The 6th IEEE International Conference on Predictive Control of Electrical Drives and Power Electronics

Current-Sensorless Model Predictive Control of

Dual Active Bridge Converters with Kalman Filter
2021 IEEE International Conference on Predictive Control of Electrical Drives and Power Electronics (PRECEDE) | 978-1-6654-2557-5/21/$31.00 ©2021 IEEE | DOI: 10.1109/PRECEDE51386.2021.9681004

Minghao Han1 , Han He1 , Xingtao Wang2 , Zheng Dong1 , Zhenbin Zhang1?

1: School of Electrical Engineering, Shandong University, Jinan 250061, China, [email protected]

2: Shandong Labor Vocational and Technical College (Shandong Labor Technician College), Jinan 250300, China

Abstract—Dual active bridge (DAB) converter has become the electrical systems applications, the primary goal of control
critical equipment in fuel cell vehicle (FCV) electrical systems. strategy is to regulate the output voltage to the reference value
For FCV electrical systems, the DAB converter needs to obtain with fast dynamics [5].
constant output voltage rapidly and accurately under all working
conditions. Model predictive control (MPC), well-known for its Model predictive control (MPC) is a good choice for power
fast dynamics, is a particularly promising strategy to achieve this. converter because of its fast dynamics and multi-objective
However, MPC relies on accurate system model, parameters mis- control capability [6, 7]. Within MPC schemes, finite-control-
match will affect the control performance considerably. Besides, set model predictive control (FCS-MPC) is simple and flexible
additional current sensor is also introduced when using MPC as it leaves out the separate modulator. However, when applied
method, which increases the hardware cost. To tackle this issue,
in this work, we propose a current-sensorless MPC method based to DAB converters, it ignores the advantages of phase-shifting
on Kalman filter. The current estimated by the Kalman filter is modulation strategies and brings large computational burden to
used instead of the sampling one. Compared with traditional the controller. Continuous-control-set model predictive control
MPC, the proposed method depends little on the accurate model (CCS-MPC) can be easily applied to power converters, pro-
parameters. Simulation data verify that, the proposed control viding fixed switching frequency according to the modulation
strategy (i) is able to reduce the current sensor, (ii) has strong
robustness, (iii) does not sacrifice control dynamic performance. strategy adopted. Therefore, CCS-MPC has received more
attention in the field of DAB converter [8, 9].
Index Terms—Dual active bridge converter, predictive control, Generally, in order to enhance the dynamic performance,
Kalman filter, fuel cell vehicle. CCS-MPC introduces an output current sensor to obtain more
system state information. The additional sensors increase the
I. I NTRODUCTION system burden and the hardware cost [10]. Besides, these
In recent years, the growing public attention to global cli- physical sensors are easily subject to noise, requiring addi-
mate change and local environmental pollution pushes govern- tional calibration and filtering processes. According to [11],
ments to investigate an alternative to fossil fuels [1]. Fuel cell the sensors have limited lifespan and more than 40% of system
vehicle (FCV) utilize hydrogen to generate electrical power, failures are related to their malfunction which deteriorates the
has been considering as a potential candidate for reducing car- system reliability.
bon emissions. Besides, FCV also has such advantages as fast- In addition, the mismatch between the actual value and the
refueling time, high safety, and long cruising range. However, one used in the prediction equation is inevitable in practical in-
there are many technical challenges interfacing fuel cells to dustrial applications. However, MPC strategy based on discrete
the motor drive system. For example, the output voltage of system model is sensitive to the parameters. This mismatch
the fuel cell is usually in the range of 60-120V, much lower of system parameters such as inductor value and transformer
than the DC bus voltage, which is typically more than 600V turns ratio will deteriorate the control precision and eventually
[2]. Furthermore, the fuel cell has soft output characteristics: causes system instability.
its output voltage decreases greatly with the increase of power Therefore, sensorless control has been gradually promoted
[3]. To this end, a DC-DC converter with wide voltage gain in DAB converters, which can reduce cost and enhance system
range is required to solve voltages mismatch. reliability. A sensorless control method is proposed in [12],
Dual active bridge (DAB) converters have become a pre- which estimates the load current based on the nonlinear dis-
ferred topology for FCV electrical system, due to the ad- turbance observer and utilizes the idea of feedforward control
vantages of symmetric and modular topology structure, high to improve the dynamic performance. In [13], an energy
power density, and galvanic isolation capability [4]. In FCV feedforward scheme with no extra current sensor is proposed,
considering the change of inductor energy. In [14], a sensor-
The corresponding author of this work is Dr.-Ing. Zhenbin Zhang (e-mail:
[email protected]). This work is financially supported by Shandong Natural less control method for input-independent output-parallel dual
Science Foundation (ZR2019QEE001), Shandong Provincial Key Research active bridge converters is proposed. In addition, the current
and Development Program (Major Scientific and Technological Innovation sharing control of each power module is also realized in this
Project NO.2019JZZY020805), General Program of National Natural Science
Foundation of China (51977124) and Natural Science Foundation of Jiangsu method. Nevertheless, some of the above sensorless control
Province (BK20190204). methods rely on integrators, reducing the dynamic response.

