You are on page 1of 8

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 1

Analysis of Finite-Control-Set Model Predictive


Current Control with Model Parameter Mismatch in
a Three-Phase Inverter
Hector A. Young, Member, IEEE, Marcelo A. Perez, Senior Member, IEEE, Jose Rodriguez, Fellow Member, IEEE

Abstract—It is well known that predictive control methods can only a finite number of output states is available. These states
be affected by the presence of modeling errors. The extent to are evaluated in order to select the optimal one in terms of
which Finite-Control-Set Model Predictive Control (FCS-MPC) the control objectives and constraints of the application. This
is influenced by parametric uncertainties is a recurrent concern
at the moment of evaluating the viability of this method for approach is known as finite-control-set MPC (FCS-MPC) [4].
power electronics applications. This paper proposes an analytic Applications of FCS-MPC have been explored in power
approach to examine the influence of model parametric uncer-
tainties on the prediction error of FCS-MPC for current control electronics systems where it contributes with several advan-
in a three-phase two-level inverter. The analysis shows that the tages such as fast transient response, simple implementa-
prediction error is not only determined by parametric mismatch tion and straightforward handling of non-linearities and con-
but also by the instantaneous values of load current and inverter straints [5]. However, the direct use of system models for
output voltage. This implies that within each sampling period of selecting the optimal control actions make predictive methods
the predictive algorithm several conditions of prediction error are
generated, as multiple voltage vectors are evaluated. Simulation prone to changes in their performance when facing modeling
and experimental results are provided and discussed, showing errors or parameter variations. Therefore, the influence of
the effects of inaccuracies in the modeling of load resistance errors in the prediction model remains as a major concern
and inductance parameters on the performance of FCS-MPC. in FCS-MPC [6].
Even though steady-state performance is noticeably affected
with parameter changes, especially when the load inductance Due to the non-linear nature of FCS-MPC, it is not possible
is overestimated by the model, its fast transient step response is to assess the effect of the parameter variation using the same
less affected by parameter changes. well-established analysis methods as in linear systems. Hence,
Index Terms—Power electronics, current control, predictive in previous research the effect of model parameter mismatch
control. has been evaluated empirically by studying the behavior of
FCS-MPC under different uncertainty scenarios. These studies
I. I NTRODUCTION have been carried out in applications involving current pre-
During the last decade, model predictive control (MPC) diction and control with three-phase two-level inverters [7]–
has received increased attention from the research community [9], active-front end rectifiers [6], multi-level voltage source
and its feasibility for a wide range of applications in power converters [10] and multiphase electric drives [11].
converters and drives has been demonstrated [1]. Within this Qualitative observations have led to practical guidelines
area of application, MPC techniques can be separated in two for adjusting the value of the modeled inductance in order
categories depending on how the switching signals that control to reduce the effects of possible mismatches [6], [7]. As
the operation of the power converters are generated [2]. In one for uncertainty in the load resistance parameter, its influence
approach the predictive algorithm yields continuous outputs on the operation of FCS-MPC has been often regarded as
which serve as references for a modulation stage [3]. The other less relevant or negligible [6], [7], [11]. Other studies have
class of MPC does not employ a modulation stage but relies focused on the separate influence of resistance and inductance
instead on the switched nature of power converters, where uncertainties on the prediction model of FCS-MPC [8], [9].
Manuscript received May 31, 2015; revised August 31, 2015 and November These works concluded that resistance variations are related to
5, 2015; accepted December 17, 2015. steady-state error in the predicted current, whereas inductance
Copyright (c) 2015 IEEE. Personal use of this material is permitted. variations are responsible for errors in both the predicted
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to pubs-permissions@ieee.org. current steady-state and dynamics, leading to increased ripple
The authors acknowldege the support of Centro Basal FB0008 Advanced in the controlled current.
Center for Electrical and Electronics Engineering (AC3E), Universidad An-
dres Bello and the Chilean National Fund for Scientific and Technological Previous research has relied mostly on empirical methods
Development (FONDECYT), under grant No. 1150829. to study the performance of FCS-MPC under parametric
H.A. Young is with the Department of Electrical Engineering, Universidad
de La Frontera, Temuco 4811230, Chile, (e-mail: hector.young@ufrontera.cl). uncertainties in the prediction model. This work is aimed
M.A. Perez is with the Electronics Engineering Department, Universi- to propose a mathematical analysis of the prediction error
dad Técnica Federico Santa Marı́a, Valparaiso 2390123, Chile, (e-mail: under uncertainty of the modeled load parameters. The study
marcelo.perez@usm.cl).
J. Rodriguez is with the Faculty of Engineering, Universidad Andres Bello, is complemented by simulation and experimental results using
Santiago, Chile, (e-mail: jose.rodriguez@unab.cl). a 5.5 kW two-level three-phase inverter.

