724 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 26, NO.

2, MARCH 2018

Adaptive Reference Model Predictive Control With Improved

Performance for Voltage-Source Inverters
Yun Yang, Student Member, IEEE, Siew-Chong Tan, Senior Member, IEEE, and Shu Yuen Ron Hui, Fellow, IEEE
Abstract— Power converters under the model predictive con- control strategy in which the future control inputs and system
trol (MPC) inherently suffer from nonignorable steady-state response are predicted using a system model and optimized at
residuals in its control outputs when it exists a mismatch in the regular intervals with respect to a performance index [4]–[6],
parameters between the actual system in control and the system’s
model adopted in the control. In this brief, an adaptive reference which appears to be a possible alternative for the control
MPC (ARMPC) is proposed in response to this issue. Unlike of power electronics [7]–[10]. MPC includes a wide fam-
those conventional derivatives of MPC, the ARMPC is designed ily of controllers [5], [11], [12]. In particular, one of the
to track the so-called virtual references instead of the actual most attractive predictive strategies for power electronics is
references. The virtual references are generated by a flexibly the finite-control-set MPC (FCS-MPC) [7]–[10]. Unlike the
modeled virtual multiple input multiple output system. Conse-
quently, additional tuning is not required for different operating conventional MPC, the calculation of the optimal actuation
conditions. ARMPC has been applied to a single-phase full- of FCS-MPC is achieved through the online evaluation of
bridge voltage-source inverter with both resistive and resistive- the finite number of switching states in the system, which
inductive loads. It is experimentally verified that the proposed requires no modulator. FCS-MPC inherits several advantages.
ARMPC can significantly attenuate the steady-state offsets in the First, the concept of FCS-MPC is intuitive. Second, the
environment of model mismatch (which is an inherent problem of
MPC without significantly sacrifice transient performance). Also, criteria of FCS-MPC to minimize the cost function are quite
a demonstration that ARMPC renders a consistent attenuation flexible and easily programmable. Third, FCS-MPC provides
of steady-state errors than the conventional MPC with integrator a systematic method of dealing with constraints on inputs and
is provided. More importantly, ARMPC shows better transient states, which is important for the control of complex power
performance than the MPC with integrator for some cases. electronic systems with multiple variables. Besides, FCS-MPC
Index Terms— Adaptive reference model predictive can improve the operation of the system by only including
control (ARMPC), model predictive control (MPC), MPC restrictions on some variables.
with integrator, virtual multiple-input multiple-output (MIMO) An eligible FCS-MPC requires the actual system model.
system, virtual references.
Practically, most predictive models are derived from the
nominal model of the system that expresses the circuit in
I. I NTRODUCTION terms of a fixed set of components and parameters specified
in the design, which may deviate from the actual system
T HE technology of power converters has gone through
rapid advancement during the last five decades and its
applications are fast expanding, ranging from industrial to
model. Furthermore, the nominal model is typically a sim-
plified approximated model of the actual system as precise
residential applications [1], [2]. Power electronics circuits are description of the actual system model in state-space form is
nonlinear systems with hybrid nature, including linear and typically impossible. Besides, power electronic systems may
nonlinear parts and a finite number of switching devices. own varying parameter values that change with temperature,
To achieve power regulations, hysteresis control, and linear operating conditions, and operating timespan. They are sus-
control with pulsewidth modulation based on the analog ceptible to external disturbance. All these possible scenarios
circuits have been widely used for decades [3]. However, lead to a parametric mismatch in the model between the
with the development of faster, cheaper, and more powerful actual power electronic system and the predictive system,
microprocessors, the implementation of new and more com- which could result in the existence of steady-state errors for
plex control schemes is possible. the actual system [10]. To overcome such steady-state errors,
Model predictive control (MPC) is a type of optimal control several schemes have been proposed [7]–[9]. In [7], MPC
strategy based on numerical optimization. It is a process with Kalman filter is proposed to deal with unmeasured load
variations and to attenuating the steady-state offsets. In [8], an
Manuscript received October 5, 2016; revised January 17, 2017; accepted adaptive robust predictive current control removes the current
February 11, 2017. Date of publication February 24, 2017; date of current ver-
sion February 8, 2018. Manuscript received in final form February 12, 2017. error by operating the control in parallel with the deadbeat
This work was supported by the Hong Kong Research Grant Council under algorithm. In [9], the proposed MPC with integrator reduces
Theme-based Research Project under Grant T23-701/14-N. Recommended by the steady-state errors at the sampling instant and during
Associate Editor A. Chakrabortty.
Y. Yang and S.-C. Tan are with the Department of Electrical and Elec- the intersampling. This brief presents an adaptive reference
tronic Engineering, The University of Hong Kong, Hong Kong (e-mail: MPC (ARMPC) that integrates the merits of the strategies
yangyun@eee.hku.hk; sctan@eee.hku.hk). in [7]–[9] and incorporates trajectory-based control [13]–[21].
S. Y. R. Hui is with the Department of Electrical and Electronic Engineering,
The University of Hong Kong, Hong Kong, and also with the Depart- It is an extended version of [22]. To further validate the utility
ment of Electrical and Electronic Engineering, Imperial College London, of ARMPC to alleviate the steady-state errors of power con-
London SW7 2AZ, U.K. (e-mail: ronhui@eee.hku.hk). verters for different cases and stress the merits of the dynamic
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. performance, a comprehensive presentation of ARMPC is
Digital Object Identifier 10.1109/TCST.2017.2670529 provided in this brief.
1063-6536 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: ARMPC WITH IMPROVED PERFORMANCE FOR VSI 725