978-1-6654-2557-5/21/$31.00 ©2021 IEEE 663

Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on May 07,2024 at 21:55:35 UTC from IEEE Xplore. Restrictions apply.
Other methods highly depend on the accuracy of the system
model with poor robustness.
To this end, we propose a current-sensorless model predic-
tive control of DAB converters with Kalman filter. The pro-
posed method estimates the load current by utilizing the output
voltage information and depends little on the accurate model
parameters. In addition, we investigate the control performance
comparisons under the load variations and model mismatches
between traditional MPC and the proposed method. Simulation
data verify that the proposed method enhances the system
reliability and not sacrifices fast dynamic response with no
current sensor.

II. S YSTEM D ESCRIPTION AND M ODELING OF DAB Fig. 2: Operation waveforms in single phase shift modulation.
The configuration of the DAB converter is depicted in
Fig. 1. The input side is connected with a dc voltage source and converter, respectively. The output y = Cx(t) denotes the
capacitor C1 . At the output side, the filter capacitor C2 is in output voltage. The system matrices are
parallel with the load resistor Ro . Two H-bridges are arranged
nSb
 
S 
on both sides of the high-frequency isolated transformer with −
 0
a
the turns ratio n. The inductance L consists of transformer L 
 , B =  L  , C = 0 1 . (3)
 
A=  nSb
leakage inductance and the auxiliary inductor. 1 
− 0
C2 Ro C2
The variable Sa and Sb represent the switches position which
are defined as
(
1, S1 and S3 are on, S2 and S4 are off
Sa =
−1, S2 and S4 are on, S1 and S3 are off
( (4)
1, S5 and S7 are on, S6 and S8 are off
Sb =
−1, S6 and S8 are on, S5 and S7 are off.
Fig. 1: Topology of dual active bridge converter. According to (1)–(4), current value of iL under steady-state
condition can be derived as
The single phase shift (SPS) modulation with one degree 
Vi + nVo
of freedom is widely used in DAB converters [15]. Assuming

 (t − t0 ) + I0 , t0 < t < t1
L



the voltage gain k = Vi /(nVo ) > 1 and forward power flow 
V − nVo
 i

(t − t1 ) + I1 , t1 < t < t2


P > 0, the theoretical waveforms of SPS modulation at steady L
state are shown in Fig. 2. We use D to denote the phase shift iL (t) = (5)
−Vi − nVo
(t − t ) + I , t < t < t


ratio, and Ts to represent the switching cycle.  2 2 2 3


 L
In SPS modulation, the duty cycles of all the switching  −Vi + nVo (t − t3 ) + I3 , t3 < t < t4 .



signals are 50%. In each phase leg of both H-bridges, the upper L
and lower switches are controlled with complementary signals Due to the symmetry of the inductor current, I0 = −I2 , I1 =
and the terminal voltages u1 , u2 are symmetrical square waves. −I3 , the inductor current value at the switching instants are
The phase-shift ratio D of the two H-bridges is controlled to
T

adjust the power flow direction and magnitude. I0 = −I2 = − s [Vi + nVo (2D − 1)]

4L (6)
The state-space model of DAB converter under SPS modu-
lation can be obtained as I1 = −I3 = Ts [Vi (2D − 1) + nVo ].