0278-0046 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 2

work as a simple way to compensate for the unavoidable cal-


Vdc L0 R0 v0 culation delay in the digital implementation of the predictive
v i controller [12]. A cost function, designed according to the
n
required objectives and constraints, is then applied to the set
of predictions and a discrete optimization process yields the
control action to be applied to the converter. The cost function
employed in this work aims for load current reference tracking
Fig. 1. Generic voltage-source converter model.
and is given by:
k+2
gFCS = |ek+2
α | + |eβ |, (3)
II. S YSTEM MODEL AND CURRENT CONTROL METHODS
In this section the model of a generalized three-phase con- where ek+2
α and ek+2
β are the predicted tracking errors of the
verter is presented along with details of the model predictive load current in the stationary reference frame at the sampling
control current strategy. Additionally, for comparison purposes instant k + 2. These errors are calculated as:
the description of a linear PI controller with space vector ek+2 =i∗,k+2 − ik+2 (4)
α α α
modulation strategy is given.
ek+2
β =i∗,k+2
β − ik+2
β , (5)

A. Generalized voltage source converter model using the real and complex components of the current refer-
ence i∗,k+2
α , i∗,k+2
β and the prediction ik+2 = ik+2
α + jik+2
β ,
The generalized voltage source converter model is shown
which is obtained by iterating (2):
in Fig. 1, whose dynamical model is given by:  
R0 k+1 Ts k+1
ik+2 = 1 − Ts − v0 k+1 ,

d i + v (6)
v = R0 i + L0
i + v0 (1) L0 L0
dt
The voltage-source converter described by this equation can where vk+1 is the control action evaluated during the ex-
have any number of levels and phases because it takes into ecution of the predictive algorithm. Moreover, assuming a
account only the output voltage which changes accordingly to sampling frequency much higher than that of the load voltage
the converter topology. For example, in a two-level three-phase source, the latter can be considered constant between consec-
converter this voltage corresponds to a three-phase vector utive sampling instants so that v0k+1 ≈ v0k .
v = [va vb vc ] which can produce in any phase two voltage
levels, hence a total of 8 different voltage states (6 active and C. Linear current controller
2 zero) can be produced. Furthermore, in the case of a three- In order to assess the performance of the FCS-MPC subject
level single-phase neutral point converter the output voltage is to modeling error, its behavior is compared to that of a
a scalar value which can have three voltage values per arm, widely-accepted control strategy: a linear proportional-integral
hence a total 9 voltages states (5 output voltages levels with controller with space vector modulation (PI-SVM). The linear
redundancies) can be generated. current controller has been implemented in the rotating syn-
On the other hand, the model is representative whether an chronous dq frame allowing operation with zero steady-state
AC machine load or a grid connection is considered. In the error (SSE), and includes feed-forward terms to overcome
first case the inductive-resistive elements model the stator and cross-coupling effects caused by the inductance [13]. The
rotor equivalent and the voltage source v0 corresponds to the transfer function of the PI controller in the Laplace domain is
induced electromagnetic force. As for the second case, the the following:
inductive-resistive elements can model the input filter or the
AC line and the voltage source represents the grid voltage. 1 + sTi
C(s) = Kp (7)
Using the forward Euler method with a sampling time Ts the sTi
following discrete model is obtained: A well-known procedure, the magnitude optimum (MO)
 
R0 k Ts k technique [14], was employed for designing the parameters
k+1
v − v0 k .

i = 1 − Ts i + (2) of the PI-SVM controller. The design objective is to extend
L0 L0
the bandwidth of the closed-loop system as much as possible
given the constraints imposed by dynamics and delays of the
B. Predictive control algorithm system. This allows a reasonable comparison with the FCS-
The FCS-MPC algorithm works by determining, at each MPC, which features a particularly fast transient response. For
sampling interval, the optimal combination of switching states the tuning of the linear controller, the dynamics given by the
to be applied to the power converter. This is accomplished nominal load parameters (R0 and L0 ) have been considered.
by using a mathematical model of the system to predict its The value of the integral time is set to Ti = L0 /R0 , which
behavior for each of the available control actions, which are aims to cancel the dominant pole of the plant with the
the output voltage states of the converter. The predictions are controller’s zero. The proportional gain is then calculated by
made within a time horizon of usually one or two sampling Kp = L0 /2T0 . In this last expression, T0 accounts for the
periods [5]. The latter option, which implies evaluating pre- equivalent delay introduced by the modulation and the digital
dictions at the sampling instant k + 2, is employed in this implementation of the control algorithm. This delay can be