ARMPC consists of two major units, virtual reference and

virtual multiple input multiple output (MIMO) system. Instead
of tracking the actual references for the conventional MPC, vir-
tual references are used for the predictive strategy in ARMPC.
Virtual MIMO system derives the virtual references as outputs
that are based on the trajectories in the states plane, which
can be flexibly modeled. According to the severity of the
model mismatch, the complexity and accuracy of the algorithm
for the virtual MIMO system are selected appropriately. The
advantages of ARMPC over the derivatives of the conventional Fig. 1. Single-phase full-bridge VSI with LC filter.
MPC, including MPC with Kalman filter, adaptive robust
predictive current control, and MPC with integrator, are simple
algorithm, easy implementation in low-price microcontrollers, Then, it is defined that x(k) = x̂(k) − x(k), u(k) =
less tuning, and inferior steady-state performance [22]. u(k)−uc (k) and the offset δ(k) = y(k)−yr (k). By subtracting
In this brief, ARMPC is applied to a single-phase full-bridge (3) and (4)
voltage-source inverter (VSI) with both resistive and resistive-     
inductive (RL) load. It is experimentally verified that ARMPC A − zI B x(k) Bd d(k)
= (5)
owns better steady-state performance than the conventional C O u(k) δ(k) + Cd d(k)
MPC and MPC with integrator. The comparisons of ARMPC, can be obtained.
MPC, and MPC with integrator in transient are also provided. Based on (5), we can derive
Despite the illustration of the adoption of ARMPC in this brief
is only on the VSI, the proposed ARMPC has potential to be δ(k) = Cx(k) − Cd d(k). (6)
widely applied to other power electronics devices.
Apparently, δ(k) = O, which denotes the existence of steady-
state errors when the conventional MPC is adopted to regulate
II. S TEADY-S TATE E RRORS OF E XISTING MPC S YSTEMS
a power electronics system.
Consider a discrete linear time-invariant power electronics The single-phase full-bridge VSI is a very well-known
system with disturbance [23] power electronics device that can be used as an ac source
⎧ or in applications such as the uninterruptable power supply,
⎪
⎨x(k + 1) = Ax(k) + Bu(k) + Bd d(k) active filter, and grid-tied inverter, etc. [2]. Most previous
y(k) = Cx(k) + Cd d(k) (1)
⎪
⎩ works on VSI that is controlled by MPC are limited to current
u(k) ∈ P := {u 1 , u 2 , . . . , u p } control and the load is an RL load [24]–[27]. Even though
where x(k) ∈ Rn denotes the vector of discrete-time state the voltage control version of MPC is studied in [28], it is
variable; y(k) ∈ Rm is the output vector; u(k) is the switching noted that a current observer is still needed. However, in many
signal; d(k) ∈ Rd is the constant disturbance vector, i.e., applications, it is necessary to control the output voltage to
d(k + 1) = d(k); A ∈ Rn×n , B ∈ Rn× p , C ∈ Rm×n , directly follow a reference and an LC filter is installed between
Bd ∈ Rn×d , Cd ∈ Rm×d , (A, B) is controllable and the bridge and the RL load, as illustrated in Fig. 1. This
(C, A) is observable. p is the number of modes of the switched makes the use of predictive current control cumbersome since
power electronics system. The objective is to design an MPC a cascaded structure or lookup table for deriving the voltage
controller based on a linear system model to have the output reference will be required. Therefore, MPC that is based on
y(k) track the reference yr (k), where yr (k) ∈ Rm is an voltage control is typically more applicable for VSI with an
asymptotically constant reference vector. LC filter.
However, a simplified linear model for the MPC design as The discrete model of the VSI with purely resistive load is
 provided in [22]. The discrete model of the VSI with RL load
x(k + 1) = Ax(k) + Bu(k) using Euler’s forward method is
(2) ⎧
y(k) = Cx(k) ⎪ Ts Ts
⎪
⎪ v C (k + 1) = v C (k) + i L (k) − i R (k)
⎪
⎪ C C
is generally adopted as the predicted equation, which derives ⎨ Ts Vdc Ts
     i L (k + 1) = − v C (k) + i L (k) + i R (k) (7)
A − zI B x̂(k) O ⎪
⎪ L L
= (3) ⎪
⎪ T RT
C O u(k) y(k) ⎪
⎩i R (k + 1) =
s
v C (k) + 1 −
s
i R (k)
L1 L1
where x̂(k) is the predicted states; I is an n-dimensional unity
matrix, O is a null matrix, and z is the Z-transform variable. where Ts is the sampling time; L 1 and R are the inductance
An ideal controller is required for the output of the system to and resistance of the load, respectively. Note that with the
track the reference such that fast computational capability of emerging controller hardware
     (e.g., DSP TMS320F28069 with a 32-b CPU of 90 MHz) and
A − zI B x(k) −Bd d(k)
= (4) the system’s order of the VSI with an RL load being only
C O uc (k) yr (k) − Cd d(k)
three, the discretization introduces very small quantization
where uc (k) is the desired control signal. errors [10].