 4L
 dx(t) Assuming the output voltage uo is constant with the value Vo ,
= Ax(t) + Bui (t)
dt (1) the power transmission under steady-state condition is
y(t) = Cx(t),
 Z t2
2 nVi Vo Ts
P = u1 (t)iL (t)dt = D(1 − D). (7)
where Ts t0 2L
>
It should be noted that the expression of transmission power

x(t) = iL (t) uo (t) , (2)
is suitable for both P > 0 and P < 0. For forward power
is the state vector, iL (t) represents the inductor current, uo (t) transmission, the value of the phase shift ratio D ranges from
and ui (t) represent output voltage, input voltage of the DAB 0 to 0.5 due to the current stress restrictions.

664
Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on May 07,2024 at 21:55:35 UTC from IEEE Xplore. Restrictions apply.
 III. C ONTROL S TRATEGY B. Current-sensorless method based on Kalman filter
A. CCS-MPC for DAB converter As can be seen in (10), output voltage prediction requires
output current and the accurate inductor value in FCV applica-
For employing CCS-MPC, it is of fundamental importance tions, due to the time-varying load. Thus, an extra estimation
that the derivation of the prediction model in discrete format method should be added to achieve current sensorless control
of the DAB converter. The output current of H-bridge on the and eliminate the error due to the inaccurate model.
secondary side i2 under steady-state condition can be obtained To this end, we design a Kalman filter in this section which
by (7) estimates the state variables at the sampling instant, achieves
P nVi Ts current-sensorless CCS-MPC scheme for DAB converter.
i2 = = D(1 − D). (8)
Vo 2L It should be noted that, the state-space model (1) can hardly
Applying Kirchhoff current law, we establish the differential be directly utilized, due to its multi-operating modes and
equation of the output side different switch times. For this, we reformulate the discrete-
time model based on (10), which is essentially a reduced-order
duo state space average model. This process involves the following
C2 = ic = i2 − io . (9)
dt two steps.
Thus, the discrete-time output voltage expression of DAB First, the output current of DAB converter io is introduced
converter can be derived by utilizing the first order forward into the model as another state variable, while the state variable
Euler approximation approach iL is discarded because it averages zero over sampling periods.
The state variable uo , together with io form the augmented
nVi Ts2 Ts io (k) state vector
uo (k + 1) = uo (k) + D(1 − D) − . (10)  >
2LC2 C2 xa = uo io . (15)
The main control objective of DAB converter is to ensure its Second, the linear stochastic discrete-time state equations of
output voltage track the reference value rapidly and accurately. the augmented model can be given as
Therefore, the cost function is built to minimize the output
voltage error between the actual and reference value, which is xa (k) = Aa xa (k − 1) + Ba ua (k − 1) + w(k − 1), (16)
chosen as the measured state vector is obtained by
J = (u∗o − uo (k + 1))2 . (11)
za (k) = Ha xa (k) + v(k), (17)
Note that D, which is continuous in nature, ranges from 0
to 0.5 due to the power flow and the current stress restrictions. where
Therefore, the optimization problem under CCS-MPC method ua (k − 1) = Vi D(k − 1)(1 − D(k − 1)), (18)
cannot use enumeration, makes it a challenging task. The
common solution of constrained optimization problem is to is the reformulated input vector, which reflect the effect of
use Lagrange multipliers, and the phase shift ratio D will be phase shift ratio D on the output voltage uo and the coefficient
calculated as matrices are
 