0278-0046 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 3

approximated by T0 = 1.5TSVM [15], the latter parameter 18


being the sampling period of the SVM modulation. In order 16 R/R0 = 0.5
to improve the transient performance of the PI-SVM, an anti- R/R0 = 0.75

Prediction error |∆|/|ik |(%)


14 R/R0 = 1
windup scheme based on the feedback of constrained actuation
signals through a transfer function containing the dynamics of 12 R/R0 = 1.25
the controller [16] has been implemented. 10 R/R0 = 1.50
As described above, the design procedure of the PI-SVM
depends on the knowledge of the model parameters. There- 8
fore the presence of parametric uncertainties will affect the 6
performance of the linear controller. As for the steady-state
4
performance, the integral component of the PI-SVM will
account for changes in the load parameters, allowing to obtain 2
zero steady-state error. However, the transient response is 0
expected to deteriorate since the pole-zero cancellation in the 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
design stage of the controller no longer holds when the load Inductance (L/L0)
parameters differ from their nominal values. This can give rise
Fig. 2. Magnitude of the prediction error with load inductance between 50%
to the appearance of higher order dynamics in the closed loop and 150% of its modeled value and different degrees of uncertainty in the
response, as will be shown later in Section IV. load resistance.

III. P REDICTION ERROR IN FCS-MPC DUE TO MODEL


PARAMETER MISMATCH
Equation (10) can be used to calculate the prediction error
due to any combination of values for the model parameter
Tolerances in the manufacturing process can cause small mismatches R̃ and L̃. The prediction error depends on ik ,
unbalances in loads with three or more phases, which can be an which is measured at each sampling step of the predictive
issue for particular applications [17]. However, the assumption algorithm. It also depends on the load voltage source v0 k ,
of balanced loads and simultaneous variation of the load pa- which is not affected by changes in the load inductance and
rameters is widely accepted in studies on parameter estimation resistance values. As for the inverter output voltages vk ,
and sensitivity analysis [6], [11], [18]. Consequently, this they depend only on the switching state of the inverter and
approach is also adopted in this work. its dc-link voltage. Therefore, the prediction error does not
In order to develop an analysis of the effect of a mismatch depend on the prediction error at previous steps, thus having
between the parameter values and the real system, let us review a non-cumulative nature.
the current prediction model using load parameters R0 and It is worth highlighting that the prediction error (10) is a
L0 for resistance and inductance, respectively. It is possible to complex vector that depends on the instantaneous values of
represent the parameter variation by adding uncertain compo- load current ik and voltage v∆ k
. As defined before, this latter
nents R̃ and L̃ to the resistance and inductance values used variable is the voltage difference between the inverter output
in the model. Therefore, the real load values are given by and the load voltage source. Within each sampling period the
R = R0 + R̃ and L = L0 + L̃ for resistance and inductance, load voltage source remains constant whereas a number of
respectively. Including this in (2) and introducing the variable inverter output voltages need to be evaluated, creating different
k
v∆ = vk − v0 k , the following modified prediction equation prediction error scenarios.
is obtained:
! This study considers a two-level three-phase inverter with
k+1 (R0 + R̃)Ts k Ts k a total of seven different output voltages to be evaluated on
ĩ = 1− i + v∆ , (8)
L0 + L̃ L0 + L̃ each sampling period. This leads to a wide range of cases
depending on the magnitude of voltage and current vectors
where ĩk+1 is the current prediction to the next sampling and their relative orientation in a complex plane, as well as
instant subject to parameter variations. In the following anal- the amount of parametric error in the load resistance and
ysis, for the sake of simplicity and compact notation, one- inductance. In order to limit the number of scenarios in
step-ahead predictions are used. However, the results can be which the prediction error is analyzed, a fixed amplitude of
k
extended to the case with two-step-ahead predictions. v∆ with a relative angle of 5π/6 rad with respect to the
The prediction error ∆ is defined as the difference between instant current is employed. This angle was selected because it
the prediction given by the error-free model and the one implies a condition where high prediction errors are generated,
subject to parameter uncertainty: as will be discussed later in this section. Fig. 2 shows the
prediction error for a range of load inductance mismatch and
∆ = ik+1 − ĩk+1 . (9) different degrees of uncertainty in the load resistance. The
By substituting (2) and (8) in the above equation, the following error is presented as the magnitude ratio between prediction
expression is obtained: error and the error-free current prediction (2). As expected,
the case where L/L0 = 1 and R/R0 = 1 yields error-free
Ts h i
∆= (R̃L0 − R0 L̃)ik + L̃v∆k
. (10) predictions. Moreover, the dependence of the prediction error
L0 (L0 + L̃) on changes in the load inductance is remarkably asymmetric:

0278-0046 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 4

Prediction error |∆|/|ik |(%) 20 RL load


Grid
15 L/L0 = 0.5 Vdc
R/R0 = 1.5 L/L0 = 1.5 +
10 R/R0 = 0.5
L/L0 = 1 3
R/R0 = 1.5 Voltage-source inverter
Switching Current/voltage
5 Danfoss VLT5008
signals
Vdc measurements

0
0 π/2 π 3π/2 2π
Relative angle between ik and v∆
k
(rad)
Fig. 3. Magnitude of the prediction error as a function of the relative angle PC dSPACE 1103
between instantaneous current and voltage vectors.
Fig. 4. Schematic layout of the experimental setup.

for a real inductance lower than the modeled one (L/L0 < 1), TABLE I
the effect on the prediction error is noticeably heavier than PARAMETERS USED FOR THE EXPERIMENTAL RESULTS .
in the opposite situation. This holds true with all the tested Parameter Value Unit Description
amounts of simultaneous resistance mismatch, although the Ts 50 µs Sampling period for FCS-MPC
TSVM 250 µs Sampling period for PI-SVM
effect is increased with higher R/R0 ratios. As an effect of inom 5.5 A Nominal load current
this asymmetry, the prediction error is more severe when the f∗ 50 Hz Current reference frequency
model overestimates the load inductance. 5 Ω Min. load resistance
R 10 Ω Nominal load resistance (R0 )
As can be deduced from (10), not only the parametric 20 Ω Max. load resistance
uncertainties have an effect on the prediction error but also the 5 mH Min. load inductance
k 10 mH
instant values of the load current and v∆ . Since the amplitude L Nominal load inductance (L0 )
15 mH Max. load inductance
of the voltage vectors that can be generated by the power
converter are fixed and the load voltage source is a slow-
varying disturbance, it is convenient for this analysis to con-
inverter (Danfoss VLT5008) with a dc-link voltage of Vdc =
sider as a variable only the angle between the aforementioned
100 V and a three-phase RL load. A dSPACE1103 platform
vectors. Fig. 3 shows the prediction error as a function of this
connected to a personal computer was used for implementing
angle for three cases of parametric mismatch. It is possible
the digital control algorithms.
to observe that the maximum prediction errors occur when
The SVM used with the PI controller has a switching
the angle between both vectors is close to π rad, i.e. they
period TSVM = 250 µs and is generated by the slave DSP
are opposed to each other. One of the chosen combination
unit of the dSPACE1103, a TMS320F240 with 16 bits and a
of parameter uncertainties is the one that leads to the highest
resolution of 50 ns. For the predictive controller, a sampling
prediction errors in Fig. 2 (L/L0 = 0.5 and R/R0 = 1.5),
time Ts = 50 µs has been selected. This makes possible to
confirming that these results are the worst case condition
obtain an average switching frequency comparable to that of
(≈ 18% of prediction error). The combination L/L0 = 1.5
the PI-SVM under nominal load parameter conditions.
and R/R0 = 0.5 also leads to a bell-shaped curve with a
The results are presented for operating points defined by
maximum around π rad. In the case of an exact modeling
the peak current reference i∗ in per-unit notation, referred to
of the inductance (L/L0 = 1, L̃ = 0), the dependence on
the nominal current inom . Current references with a frequency
the instantaneous angle between current and voltage is largely
f ∗ = 50 Hz and amplitudes ranging between 0.05 and 0.85 pu
reduced. This follows from (10) where L̃ links the inverter
(0.275 and 4.675 A) have been used. The parameters consid-
output voltage vector with the prediction error.
ered in the experiments are presented in Table I.

