IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO.

5, OCTOBER 2007

2697

Predictive Control of a Three-Phase

Neutral-Point-Clamped Inverter
Ren Vargas, Student Member, IEEE, Patricio Corts, Student Member, IEEE, Ulrich Ammann, Member, IEEE,
Jos Rodrguez, Senior Member, IEEE, and Jorge Pontt, Senior Member, IEEE

AbstractA new predictive strategy for current control of a

three-phase neutral-point-clamped inverter is presented. The algorithm is based on a model of the system. From that model, the
behavior of the system is predicted for each possible switching
state of the inverter. The state that minimizes a given quality function g is selected to be applied during the next sampling interval.
Several compositions of g are proposed, including terms dedicated
to achieve reference tracking, balance in the dc link, and reduction
of the switching frequency. In comparison to an established control
method, the strategy presents a remarkable performance. The
proposed method achieves comparable reference tracking with
lower switching frequency per semiconductor and similar transient behavior. The main advantage of the method is that it does
not require any kind of linear controller or modulation technique,
achieving a different approach to control a power converter.
Index TermsCurrent control, DCAC power conversion,
inverters, predictive control, switching frequency, voltage control.

I. I NTRODUCTION

HREE-LEVEL neutral-point-clamped (NPC) inverters are

widely used in industry for medium- and high-voltage
power conversion and drives [1], [2]. Topics related to power
losses due to commutations and quality of the output current
are relevant issues at this power range [3][5]. The neutral-point
balancing problem in this topology is another subject that has
been studied in recent years [6][8]. Between the most common
control methods for this converter, literature states nonlinear
techniques, such as hysteresis control, and linear methods,
such as the use of proportionalintegral (PI) controllers in
conjunction with pulsewidth modulation (PWM) [9][12].
Predictive control is a control theory that was developed
at the end of the 1970s [13]. Variants of this type of control
strategy have found application in power converters. Predictive
control has been used in current control [14], drives [15][17],
power factor correction [18], and active filters [19], [20]. All
of these works consider linear models and use modulation
techniques for voltage generation. As classic solutions, the
Manuscript received May 5, 2006; revised January 26, 2007. This work
was supported in part by the Chilean Research Fund CONICYT under Grant
1050549, by the Industrial Electronics and Mechatronics Millennium Science
Nucleus, by the German Academic Exchange Service, and by the Universidad
Tcnica Federico Santa Mara. This paper was presented in part at the Power
Electronics Specialists Conference, PESC05, Recife, Brazil, June 2005.
R. Vargas, P. Corts, J. Rodrguez, and J. Pontt are with the Department of Electronics Engineering, Universidad Tcnica Federico Santa Mara,
Valparaso 2390123 Chile (e-mail: patricio.cortes@elo.utfsm.cl; rene.vargas@
usm.cl).
U. Ammann is with the Institute of Power Electronics and Control Engineering, Universitt Stuttgart, 70569 Stuttgart, Germany.
Digital Object Identifier 10.1109/TIE.2007.899854

Fig. 1. Circuit of a three-phase NPC inverter connected to a resistive

inductiveactive load.

basic idea under these methods is to consider the converter as a

linear system instead of taking advantage of the discrete nature
of the inverter and its control processor. Many of those classic
methods, such as PWM, were adapted to be used with digital
control platforms but were originally conceived from analog
electronic devices.
Model predictive control (MPC) is a subset of predictive
strategies that generate predictions from a model of the system.
A quality function is evaluated based on those predictions over
a finite receding horizon. In [21] and [22], a new variant of MPC
is used to control a matrix converter and a three-phase twolevel inverter, respectively. In both cases, the idea is to apply
the switching state that minimizes a given quality function over
a one-step receding horizon, taking into account the discrete
nature of converters. In this paper, a similar technique is developed, achieving a new control method for a three-phase NPC
inverter. Several variations of the algorithm are studied and
compared with classic PWM control, including features such
as reference tracking, balance in the dc link, and reduction of
the switching frequency.
II. M ODEL OF THE S YSTEM
Fig. 1 shows a model of the system. It includes a three-phase
three-level inverter and a resistiveinductiveactive load. The
reason to use this load is because it represents one of the most