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.
726 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 26, NO. 2, MARCH 2018

Fig. 2. Control block diagram of MPC for VSI with RL load.

By adopting the two-step ahead prediction for compensating

the calculation delay of the MPC [29], the predicted voltage
at the sampling time k + 1 will be used for the optimization
based on the measurements of v C (k − 1), i L (k − 1), i R (k − 1),
and the applied control input u(k − 1), which is given as
Ts2 T2 2Ts
v̂ C (k + 1) = 1− − s v C (k − 1) + i L (k − 1)
LC L 1C C
Fig. 3. State trajectory and the corresponding output voltage of VSI
RTs2 2Ts Vdc Ts2 with RL load before and after the appearance of model mismatch. (a) State
+ − i R (k − 1) + u(k − 1). trajectory. (b) Corresponding output voltage.
L1C C LC
(8)
The cost function is designed to contain two terms. One In trajectory-based control theory, the natural trajectories of
is called the primary term, which is established in order a system will be corresponding to the actual state-space model
to provide a proper system behavior. The primary term is of that system [13]. If there exists model mismatch, the natural
expressed in a general way as the error between the reference trajectories of the system will be changed, which is illustrated
voltage and the predicted voltage at the sampling time k +1, as by the simulation results of VSI with purely resistive load
derived in (8). The other is named the secondary term, which is in [22]. In this brief, simulation results of the state trajectory
established to improve system’s performance. The secondary (for state i L and v C ) and the corresponding output voltage of
term is expressed as the number of switching actions taken the VSI with RL load are given in Fig. 3, which presents the
from sampling time k −1 to sampling time k and the objective state trajectory and the corresponding output voltage before
of the cost function is to minimize the required switching and after the appearance of model mismatch.
frequency so as to improve the converter’s efficiency. The In Fig. 3(a), when the system is introduced a model mis-
expression of the cost function is match, the state trajectory changes from circle A to circle B.
Fig. 3(b) shows the output voltage is incapable of tracking
J =  y [v ref (k + 1) − v̂ C (k + 1)]2 + u [u(k) − u(k − 1)]2
the reference accurately, i.e., the difference of the root-mean-
(9) square (rms) value between the reference and the output
where v ref (k + 1) is the actual reference at the sampling time voltage is increased from 0 to 8.5 V.
k + 1.  y and u are the weighting factors.
Then, the trajectory for tracking the reference can be found III. P ROPOSED ARMPC FOR P OWER E LECTRONICS
by solving Consider that the reference yr (k) in (4) is replaced by the
virtual reference yr (k) as
u ∗ (k) = arg min J. (10)     
u∈{−1,0,1} A − zI B x(k) −Bd d(k)
= . (11)
Unlike those traditional optimization solvers for quadratic C O uc (k) yr (k) − Cd d(k)
programming problems, such as the active set method and We define δ  (k) = yr (k) − yr (k). By subtracting (3) and (11),
the interior point method, the decision making of (10) is we obtain
algorithmic. In particular, in the period of the sampling time     
k − 1 and k, three possible control input values are substituted A − zI B x(k) Bd d(k)
= .
in (10) to check which control input gives the minimum value C O u(k) δ(k) + δ  (k) + Cd d(k)
of the objective function. The selected input is named as the (12)
optimal control signal at the sampling time k from all the Apparently, if the virtual references yr (k) are designed prop-
possible control inputs of u = −1, 0, and 1, which is denoted erly, such that
as u ∗ (k). The corresponding control diagram can be depicted
as shown in Fig. 2. δ  (k) = Cx(k) − Cd d(k) (13)

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: ARMPC WITH IMPROVED PERFORMANCE FOR VSI 727

Fig. 4. Outputs of the system with model mismatch and virtual references.