IV. E XPERIMENTAL RESULTS


B. Steady-state performance under modeling errors
In this section, experimental results are presented to illus-
trate the effects of mismatches between the modeled values The performance of FCS-MPC and the linear controller in
of the load resistance and inductance on the performance of terms of the current steady-state error (SSE) with different
FCS-MPC. The behavior of a linear PI-SVM controller under amounts of error between the modeled and the real load
similar operating conditions is also presented and discussed inductance is given by Fig. 5. The modeled inductance value is
for comparison purposes. L0 = 10 mH whereas the actual load inductances considered
were 5, 10 and 15 mH. From the results, it is clear that the PI-
SVM is able to achieve zero SSE with all the tested inductance
A. Experimental setup changes. On the other hand, FCS-MPC exhibits large steady-
The experimental setup used for this study is represented in state error for current references below 0.2 pu, even when
Fig. 4. The power stage consists of a two-level voltage-source there is no mismatch between the modeled and the real values

0278-0046 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 5

40 R/R0 = 2 L/L0 = 1.5


PI-SVM (all tests) 25

WTHD (%)
20 PI-SVM
FCS-MPC L/L0 = 1
30 15 FCS-MPC
Current SSE (%)

FCS-MPC L/L0 = 1.5 10


FCS-MPC L/L0 = 0.5 5
0
20 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a)
R/R0 = 1 L/L0 = 1
10
25

WTHD (%)
20
0 15
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 10
5
i∗ (pu) 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 5. Current steady-state error for different combinations of load inductance (b)
mismatch with FCS-MPC and PI-SVM controllers.
R/R0 = 0.5 L/L0 = 0.5
25

WTHD (%)
PI-SVM R/R0 = 1 and R/R0 = 0.5 20
PI-SVM R/R0 = 2 15 60%
10
5
50 0
FCS-MPC R/R0 = 2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Current SSE (%)

40 i∗ (pu)
FCS-MPC R/R0 = 1 (c)
30
FCS-MPC R/R0 = 0.5 Fig. 7. WTHD for different combinations of error between modeled (R0 ,
20 L0 ) and real (R, L) load parameters.

10
load resistance throughout all the operating points considered
0 in the experiment. This is consistent with the analysis of the
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 effect of resistance changes developed in Section III.
i∗ (pu) The performance of the predictive controller in terms of
weighted total harmonic distortion (WTHD) has also been as-
Fig. 6. Current steady-state error for different combinations of load resistance sessed under a range of modeling errors of the load parameters.
mismatch with FCS-MPC and PI-SVM controllers.
The results presented in Fig. 7 show three particular scenarios:
• load resistance and inductance higher than the modeled
of the load parameters. This behavior is characteristic of the values (Fig.7a),
operation of FCS-MPC considering only current tracking as a • accurate load parameters in the model (Fig.7b), and
control objective [19], [20]. Moreover, errors in the modeled • load resistance and inductance lower than the modeled
load inductance have a moderate effect on the SSE with values (Fig.7c).
the predictive controller, specially for cases where the real Firstly, it is possible to observe that changes in the load
inductance is lower than the modeled value. parameters affect the WTHD with both controllers. This is
A similar set of tests was carried out for changes in the because the resistance and inductance parameters determine
load resistance, considering a modeled resistance value of the dynamics of the load current and, therefore, the ripple
R0 = 10 Ω and actual load resistances of 5, 10 and 20 Ω. characteristics which are closely related with the harmonic
The results are presented in Fig. 6. Unlike the case where content in the signals. Taking as a reference the WTHD with
inductance modeling errors have been considered, here the PI- accurate model parameters (Fig. 7b) an unmodeled increase
SVM controller presents a high non-zero SSE which can be in the load resistance and inductance can actually decrease
explained by the saturation of the linear controller under the the WTHD, as can be seen in Fig. 7a. Nevertheless this
increase of the load resistance. As expected, this phenomenon condition leads to an undesirable increase in the SSE, as
appears with reference amplitudes higher than 0.5 p.u. when discussed earlier. The detrimental effect of modeling errors on
the load resistance is doubled. It is worth noting that under the WTHD with FCS-MPC is remarkably higher when the load
these conditions, FCS-MPC achieves a lower SSE than PI- inductance is lower than the modeled value. Overestimating
SVM since it is not restricted to the linear modulating range the resistance of the load also increases the WTHD. Therefore,
of the latter control strategy. However, the behavior of the SSE the worst case for the FCS-MPC is when both the inductance
with FCS-MPC is strongly dependent on the variations of the and the resistance values of the load are lower than the values