0278-0046/$25.00 2007 IEEE

2698

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 5, OCTOBER 2007

be described as




L
Ts
i(k) + v(k + 1) e(k + 1) .
i(k + 1) =
RTs + L Ts

(6)

Equation (6) is used to obtain predictions for the future value

of the load current i(k + 1), considering all possible voltage
vectors v generated by the inverter and measured current at the
kth sampling interval.
The control strategy also uses an estimation of the future
reference current. Depending on the sampling time applied and
the computational constrains, the estimation can be obtained by
a second-order extrapolation given by
i (k + 1) = 3i (k) 3i (k 1) + i (k 2)

Fig. 2. Possible voltage vectors and switching states generated by a three-level

inverter.

common applications for this kind of converteran induction

machine [21], [22]. Also, with this model, it is possible to characterize a wide range of applications, including passive loads
and grid-connected converters. The source of the reference
current will depend on the specific application. For example,
for field-oriented control of an induction machine, the reference
current is generated from speed and flux controllers [21].
The converter applies to the load 19 voltage vectors, which
are generated from 27 switching states, as presented in Fig. 2.
The center of an MPC algorithm is the model of the plant
from which predictions are obtained. In this case, it corresponds
to the equation of a three-phase resistiveinductiveactive load,
which fulfills
L

di(t)
= v(t) Ri(t) e(t)
dt

(1)

where R and L are the load resistance and inductance, respectively, v is the voltage vector generated by the inverter, e is the
electromotive force (EMF) of the load, and i is the load current
vector. These vectors are defined as
2
v = (Va0 + aVb0 + a2 Vc0 )
3
2
i = (ia + aib + a2 ic )
3
2
e = (ea + aeb + a2 ec )
3

or, for a sufficiently small sampling time and also to save

computational efforts, it is possible to consider i (k + 1)
i (k); thus, no extrapolation is necessary.
The current prediction in (6) also requires an estimation of
the future load back EMF e(k + 1). That value, which is analog
to the future reference current case, can be estimated using
a second-order extrapolation from present and past values or
considering e(k + 1) e(k). As mentioned, that will depend
basically on the sampling time and the platform used for implementation. Present and past estimations of e can be obtained
from the load (6) shifted backward in time and load current
measurements, as follows:
(k) = v(k) +
e

RTs + L
L
i(k 1)
i(k).
Ts
Ts

(8)

Finally, each capacitor from the dc link fulfills the following

dynamic equation:
Vc (k + 1) = Vc (k) +

1
ic (k)Ts
C

(9)

where ic (k) is the current through the capacitor, vc (k) is its

voltage, and C is the capacitance. Currents through the capacitors are obtained based on the load currents and the present
switching state; thus, no additional measurements are needed.
Using (9), it is possible to obtain predictions for the future
value of the capacitors voltage based on its present current and
voltage.

(2)
III. PWM C URRENT C ONTROL M ETHOD
(3)
(4)

where a = ej(2/3) .
Applying a sampling period Ts , the derivative form di(t)/dt
is approximated by
di(t)
i(k) i(k 1)

.
dt
Ts

(7)

(5)

Replacing (5) in (1) and shifting the discrete time one step
forward, the relation between the discrete-time variables can

Before exposing the proposed predictive control method, a

short review of classic PWM current control applied to a threephase NPC inverter is presented to obtain suitable comparisons.
The selected method involves linear controllers and a modulation strategy known as level-shifted phase disposition PWM.
This alternative was selected among other PWM strategies
because it is widely used on this kind of inverter and provides
the best harmonic profile [2].
The PWM scheme is shown in Fig. 3. The load current is
measured and compared with its reference value. Next, a PI
controller generates the reference load voltages that enter a
modulator. In this stage, each reference voltage is compared
with two triangular carrier signals (superior and inferior) that

VARGAS et al.: PREDICTIVE CONTROL OF A THREE-PHASE NPC INVERTER

2699

The objective of the second term in (10), i.e., h(V c12 , nc ),

is to take advantage of the state redundancy of a three-level
inverter, from the fact that the tracking cost f depends only on
the voltage vector selected. Its composition is as follows:
h(V c12 (k + 1), nc ) = dc |V c1 (k + 1)
V c2 (k + 1)| + n nc .
Fig. 3.

Classic PWM current control method.

Fig. 4.