Fig. 5. Modeling of virtual MIMO system in the state plane for the VSI.

is satisfied, δ(k) = O, which means the offsets can be

theoretically eliminated.
Similar to the geometric interpretation of virtual references law (8) based on the stored data of v C (k − 2), i L (k − 2),
for the operation of VSI in [22], an illustration of a general u(k − 2), and the measured output voltage v C (k) is defined as
power electronics system (1) is shown in Fig. 4 to present the e(k) = v̂ C (k) − v C (k). (14)
role of virtual references in reducing the steady-state errors.
To explicitly exhibit advantages of using virtual references in Then, 2n past data of the errors
a plane, only one output element y1 of the output vector y is
considered in Fig. 4. Apparently, the output chosen by MPC e = [e(k − 2n + 1) e(k − 2n + 2) . . . e(k)] (15)
(which is depicted as red lines) will lead to offsets after the
appearance of model mismatch. However, if the output is made are stored and used for the virtual MIMO system. In the
to track the virtual references, which is plotted in orange dot meantime, the variation of virtual references for the last
lines instead of the actual reference, the steady-state errors in n sampling is
the system controlled by MPC can be decreased.
vref = [v ref (k − n + 1) v ref (k − n + 2) . . . v ref (k)].
Then, developing a virtual MIMO system from the basis of
trajectory-based control [13]–[21] to generate the virtual ref- (16)
erences is critical for the design of ARMPC. The modeling of
Since linear fitting method is adopted for the modeling of
virtual MIMO system can be diversified, but is highly affected
the virtual MIMO system, the mathematical expression of the
by the specific characteristics of the objective system. In [22],
virtual MIMO system is
a linear fitting method is adopted for virtual MIMO system
considering the scale of the model mismatch of VSI in the vT
experiments. In this brief, inheriting the simplicity and utility ⎡
ref ⎤
e(k − 2n + 1) e(k − 2n + 2) · · · e(k − n)
of the virtual MIMO system in [22], more stringent descrip- ⎢e(k − 2n + 2) e(k − 2n + 3) · · · e(k − n + 1)⎥
⎢ ⎥ T
tions of the proposed virtual MIMO system with boundary =⎢ .. .. .. .. ⎥K
settings are provided. In this work on VSI, however, it is only ⎣ . . . . ⎦
necessary to use a virtual multiple input single output (MISO) e(k − n) e(k − n + 1) · · · e(k − 1)
system to accomplish the function of generating the required (17)
virtual references. Nevertheless, for the sake of generality,
the term virtual MIMO system will be adopted hereinafter where K = [K 1 K 2 . . . K n ] represents the weighting factors
as a common terminology for systems adopting this approach. of the errors on the variation of virtual references. The error
A geometric illustration of the virtual MIMO system for the matrix is nonsingular, because discretization via Euler’s for-
VSI studied in this brief can be plotted in Fig. 5. To reduce ward method introduces nonzero errors between the predicted
the steady-state errors between the operating trajectory and output voltage v̂ C and the measured output voltage v C even
the actual reference trajectory, the virtual reference trajectory if the system is free of model mismatch. For the system with
outside the actual reference trajectory is tracked, such that model mismatch, the errors in the matrix can be larger, because
the operating trajectory can be located to operate on the the disturbance factor has not been included in (8). Besides,
actual reference trajectory. A converse case that the operating since the sampling frequency of the controllers used in power
trajectory is located outside the actual reference trajectory electronics is much higher than the frequency of v C [13].
can be found in [22]. Both cases exhibit the function of As shown in Fig. 6, from e(k − 2n + 1) to e(k − 1), the
virtual MIMO system to make the consistency of the operating errors are in the same period of v C , such that
trajectory and the actual reference trajectory. e(k − n) e(k − n + 1) e(k − 1)
The implementation of the virtual MIMO system is based = = · · · = . (18)
e(k − 2n + 1) e(k − 2n + 2) e(k − n)
on several past data including states of the state variables
and virtual references. For example, the error between the Because of (18), the columns of the error matrix are linearly
predicted output voltage v̂ C (k) derived by the predicted control independent. Therefore, the error matrix can be considered

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.
728 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 26, NO. 2, MARCH 2018