D∗ (k) = arg min J Ts  nTs2

(12) 1 −
C2  , Ba = 
 
s.t. 0 ≤ D∗ (k) ≤ 0.5. Aa =   2LC2  , Ha = 1 0 . (19)

0 1 0
where D∗ (k) is the optimal phase shift ratio. Obviously, when
the output voltage considerably deviates from its reference, The random variables w = [w1 , w2 ]> ∈ R2 and v ∈ R
in order to quickly track the reference, the controller will represent the process and measurement noise, respectively.
change the phase shift ratio into the maximum or minimum Both are assumed to be independent of each other and with
value(D∗ (k) = 0 or D∗ (k) = 0.5). Once the output voltage normal probability distributions. The covariances matrices are
reaches or approximates its reference, the phase shift ratio will semi- and positive definite and given by the constant values
change into the level that corresponds to the steady state power Q = E[ww> ] and R = E[vv > ] for simplicity.
balance, which can be derived as The complete recursive estimation algorithm of the Kalman
s filter is presented by Algorithm 1. Accordingly, the equation
1 1 2L [Ts io (k) + C2 u∗o − C2 uo (k)] for the estimated state x̂a can be derived as
D∗ (k) = − −
2 4 nTs2 ui (13) (
x̂a (k) = x̂− −
a (k) + K(k)[za (k) − H x̂a (k)]
(0 ≤ D∗ (k) ≤ 0.5). (20)
x̂−
a (k) = Aa x̂a (k − 1) + Ba ua (k − 1),
Eq. (13) has a real solution if
where I is the two-dimension identity matrix, K(k) is Kalman
2L [Ts io (k) + C2 u∗o − C2 uo (k)] 1 gain, x̂−
a is the priori state estimate (as defined in Algorithm 1).
≤ . (14)
nTs2 ui 4

665
Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on May 07,2024 at 21:55:35 UTC from IEEE Xplore. Restrictions apply.
Algorithm 1 Recursive estimation algorithm of Kalman filter TABLE I: System Parameters.
1: if k = 0 then Parameters Values
2: State Initialization
Transformer Turn Ratio n = 0.6
x̂a (k) = [0, 0]> , P (k) = I Transformer Leakage Inductance L = 50µH
3: end if Output-side Capacitor C2 = 500µF
4: if k ≥ 1 then Resistive Load Ro = 30Ω
5: Time Update (Predict) Sampling Time Ts = 50µs
6: (a) Predict the state Input Voltage ui = 60V
x̂− Output Voltage Reference u∗o = 100V
a (k) = Aa x̂a (k − 1) + Ba ua (k − 1)
Covariance Matrix of Process Noise Q = diag{0.001, 1}
7: (b) Predict the error covariance matrix
Covariance Matrix of Measurement Noise R = 20
P − (k) = Aa P (k − 1)A> a +Q
8: Measurement Update (Correct)
9: (a) Compute the Kalman gain
Besides, a step-down change in the load from 40Ω to 30Ω
K(k) = P − (k)H > [HP − (k)H > + R]−1
occurs at 0.05s. It can be seen that the proposed method
10: (b) Update the estimation with measurement
has fast dynamic performance as the output voltage remains
x̂a (k) = x̂− −
a (k) + K(k)[za (k) − H x̂a (k)]
practically unaffected in respond to the change of input voltage
11: (c) Update the error covariance matrix
and load.
P (k) = [I − K(k)H]P − (k)
12: end if
13: k = k + 1; Output x̂a (k)