Predictive current control method.

are arranged in identical phase disposition. The switching state

applied to the inverter is selected according to the results of the
comparisons. For more details, see [2] and [9][11].
IV. P REDICTIVE C URRENT C ONTROL M ETHOD

g = f (i (k + 1), i(k + 1)) + h(V c12 (k + 1), nc )

(10)

where nc is the number of commutations of the power semiconductors to get to the switching state under evaluation. The
first term in (10), i.e., f (i , i), is dedicated to achieve reference tracking, quantifying the difference between the reference
current and current prediction on the next sampling time, for
a given switching state. The following composition of f , or
tracking cost, is proposed:
f (i (k + 1), i(k + 1)) = |i (k + 1) i (k + 1)|


+ i (k + 1) i (k + 1)

The first element in h adds up to g a term proportional to

the absolute difference between both capacitors voltage predictions. A switching state that generates smaller differences
will be preferred. The second element in h is proportional to
the number of commutations to get to the next switching state
nc . A switching state that implies fewer commutations of the
power semiconductors will be preferred. In this manner, the use
of h will have a direct effect in the switching frequency of the
converter. The weighing factors dc and n handle the relation
between terms dedicated to reference tracking, voltage balance,
and reduction of switching frequency within g. A large value of
a certain implies greater priority to that objective.
To measure the effect of the control strategy on the switching
frequency and reference tracking performance, it is important to define some performance variables. In the first place,
the average switching frequency per semiconductor fs will
be defined as the average value of the switching frequencies
of the 12 controlled power semiconductors in the converter
circuit. Thus
fs =

Fig. 4 shows a scheme that summarizes the implemented

control strategy. The future value of the load current and
voltages in the capacitors are predicted for the 27 switching
states generated by the inverter, by means of (6) and (9). For
this purpose, it is necessary to measure the present load current
and voltages in the capacitors. After obtaining the predictions,
a quality function g is evaluated for each switching state. The
switching state that minimizes g is selected and applied during
the next sampling period.
The proposed quality function has the following composition:

(12)

4

fsai + fsbi + fsci
12
i=1

(13)

where fski is the average switching frequency during a time

interval of the power semiconductor number i of phase k, with
i {1, 2, 3, 4} and k {a, b, c}. A reduction of the switching
frequency of the inverter will imply a reduction of fs . As the
reader can observe, fs was defined as an average between
switching frequencies. Not all 12 power semiconductors will
present the same switching frequency. Moreover, transitions
will occur in general with different current values; thus, fs will
not be directly proportional to the power losses in the converter.
However, it will allow us to measure or have indication of the
switching frequency of the inverter and the power losses due to
commutations.
The mean absolute reference tracking error e will be defined
as the mean value of the absolute difference between the reference current and the load current, within a given time interval.
As a difference between current variables, it will be measured in
[A] and will also be expressed as a percentage of the amplitude
of the reference current.
V. S IMULATION R ESULTS

(11)

where i and i are the real and imaginary components of

current vector i, respectively, and i and i are the real and
imaginary components of the reference current vector i .

Simulation results for the proposed predictive current control

strategy are presented. Within this section, the total dc-link
voltage will be maintained at 533 V, following a previous
rectification stage. The sampling period applied is Ts = 100 s.
Values of the dc link and load parameters (Fig. 1) are shown in
Table I.

2700

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 5, OCTOBER 2007

TABLE I
CIRCUIT PARAMETERS

Fig. 6.

Load voltage on phase a (simulation). (a) PWM. (b) Predictive.

Fig. 7.

Load voltage spectrum (simulation). (a) PWM. (b) Predictive.

Fig. 5. Load current response (simulation). (a) PWM. (b) Predictive.

A. Reference Tracking
The performance of the proposed strategy was analyzed
and compared with PWM current control. The algorithm was
implemented using the following quality function:
g = |i (k + 1) i (k + 1)|


+ i (k + 1) i (k + 1) + n nc .