Fig. 6. Two adjacent samples of v C and i L of the VSI. (a) Output

voltage v C . (b) Inductor current i L .

nonsingular. Then, K can be derived from (17) as

K = vref
⎛⎡ ⎤−1 ⎞T Fig. 7. Flowchart of ARMPC for the VSI with RL load.
e(k −2n +1) e(k −2n +2) · · · e(k −n)
⎜⎢e(k −2n +2) e(k −2n +3) · · · e(k −n +1)⎥ ⎟
⎜⎢ ⎥ ⎟
×⎜⎢ .. .. .. .. ⎥ ⎟ . for the virtual references are set within a feasible zone for
⎝⎣ . . . . ⎦ ⎠ a heuristic stability assurance. Virtual references of ARMPC
e(k −n) e(k −n +1) · · · e(k −1) are required to be located within the feasible zone. If the
(19) virtual reference is outside this zone, which corresponds to
the rms value of v c being larger than v HB (rms) or smaller
By multiplying the derived matrix K in (19) and than v LB (rms), the virtual reference will be set to v HB (rms)
[ e(k − n + 1) e(k − n + 2) . . . e(k) ]T , the variation of the and v LB (rms), respectively. This boundary setting of virtual
virtual reference v ref (k + 1) can be obtained as references ensures that the state variables are within their
v ref (k + 1) = K[e(k − n + 1) e(k − n + 2) . . . e(k)]T . acceptable ranges which keep the VSI system working stably.
(20) A 10% tolerance for the output voltage is set in the experiment,
which is validated by numerous tests.
Then, the first row of the error matrix
[ e(k − 2n + 1) e(k − 2n + 2) . . . e(k − n) ] and v ref (k − IV. E XPERIMENTAL V ERIFICATION
2n + 1) in the vector vref in (19) are replaced by The experiment is conducted on a Texas Instruments’
[ e(k − n + 1) e(k − n + 2) . . . e(k) ] and v ref (k + 1), C2000 Solar dc/ac single phase inverter [30]. The DSP used
respectively, to derive a new K for the next iteration. The for the control is TI’s F28069 Piccolo controlCARD. The
whole process detects the information of model mismatch of dc power supply is the programmable source CSW5550 from
VSI instantly and generates variation of the virtual references California Instruments. The purely resistive load used is three
simultaneously. incandescent bulbs, at nominal power of 20 W and nominal
The objective function of ARMPC for VSI with RL load is voltage of 110 V. In the nominal model of the VSI system
J =  y [v̂ ref (k + 1) − v̂ C (k + 1)]2 + u [u(k) − u(k − 1)]2 adopted in this experiment, the bulbs are connected in parallel
(21) as the nominal load. The RL load used are Kikusui’s electronic
load PLZ303WH in series with a 0.249 H inductor. The
where v̂ ref (k + 1) is the virtual reference and nominal dc input voltage of VSI with both resistive and RL
load Vdc is 165 V. The reference of VSI with both resistive
v̂ ref (k + 1) = v̂ ref (k) + v ref (k). (22)
and RL load v ref in rms is 110 V. The specifications of the
The flowchart of the algorithm of ARMPC for the VSI are given in Table I. Detailed values of the components
VSI with RL load is shown in Fig. 7. The optimization process can be found in the datasheets [30]. The number of the
of ARMPC is the same as the optimization process in (10). restored data vref for the virtual MIMO system is 3. The
To guarantee the stability of the controlled VSI, the boundaries lower bound of the virtual reference v LB (rms) is set at 100 V

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: ARMPC WITH IMPROVED PERFORMANCE FOR VSI 729