0278-0046 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 6

L/L0 = 1.5 L/L0 = 1.5


0.5 0.5
ia i∗a
i∗a
i∗a , ia (pu)

i∗a , ia (pu)
0 0
ia
−0.5 −0.5
0.04 0.06 0.08 0.1 0.12 0.04 0.06 0.08 0.1 0.12
(a) (a)
L/L0 = 1 L/L0 = 1
0.5 0.5
i∗a , ia (pu)

i∗a , ia (pu)
0 0

−0.5 −0.5
0.04 0.06 0.08 0.1 0.12 0.04 0.06 0.08 0.1 0.12
(b) (b)
L/L0 = 0.5 L/L0 = 0.5
0.5 0.5
i∗a , ia (pu)

i∗a , ia (pu)
0 0

−0.5 −0.5
0.04 0.06 0.08 0.1 0.12 0.04 0.06 0.08 0.1 0.12
i (pu)

i (pu)

(c) (c)

Fig. 8. Transient response in phase a with PI-SVM under load inductance Fig. 9. Transient response in phase a with FCS-MPC under load inductance
modeling errors. modeling errors.

used in the model, as can be seen in Fig. 7c. response of the PI-SVM it can be observed that changes in
the load inductance have a detrimental effect on its transient
C. Transient performance under modeling errors step response. This is natural since the tuning of the PI-SVM
considers specific values of the load parameters to adjust the
The effect of uncertainty of the load parameters on the tran- desired closed-loop response.
sient behavior of FCS-MPC has been experimentally evaluated Figs. 10 and 11 present the transient responses of the
and compared with the response of the traditional PI-SVM. PI-SVM and the FCS-MPC controllers with load resistance
This was achieved by examining the response of the controllers changes, for phase a of the current. The results show that
to step changes in the current reference under modeling error the transient response of FCS-MPC is less affected by the
conditions. resistance changes than the PI-SVM. In the case where the
The transient responses of both the studied controllers load resistance is twice the modeled value (Fig. 10a), the linear
considering modeling errors of the load inductance are plotted controller exhibits a slow transient response as compared with
in Fig. 8 (PI-SVM) and in Fig. 9 (FCS-MPC). Changes in the nominal case (Fig. 10b). On the other hand, when the
the load inductance have an important effect on the current load resistance is only 50% of its nominal value (Fig. 10c),
ripple due to its filtering effect. Therefore, a reduction in the response of the PI-SVM presents higher overshoot and
the inductance value gives rise to higher current ripple, as oscillations during the transients. Even though the response
presented in Figs. 8c and 9c. This can be observed for both of the FCS-MPC is fast with all the combinations of load
controllers, but it is most noticeable with the FCS-MPC. As resistance, there exists non-zero steady-state error when the
discussed in Section III, the effect of the inductance on the load resistance does not match the modeled value (Figs. 11a
prediction error of FCS-MPC is stronger when the real load and 11c).
inductance is overestimated by the model, as can be confirmed
with these experimental results. Despite the significant current
ripple that appears with FCS-MPC as the load inductance V. C ONCLUSIONS
decreases from its nominal value, the transient step response In this work the influence of model parameter mismatch on
of the predictive controller is practically unaffected. From the the prediction error of FCS-MPC has been analytically studied

0278-0046 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 7

R/R0 = 2 R/R0 = 2
0.5 0.5
i∗a i∗a

i∗a , ia (pu)
i∗a , ia (pu)

0 0
ia
ia
−0.5 −0.5
0.04 0.06 0.08 0.1 0.12 0.04 0.06 0.08 0.1 0.12
(a) (a)
R/R0 = 1 R/R0 = 1
0.5 0.5

i∗a , ia (pu)
i∗a , ia (pu)

0 0

−0.5 −0.5
0.04 0.06 0.08 0.1 0.12 0.04 0.06 0.08 0.1 0.12
(b) (b)
R/R0 = 0.5 R/R0 = 0.5
0.5 0.5

i∗a , ia (pu)
i∗a , ia (pu)

0 0

−0.5 −0.5
0.04 0.06 0.08 0.1 0.12 0.04 0.06 0.08 0.1 0.12
i (pu)

i (pu)

(c) (c)

Fig. 10. Transient response in phase a with PI-SVM under load resistance Fig. 11. Transient response in phase a with FCS-MPC under load resistance
modeling error. modeling error.