For Kalman filter, it is important to choose the initial error

covariance matrix P (0) and noise covariance matrices Q and
R. The matrix P (0) only influences the initial estimation at
startup process, while no impact on transient and steady-state.
The matrices Q and R are highly correlated to the credibility
with the system model and the sampling value, respectively.
In current-sensorless DAB converter, the state uo has high
confidence as it is obtained by actual physical model, while
the state io is of little confidence due to the augmented model.
Finally, we introduce the estimated states îo and ûo to the
controller as inputs, instead of the measured value io and uo . Fig. 4: [Simulation results:] overall performance of proposed
The block diagram of the entire control scheme is illustrated method, from up to down are output voltage, inductor current.
in Fig. 3.
Fig. 5 shows the estimation performance of the Kalman
filter. The load of DAB converter changes from 40Ω to 30Ω
Voltage Cost Function at 0.01s. According to Fig. 5, Kalman filter tracks the load
Prediction Minimization current fast and accurately without a current sensor. The
steady-state error between actual and estimated values closes
Control to zero. During transient, the converter reaches the new steady-
Kalman state operating point in about 2ms without overshoot.
Variable
Filter
Generation

Fig. 3: Block diagram of continuous-control-set model predic-

tive control scheme with Kalman filter.

IV. S IMULATION VALIDATION

In this section, a simulation model of the proposed current-
sensorless CCS-MPC with Kalman filter is established to
verify the performances. System parameters are collected in
Table. I. Fig. 5: [Simulation results:] estimation performance of Kalman
Fig. 4 shows the simulation results of the overall perfor- filter under step-changes in load.
mance with the proposed scheme. In FCV applications, the
controller is given a constant voltage reference as u∗o = 100 V. To examine the robustness of the proposed method, we
The input voltage is dropped from 60V to 55V at 0.025s. add white noise on the voltage sensor, which mean is 0V,

666
Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on May 07,2024 at 21:55:35 UTC from IEEE Xplore. Restrictions apply.
standard deviation is 0.5V. Fig. 6 depicts the steady-state In this work, we propose a current-sensorless MPC method
control performance of the output voltage between classic for DAB converters, which estimates the output current sen-
and the proposed schemes in presence of sensor noise. The sor utilizing Kalman filter. The proposed method eliminates
proposed scheme traces the voltage reference well and its current sensor and the error due to the inaccurate model.
ripple is lower compared with the classic scheme. Simulation results show that the proposed method achieves
fast and accurate control performance. Finally, a comparison
Classic scheme
between the proposed method and classic MPC is investigated,
confirming that the proposed method has strong robustness.

