ELECTRONICS, IEEE OPEN JOURNAL OF THE INDUSTRIAL
pp. 1158–1169, Feb. 2024. ELECTRONICS SOCIETY, and IEEE INDUSTRIAL ELECTRONICS TECHNOLOGY
[21] Y. Xie et al., “A simplified algorithm of finite set model predictive NEWS, and Guest Editor for IEEE TRANSACTIONS ON INDUSTRIAL INFORMAT-
control for three-phase CHB-based BESSs,” IEEE J. Emerg. Sel. Topics ICS.
Power Electron., vol. 12, no. 2, pp. 1203–1214, Apr. 2024.
[22] D.-J. Kim, B. Kim, C. Yoon, N.-D. Nguyen, and Y. I. Lee, “Disturbance
observer-based model predictive voltage control for electric-vehicle HAITHAM ABU-RUB (Fellow, IEEE) received
charging station in distribution networks,” IEEE Trans. Smart Grid, the M.Sc. degree from Gdynia Maritime Academy,
vol. 14, no. 1, pp. 545–558, Jan. 2023. Gdynia, Poland, 1990, and the Ph.D. degree from
[23] A. Sharida, S. Bayhan, and H. Abu-Rub, “Voltage-sensorless open- Technical University of Gdansk, Poland, in 1995,
switch fault tolerant control of three-phase T-type rectifier,” IEEE Trans. both in electrical engineering, and another Ph.D.
Power Electron., vol. 38, no. 12, pp. 15365–15376, Dec. 2023. degree in humanities from Gdansk University,
[24] S. Bayhan and H. Komurcugil, “Sliding-mode control strategy for three- Gdansk, Poland, in 2004.
phase three-level T-type rectifiers with DC capacitor voltage balancing,” Since 2006, he has been with Texas A&M Uni-
IEEE Access, vol. 8, pp. 64555–64564, 2020. versity at Qatar, Doha, Qatar, where he is currently
[25] IEEE Recommended Practices and Requirements for Harmonic Control a Professor and Managing Director of the Smart
in Electric Power Systems, IEEE Standard 519-2014, 2014. Grid Center Extension. He has co-authored more
than 550 journal and conference papers, six books, and six book chapters.
ALI SHARIDA (Student Member, IEEE) received His main research interests include energy conversion systems, smart grid,
the B.E. degree in mechatronics engineering from renewable energy systems, electric drives, and power electronic converters.
Palestine Technical University (PTUK), Tulkarm, Dr. Abu-Rub was the recipient of many prestigious national and interna-
Palestine, in 2013, and the M.Sc. degree in mecha- tional awards and recognitions, such as the American Fulbright Scholarship
tronics engineering from Palestine Polytechnic and the German Alexander von Humboldt Fellowship. He is the Co-Editor in
University, Hebron, Palestine, in 2020. He is cur- Chief of IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS.
rently working toward the Ph.D. degree in electri-
cal engineering - fast electric vehicle chargers from
Texas A&M University, College Station, TX, USA.
In 2014 he joined PTUK as a T.A. In 2022,
worked with Texas A&M University at Qatar as
an Associate Research Assistant. His current research interests include in-
telligent systems, systems identification, power converters, power converters
control, and adaptive control.

ABDULLAH BERKAY BAYINDIR (Student Mem-

ber, IEEE) received the B.Sc. degree in electrical
and electronics engineering from TOBB Economy
and Technology University, Ankara, Türkiye, in
2023. He is currently working toward the Ph.D.
degree in grid ancillary services with Texas A&M
University, College Station, TX, USA.
His research interests include power converters,
advanced control , renewable energy, and embed-
ded software programming.

806 VOLUME 5, 2024