for current control in a three-phase two-level inverter. It has resistance responsible for the increase in the steady-state error,
been shown that the prediction error depends not only on the while changes in the load inductance are related to the increase
uncertainty of the load parameters in the model but also on the of current ripple and harmonic distortion.
instantaneous values of load current and output voltage of the
inverter. Particularly, the relative angle between both vectors R EFERENCES
(when represented in stationaty αβ coordinates) determine the [1] S. Vazquez, J. I. Leon, L. G. Franquelo, J. Rodriguez, H. A. Young,
magnitude of the prediction error, which is maximum when A. Marquez, and P. Zanchetta, “Model Predictive Control: A Review of
both vectors are in opposition. Therefore, the prediction error Its Applications in Power Electronics,” IEEE Ind. Electron. Mag., vol. 8,
no. 1, pp. 16–31, Mar. 2014.
scenario changes constantly even within a single sampling [2] P. Cortes, M. P. Kazmierkowski, R. M. Kennel, D. E. Quevedo, and
period since the FCS-MPC algorithm computes predictions for J. Rodriguez, “Predictive Control in Power Electronics and Drives,”
all the available voltage vectors at each iteration, making the IEEE Trans. Ind. Electron., vol. 55, no. 12, pp. 4312–4324, Dec. 2008.
[3] Z. Ma, S. Saeidi, and R. Kennel, “FPGA Implementation of Model
analysis a challenging task. This effect is mainly present when Predictive Control With Constant Switching Frequency for PMSM
inductance mismatches occur, being negligible under pure Drives,” IEEE Trans. Ind. Informatics, vol. 10, no. 4, pp. 2055–2063,
resistive parametric changes. It has also been demonstrated Nov. 2014.
[4] S. Kouro, P. Cortes, R. Vargas, U. Ammann, and J. Rodriguez, “Model
that the effect of inductance changes on the performance of Predictive Control—A Simple and Powerful Method to Control Power
FCS-MPC is asymmetric, being more detrimental the case Converters,” IEEE Trans. Ind. Electron., vol. 56, no. 6, pp. 1826–1838,
when the prediction model overestimates the real inductance Jun. 2009.
[5] J. Rodriguez, M. P. Kazmierkowski, J. R. Espinoza, P. Zanchetta,
value. H. Abu-Rub, H. A. Young, and C. A. Rojas, “State of the Art of Finite
The performance of the FCS-MPC has been experimentally Control Set Model Predictive Control in Power Electronics,” IEEE Trans.
evaluated and compared to that of the PI-SVM under a number Ind. Informatics, vol. 9, no. 2, pp. 1003–1016, May 2013.
[6] S. Kwak, U.-C. Moon, and J.-C. Park, “Predictive-Control-Based Direct
of changes in the load parameters. The results show that Power Control With an Adaptive Parameter Identification Technique for
the transient step response of the FCS-MPC is less affected Improved AFE Performance,” IEEE Trans. Power Electron., vol. 29,
than that of the linear PI-SVM controller under changes in no. 11, pp. 6178–6187, Nov. 2014.
[7] J. Rodriguez, J. Pontt, C. A. Silva, P. Correa, P. Lezana, P. Cortes, and
the load parameters. However, the steady-state performance U. Ammann, “Predictive Current Control of a Voltage Source Inverter,”
of the FCS-MPC deteriorates, being the changes in the load IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 495–503, Feb. 2007.

0278-0046 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIE.2016.2515072, IEEE
Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 8