TABLE I
S PECIFICATIONS OF THE N OMINAL VSI

and the upper bound of the virtual reference v HB (rms) is set

at 120 V based on numerous tests under satisfied tolerances.
After comparisons, the weighting factors are set as  y = 0.9
and u = 0.1. The switching frequency of VSI in all the
scenarios in the experiment is set as 20 kHz. The Keysight’s
34401A Multimeter is used to measure the values of v C for
all the cases in this brief. In addition, it is emphasized that the
errors of the error matrix in the experiment are bigger than the
theoretical ones (because of unavoidable hardware delays and
calculation errors of the digital controller), which is verified
in experiment.
Fig. 8. Transient performance of VSI with load change (three bulbs to
two bulbs) controlled by (a) MPC, (b) MPC with integrator, and (c) ARMPC
A. Results of VSI With Resistive Load based on the tuned model.
1) Steady-State Performance: Comprehensive comparisons
of the conventional MPC, MPC with integrator, and ARMPC mainly shows two major advantages of ARMPC over MPC
being applied to the VSI with resistive load for the case of and MPC with integrator.
load variation and input voltage Vdc fluctuation are provided 1) ARMPC can significantly reduce the steady-state errors
in [22]. In summary, the comparative results can be concluded of VSI with resistive load when different types of model
in Table II. mismatch occurs as compared to the conventional MPC.
Here, the nominal model is the predictive model of using 2) ARMPC can be considered as an “auto-tuning” MPC
the parametric values provided by the TI’s datasheet. In reality, with integrator, which extensively reduces the steady-
model mismatch exists between the nominal model and the state residuals over a wider operating range as compared
actual VSI with resistive load. If the nominal model is adopted to the MPC with integrator.
for the MPC, steady-state errors will exist. Tuned model is a 2) Dynamic Performance: Several tests on load and input
near optimal predictive model that the steady-state offsets of voltage changes have been conducted. Fig. 8 shows the tran-
the practical inverter are minimal even when the conventional sient performance of the VSI controlled by MPC, MPC with
MPC is adopted (negligible steady-state errors exist due to integrator and ARMPC based on the tuned model when the
quantization error of DSP). It is possible to fine-tune the load changes from three bulbs to two bulbs. Compared with
coefficients of v C (k − 1) by trial-and-error to obtain a tuned MPC (settling time of 0.8 ms), both MPC with integrator
model for the control since precise model of the actual VSI (setting time of 1.2 ms) and ARMPC (settling time of 1.3 ms)
with resistive load is unavailable. It is experimentally found have slightly longer duration of transient.
via numerous tests that the steady-state performance of the The same conclusion can also be drawn when the load
system is highly sensitive to the coefficients of v C (k − 1) but changes from three bulbs to one bulb. The transient period
is insensitive to the coefficients of i L (k − 1) and u(k − 1). of VSI controlled by both ARMPC and MPC with integrator
Hence, both the coefficients of i L (k − 1) and u(k − 1) are based on the tuned model are 1.2 ms, only slightly longer than
kept unchanged in the predicted equation for MPC. The the transient period of 1.0 ms of MPC when the loads changes
measurements show that MPC with the tuned model has an from three bulbs to one bulb. Then, experiments are carried
excellent regulation performance and the rms value of v C is out for the input voltage changing from 165 to 180 V and
about 110.04 V at steady state, which is about 0.04% deviation 165 to 195 V, respectively. When the input voltage is changed
from the reference. MPC with integrator proposed in [9] is also from 165 to 180 V, v C of VSI controlled by MPC based on
adopted with a proper tuning of K i = 0.25. By using MPC the tuned model can instantaneously reach the steady state.
with integrator, the steady-state residual is reduced to 0.02% Meanwhile, v C of VSI controlled by ARMPC based on the
for the nominal model. Comparisons of relative errors of tuned model takes about 0.6 ms to settle to the steady state.
the output voltage between MPC, MPC with integrator and Besides, v C of VSI controlled by MPC with integrator based
ARMPC when load and input voltage change in Table II on the tuned model having a transient settling time of 1.2 ms.

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.
730 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 26, NO. 2, MARCH 2018

TABLE II
C OMPARISONS OF VC (rms) B ETWEEN MPC, MPC W ITH I NTEGRATOR , AND ARMPC

Fig. 9. Comparative bar-charts of different cases versus transient period Fig. 10. Comparative bar-charts of Vdc versus relative error of v C between
between the VSI with resistive load controlled by MPC, MPC with integrator, the VSI with RL load controlled by MPC, MPC with integrator, and
and ARMPC based on the tuned model. ARMPC based on the nominal model.

Obviously, ARMPC takes only a little longer time to the

steady state than MPC and the transient duration is about half
of that of MPC with integrator. The same conclusion can also
be drawn from the results of the case that the input voltage
is changed from 165 to 195 V. An explicit comparison of the
transient period for the four cases among MPC, ARMPC, and
MPC with integrator via bar-charts is shown in Fig. 9.