R EFERENCES
Proposed scheme
[1] C. C. Chan, A. Bouscayrol, and K. Chen, “Electric, hybrid, and fuel-cell
vehicles: Architectures and modeling,” IEEE Transactions on Vehicular
Technology, vol. 59, no. 2, pp. 589–598, 2010.
[2] T. Hong, Z. Geng, K. Qi, X. Zhao, J. Ambrosio, and D. GU, “A Wide
Range Unidirectional Isolated DC-DC Converter for Fuel Cell Electric
Vehicles,” IEEE Transactions on Industrial Electronics, vol. 68, no. 7,
pp. 1–1, 2020.
[3] J. B. Benziger, M. B. Satterfield, W. H. Hogarth, J. P. Nehlsen, and I. G.
Kevrekidis, “The power performance curve for engineering analysis of
Fig. 6: [Simulation results:] comparison of output voltage fuel cells,” Journal of Power Sources, vol. 155, no. 2, pp. 272–285,
2006.
control performance between classic and proposed schemes [4] W. Song, N. Hou, and M. Wu, “Virtual Direct Power Control Scheme
in presence of sensor noise. of Dual Active Bridge DC-DC Converters for Fast Dynamic Response,”
IEEE Transactions on Power Electronics, vol. 33, no. 2, pp. 1750–1759,
2018.
Besides, the mismatch of the inductor value is also ex- [5] A. S. Samosir and A. H. M. Yatim, “Implementation of dynamic
amined in this work. Fig. 7 illustrates the steady-state per- evolution control of bidirectional DC-DC converter for interfacing
formance of the output voltage with the inductance value ultracapacitor energy storage to fuel-cell system,” IEEE Transactions
on Industrial Electronics, vol. 57, no. 10, pp. 3468–3473, 2010.
reduced by 50%. Compared with the classic scheme, the [6] P. Karamanakos, T. Geyer, and S. Manias, “Direct voltage control of DC-
proposed control strategy shows little voltage ripples, has more DC boost converters using enumeration-based model predictive control,”
robustness to the uncertainties of the inductor value. IEEE Transactions on Power Electronics, vol. 29, no. 2, pp. 968–978,
2014.
[7] M. Norambuena, J. Rodriguez, Z. Zhang, F. Wang, C. Garcia, and
Classic scheme R. Kennel, “A Very Simple Strategy for High-Quality Performance of
AC Machines Using Model Predictive Control,” IEEE Transactions on
Power Electronics, vol. 34, no. 1, pp. 794–800, 2018.
[8] L. Chen, L. Lin, S. Shao, F. Gao, Z. Wang, P. W. Wheeler, and
T. Dragicevic, “Moving discretized control set model-predictive control
for dual-active bridge with the triple-phase shift,” IEEE Transactions on
Power Electronics, vol. 35, no. 8, pp. 8624–8637, 2020.
[9] F. An, W. Song, B. Yu, and K. Yang, “Model predictive control with
Proposed scheme power self-balancing of the output parallel DAB DC-DC converters in
power electronic traction transformer,” IEEE Journal of Emerging and
Selected Topics in Power Electronics, vol. 6, no. 4, pp. 1806–1818, 2018.
[10] M. Abdelrahem, C. M. Hackl, Z. Zhang, and R. Kennel, “Robust Pre-
dictive Control for Direct-Driven Surface-Mounted Permanent-Magnet
Synchronous Generators Without Mechanical Sensors,” IEEE Transac-
tions on Energy Conversion, vol. 33, no. 1, pp. 179–189, 2018.
[11] Y. Wang, Z. Zhang, W. Huang, R. Kennel, W. Xie, and F. Wang,
“Encoderless sequential predictive torque control with smo of 3l-npc
converter-fed induction motor drives for electrical car applications,”
Fig. 7: [Simulation results:] comparison of output voltage Proceedings - PRECEDE 2019: 2019 IEEE International Symposium
on Predictive Control of Electrical Drives and Power Electronics, pp.
control performance between classic and proposed schemes 1–6, 2019.
with inductance mismatch. [12] F. Xiong, J. Wu, Z. Liu, and L. Hao, “Current Sensorless Control for
Dual Active Bridge DC-DC Converter with Estimated Load-Current
Feedforward,” IEEE Transactions on Power Electronics, vol. 33, no. 4,
pp. 3552–3566, 2018.
V. C ONCLUSION [13] J. Ge, Z. Zhao, L. Yuan, and T. Lu, “Energy Feed-Forward and Direct
In FCV electrical system, the DAB converter has become Feed-Forward Control for Solid-State Transformer,” IEEE Transactions
on Power Electronics, vol. 30, no. 8, pp. 4042–4047, 2015.
the critical equipment, requiring precise steady-state and fast [14] X. She, A. Q. Huang, and X. Ni, “Current sensorless power balance strat-
control dynamic performance. MPC, thanks to its fast dynam- egy for DC/DC converters in a cascaded multilevel converter based solid
ics response, is a good choice for DAB converters. However, state transformer,” IEEE Transactions on Power Electronics, vol. 29,
no. 1, pp. 17–22, 2014.
MPC depends on accurate system parameters and sensor data. [15] S. Dutta, S. Hazra, and Subhashish Bhattacharya, “A Digital Predictive
Parameters mismatch will deteriorate the control performance, Current-Mode Controller for a Single-Phase High-Frequency,” vol. 63,
and extra sensors increase hardware costs. no. 9, pp. 5943–5952, 2016.

667
Authorized licensed use limited to: Univ of Texas at Dallas. Downloaded on May 07,2024 at 21:55:35 UTC from IEEE Xplore. Restrictions apply.