[8] A. Flores, “Análisis de robustez del control predictivo aplicado a un Marcelo A. Perez (M’07-SM’14) was born in Con-
inversor trifásico de dos niveles,” Masters Thesis, Universidad Técnica cepción, Chile, in 1976. He received the Engineer
Federico Santa Marı́a, 2010. degree in electronic engineering, the M.Sc. degree
[9] J. Rodrı́guez and P. Cortés, Predictive control of power converters and in electrical engineering, and the D.Sc. degree in
electrical drives, 1st ed. Wiley-IEEE Press, 2012. electrical engineering from the University of Con-
[10] J. D. Barros and J. F. Silva, “Optimal Predictive Control of Three- cepción, Concepción, Chile, in 2000, 2003, and
Phase NPC Multilevel Inverter: Comparison to Robust Sliding Mode 2006, respectively. From 2006 to 2009 he holds
Controller,” in Proc. IEEE Power Electron. Spec. Conf. PESC 2007, a postdoctoral position at the Universidad Técnica
2007, pp. 2061–2067. Federico Santa Marı́a, Valparaı́so, Chile, conducting
[11] B. Bogado, F. Barrero, M. Arahal, S. Toral, and E. Levi, “Sensitivity research in the area of power converters. In 2009 he
to electrical parameter variations of Predictive Current Control in mul- becomes an associate researcher and since 2015 he
tiphase drives,” in IECON 2013 - 39th Annu. Conf. IEEE Ind. Electron. becomes an academic at the same institution. Dr. Perez has received the 2012
Soc. IEEE, Nov. 2013, pp. 5215–5220. Best paper award in the IEEE Industry Applications Magazine, he is Associate
[12] P. Cortes, J. Rodriguez, C. Silva, and A. Flores, “Delay Compensation Editor of IEEE Transaction on Power Electronics, Chapter Coordinator of
in Model Predictive Current Control of a Three-Phase Inverter,” IEEE the IEEE-IES Region 9 Latin-America and the President of the IEEE-IES
Trans. Ind. Electron., vol. 59, no. 2, pp. 1323–1325, Feb. 2012. Chapter Chile. His main interest are power converters topologies and control,
[13] M. Kazmierkowski, R. Krishnan, and F. Blaabjerg, Control in Power multilevel converters, simulation, electromobility, Smart Grids and HVDC
Electronics: Selected Problems, 1st ed. Academic Press, 2002. systems.
[14] K. Åström and T. Hägglund, PID controllers: theory, design, and tuning,
2nd ed. ISA, 1995.
[15] B.-H. Bae and S.-K. Sul, “A Compensation Method for Time Delay of
Full-Digital Synchronous Frame Current Regulator of PWM AC Drives,”
IEEE Trans. Ind. Appl., vol. 39, no. 3, pp. 802–810, May 2003.
[16] G. C. Goodwin, S. F. Graebe, and M. E. Salgado, Control System Design,
1st ed. Prentice Hall, 2000.
[17] Y. Xu, N. Parspour, and U. Vollmer, “Torque Ripple Minimization
Using Online Estimation of the Stator Resistances With Consideration
of Magnetic Saturation,” IEEE Trans. Ind. Electron., vol. 61, no. 9, pp.
5105–5114, Sep. 2014.
[18] B. Chen, W. Yao, F. Chen, and Z. Lu, “Parameter Sensitivity in Sensor-
less Induction Motor Drives With the Adaptive Full-Order Observer,”
IEEE Trans. Ind. Electron., vol. 62, no. 7, pp. 4307–4318, Jul. 2015. Jose Rodriguez (M’81-SM’94-F’10) received the
[19] R. P. Aguilera, P. Lezana, and D. E. Quevedo, “Finite-Control-Set Model Engineer degree in electrical engineering from the
Predictive Control With Improved Steady-State Performance,” IEEE Universidad Técnica Federico Santa Marı́a, in Val-
Trans. Ind. Informatics, vol. 9, no. 2, pp. 658–667, May 2013. paraı́so, Chile, in 1977 and the Dr.-Ing. degree in
[20] H. A. Young, M. A. Perez, J. Rodriguez, and H. Abu-Rub, “Assessing electrical engineering from the University of Erlan-
Finite-Control-Set Model Predictive Control: A Comparison with a gen, Erlangen, Germany, in 1985. He has been with
Linear Current Controller in Two-Level Voltage Source Inverters,” IEEE the Department of Electronics Engineering, Univer-
Ind. Electron. Mag., vol. 8, no. 1, pp. 44–52, Mar. 2014. sidad Técnica Federico Santa Marı́a, since 1977,
Hector A. Young (S’13-M’15) was born in Val- where he was full Professor and President. Since
paraı́so, Chile, in 1984. He received the B.Eng. 2015 he is the President of Universidad Andrés Bello
and M.Sc. degrees in electronic engineering from in Santiago, Chile. He has coauthored two books,
the Universidad de la Frontera, Temuco, Chile, in several book chapters and more than 400 journal and conference papers. His
2009, and the Ph.D. degree in electronic engineering main research interests include multilevel inverters, new converter topologies,
from the Universidad Técnica Federico Santa Marı́a, control of power converters, and adjustable-speed drives. He has received a
Valparaı́so, Chile, in 2014. Currently he is with number of best paper awards from journals of the IEEE. Dr. Rodriguez is
the Universidad de La Frontera as a lecturer in member of the Chilean Academy of Engineering. In 2014 he received the
the area of power electronics. His research interests National Award of Applied Sciences and Technology from the government of
include modeling and control of power converters Chile. In 2015 he received the Eugene Mittelmann Award from the Industrial
and electrical drives. Electronics Society of the IEEE.

0278-0046 (c) 2015 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.

You might also like