B. Results of VSI With RL Load

1) Steady-State Performance: The aforementioned two Fig. 11. Comparative bar-charts of resistance of RL load versus relative
error of v C between the VSI with RL load controlled by MPC, MPC with
advantages of ARMPC over MPC and MPC with integrator integrator, and ARMPC based on the nominal model.
are also experimentally verified on the VSI with an RL load.
All three controllers are initially applied based on the nominal Obviously, VSI with RL load being controlled by MPC with
model that: input voltage Vdc = 165 V, reference v ref = 110 V integrator and ARMPC have better steady-state performance
and R = 100 , L = 0.249 H. For MPC, the experimental than VSI with RL load being controlled by MPC.
results show that the output voltage is 104.291 V at steady Next, the experiments are performed on the VSI with
state, which is 5.42% deviation from the reference. For MPC RL load based on the change of resistance of the load in order
with integrator, the output voltage is 109.549 V at steady state, to observe the load effect on the output performance. The
which is 0.41% deviation from the reference. For ARMPC, the resistance of the RL load changes from 80  to 90  to 100 
output voltage is 109.637 V at steady state, which is 0.33% to 110  to 120 . The corresponding relative errors of v C
deviation from the reference. Then, the input voltage Vdc of the VSI with RL load being regulated by MPC, MPC with
changes from 173 V to 180 V to 188 V to 195 V and integrator and ARMPC are shown in Fig. 11. Both MPC with
the comparisons of relative error among MPC, MPC with integrator and ARMPC can significantly reduce the steady-
integrator and ARMPC for different Vdc is provided in Fig. 10. state residuals. Besides, ARMPC can tolerate a wider range of

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: ARMPC WITH IMPROVED PERFORMANCE FOR VSI 731

[4] E. F. Camacho and C. Bordons, Model Predictive Control. London, U.K.:

Springer, 1999.
[5] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory and
Design, Madison, WI: Nob Hill Publishing, 2009.
[6] C. Garcia, D. Prett, and M. Morari, “Model predictive control: Theory
and practice—A survey,” Automatica, vol. 25, no. 3, pp. 335–348,
May 1989.
[7] T. Geyer, G. Papafotiou, R. Frasca, and M. Morari, “Constrained optimal
control of the step-down dc-dc converter,” IEEE Trans. Ind. Electron.,
vol. 23, no. 5, pp. 2454–2464, Sep. 2008.
Fig. 12. Comparative bar-charts of different cases versus transient period [8] J. M. Espi, J. Castello, and R. Garcia-Gil, “An adaptive robust predictive
between the VSI with RL load controlled by MPC, ARMPC, and MPC with current control for three-phase grid-connected inverters,” IEEE Trans.
integrator. Ind. Electron., vol. 58, no. 8, pp. 3537–3546, Aug. 2011.
[9] R. P. Aquilera, P. Lezana, and D. E. Quevedo, “Finite-control-set model
predicitve control with improved steady-state performance,” IEEE Trans.
load deviations than the MPC with integrator without tunings Ind. Informat., vol. 9, no. 2, pp. 658–667, May 2013.
for the coefficients like the K i of the MPC with integrator. [10] J. Rodríguez and P. Cortés, Predictive Control of Power Converters and
2) Dynamic Performance: Then, the dynamic comparisons Electrical Drives, Hoboken, NJ, USA: Wiley, 2012.
[11] K.-S. Low, “A digital control technique for a single-phase PWM
among MPC, MPC with integrator and ARMPC are conducted inverter,” IEEE Trans. Ind. Electron., vol. 45, no. 4, pp. 672–674,
for the cases of input voltage change from 165 to 180 V Aug. 1998.
and 165 to 195 V, resistance of the RL load change from [12] J. Rodriguez et al., “State of the art of finite control set model predictive
control in power electronics,” IEEE Trans. Ind. Informat., vol. 9, no. 2,
100 to 80  and 100 to 120 . The results are shown in pp. 1003–1016, May 2013.
Fig. 12. Apparently, ARMPC only takes a slightly longer time [13] P. T. Krein, Elements of Power Electronics, Oxford, U.K.: Oxford Press,
(≤1.8 ms) to the steady state than the conventional MPC. 1998.
[14] R. A. DeCarlo, S. H. Zak, and G. P. Matthews, “Variable structure
Besides, ARMPC owns better dynamic performance than the control of nonlinear multivariable systems: A tutorial,” Proc. IEEE,
MPC with integrator for the input voltage change. vol. 76, no. 3, pp. 212–232, Mar. 1988.
[15] J. Y. C. Chiu, K. K. S. Leung, and H. S. H. Chung, “High-order
V. C ONCLUSION switching surface in boundary control of inverters,” IEEE Trans. Power
Electron., vol. 22, no. 5, pp. 1753–1765, Sep. 2007.
Conventional FCS-MPC will induce nonignorable steady- [16] P. K. W. Chan, H. S.-H. Chung, and S. Y. Hui, “A generalized theory
state errors of the outputs when model mismatch occurs on of boundary control for a single-phase multilevel inverter using second-
order switching surface,” IEEE Trans. Power Electron., vol. 24, no. 10,
power electronics systems. In this brief, an ARMPC that pp. 2298–2313, Oct. 2009.
is built upon the framework of MPC and trajectory-based [17] R. Alzola, P. Kslazek, M. Ordonez, and F. Blaabjerg, “Introducing
control theory is proposed and experimentally verified to state-trajectory control for the synchronous interleaved boost converter,”
in Proc. Appl. Power Electron. Conf. Expo. (APEC), Sep. 2015,
significantly alleviate the steady-state errors. ARMPC contains pp. 616–621.
two major components. One is the virtual references that [18] M. Ordonez, M. T. Iqbal, and J. E. Quaicoe, “Selection of a curved
substitute the actual references to be tracked. The other one switching surface for buck converters,” IEEE Trans. Power Electron.,
vol. 21, no. 4, pp. 1148–1153, Jul. 2006.
is the virtual MIMO system that generates required virtual [19] V. I. Utkin, Sliding Modes in Control and Optimization. Berlin, Ger-
references based on the past measured data. ARMPC can be many: Springer, 1992.
easily implemented using a low-cost digital controller. As an [20] R. Munzert and P. T. Krein, “Issues in boundary control,” in Proc. IEEE
PESC, vol. 1. Jun. 1996, pp. 810–816.
example, ARMPC is applied on a single-phase full-bridge VSI [21] S. C. Tan, Y. M. Lai, and C. K. Tse, Sliding Mode Control of Switching
with both resistive and RL. The results show that ARMPC Power Converters—Techniques and Implementation, Boca Raton, FL,
gives a better steady-state performance than the MPC when USA: CRC Press, 2012.
[22] Y. Yang, S. C. Tan, and S. Y. R. Hui, “Adaptive reference model
there exists a model mismatch, load change, and input voltage predictive control for power electronics,” in Proc. IEEE Appl. Power
change. Besides, ARMPC requires no additional tuning of Electron. Conf. Expo., Sep. 2016, pp. 1169–1175.
its control coefficients and operates efficiently over a wider [23] T. Kailath, Linear Systems, EngleWood Cliffs, NJ, USA: Prentice-Hall,
1980.
operating range as compared to the MPC with integrator. The [24] J. Rodríguez et al., “Predictive current control of a voltage source
dynamic performance of ARMPC is also attractive, since less inverter,” IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 495–503,
than 2 ms addition of the transient period are taken by ARMPC Feb. 2007.
[25] R. Vargas, P. Cortes, U. Ammann, J. Rodriguez, and J. Pontt, “Predictive
than the conventional MPC for the VSI with both resistive and control of a three-phase neutral-point-clamped inverter,” IEEE Trans.
RL load. Ind. Electron., vol. 54, no. 5, pp. 2697–2705, Oct. 2007.
[26] R. Vargas, J. Rodriguez, U. Ammann, and P. W. Wheeler, “Predictive
current control of an induction machine fed by a matrix converter with
ACKNOWLEDGMENT reactive power control,” IEEE Trans. Ind. Electron., vol. 55, no. 12,
The authors would like to thank Prof. David Hill and the pp. 4362–4371, Dec. 2008.
[27] P. Cortes, J. Rodriguez, D. E. Quevedo, and C. Silva, “Predictive current
anonymous reviewers for their feedback and suggestion. control strategy with imposed load current spectrum,” IEEE Trans.
Power Electron., vol. 23, no. 2, pp. 612–618, Mar. 2008.
R EFERENCES [28] P. Cortes, G. Ortiz, J. I. Yuz, J. Rodriguez, S. Vazquez, and
L. G. Franquelo, “Model predictive control of an inverter with output
[1] B. K. Bose, “Energy, environment, and advances in power electron- LC filter for ups applications,” IEEE Trans. Ind. Electron., vol. 56, no. 6,
ics,” IEEE Trans. Power Electron., vol. 15, no. 4, pp. 688–701, pp. 1875–1883, Jun. 2009.
Jul. 2000. [29] H. Abu-Rub, J. Guzinski, Z. Krzeminski, and H. A. Toliyat, “Predictive
[2] N. Mohan, T. M. Undeland, and W. P. Robbins, Power Electronics, current control of voltage-source inverters,” IEEE Trans. Ind. Electron.,
3rd ed. Hoboken, NJ, USA: Wiley, 2003. vol. 51, no. 3, pp. 585–593, Jun. 2004.
[3] M. P. Kazmierkowski, R. Krishnam, and F. Blaabjerg, Control in Power [30] C2000 Solar DC/AC Single Phase Inverter Schematic. (Jun. 2014).
Electronics. New York, NY, USA: Academic, 2002. [Online]. Available: http://www.ti.com/lit/df/tidr753/tidr753.pdfTI

Authorized licensed use limited to: Mahindra University - Hyderabad. Downloaded on January 27,2024 at 09:36:35 UTC from IEEE Xplore. Restrictions apply.