(14)

i.e., (10) with dc = 0. The switching frequency weighing

factor applied was n = 0.001, which is small enough to select
switching states within a given voltage vector. To generate
the same average switching frequency fs as the predictive
method of about 690 Hz, the PWM carrier frequency was set at
1380 Hz.
Waveforms obtained for the load currents, load voltage, and
voltage spectrum for both methods are presented in Figs. 57,
respectively, with a sinusoidal reference current of 10-A amplitude and 50-Hz frequency. In terms of tracking performance,
the mean absolute reference tracking error e achieved was
0.089 A or 0.89% for the predictive method and 0.113 A
or 1.13% with the PWM strategy. The load voltage spectrum
obtained with the PWM method presents the characteristic side
bands near the carrier frequency. The predictive method, on
the other hand, presents a spread spectrum, with energy in a
wider range of frequencies. However, the peak amplitude of the
spectral content of the predictive method is evidently lower.
To observe the decoupling between both components of
the load current, the amplitude of i (real component of
the reference current) was reduced from 10 to 5 A at time
t = 0.035 s. The amplitude of the imaginary component i was
left at 10 A. Results for the load current are presented in Fig. 8

Fig. 8. Load current response applying step on i (simulation). (a) PWM.

(b) Predictive.

for the PWM and predictive methods. From the presented results, it is clear that the predictive method achieves comparable
performance on reference tracking during transient response.

VARGAS et al.: PREDICTIVE CONTROL OF A THREE-PHASE NPC INVERTER

In addition, note that the proposed method presents no interaction between i and i . This decoupling is a consequence of
independently considering both components of the current vector, as shown in (14).
Note that the method does not require any kind of linear controller or modulation technique, achieving a different approach
to control a power converter.
The basic predictive strategy presented, applying quality
function (10) with dc = n = 0, requires no parameter adjustment, but only knowledge of the load. Nevertheless, to
take advantage of the possibilities offered by this method, it is
necessary to adjust parameters dc and n . No design criteria
have been established thus far for this purpose. Further research
is required. However, some hints can be mentioned. First,
the designer should consider the magnitudes of the variables
involved in g. Terms included in f (11) will usually be smaller
than terms included in h (12). According to that, if the designer
wants to give equal importance to each objective, dc and
n should be less than 1. If the designer wants to maintain
voltage balance in the dc link only by selecting the appropriate
switching state within a given voltage vector, then a small value
of dc should be used. The smallest value allowed by the implementation platform will work for that purpose. In that way, the
control method will use the state redundancy of a three-level
inverter as most established methods. The same criteria can
be applied to n . With a small value, the method will choose
the switching state that implies fewer commutations within a
voltage vector. When increasing n , the method could choose
switching states that are not within the optimal voltage vector
in terms of reference tracking but imply fewer commutations.
More details regarding the selection of n will be discussed in
the following section.
B. Reduction of the Switching Frequency
Also, using (14) as a quality function and increasing the
value of n , it is possible to considerably reduce the average
switching frequency per semiconductor, i.e., fs . Applying the
strategy with n = 0, the method presented a switching frequency fs = 820 Hz. Using a value of n = 0.001, the method
presented fs = 690 Hz, as mentioned. Increasing n to 0.16
(emphasis in reducing the switching frequency), a frequency
of 229 Hz was achieved. That represents only 27.9% of the
original average switching frequency per semiconductor and
33.1% of the fs presented by the PWM method.
As expected, applying a greater n implies a reduction in the
switching frequency. In general, the tradeoff is a slight increase
in the reference tracking error. For example, increasing n from
0.001 to 0.16, the switching frequency fs drops from 690 to
229 Hz, but the reference tracking error e increases from 0.089
to 0.189 A. To expose the possibilities of the proposed method,
a graph showing the relation between the design parameter
n and the average switching frequency per semiconductor
fs and mean absolute reference tracking error e is presented
in Fig. 9. From the figure, which is built based on several
simulations for each value of n , it is possible to confirm the
mentioned relation. Increasing n implies a reduction in the
switching frequency and increases the reference tracking error.

2701

Fig. 9. Design parameter n . (a) Relation with the switching frequency.

(b) Relation with the absolute error.

Fig. 10. Predictive strategy, passive load (simulation). (a) Load current on
phase a. (b) Load voltage on phase a.

The designer should select n and dc to fit his requirements in

terms of switching frequency and reference tracking.
The capability of the predictive method to maintain voltage
balance in the dc link was also tested in simulation, but that
topic and the performance of both methods at low switching
frequencies will be presented in Section VI.
C. Passive Load
Simulation results applying the predictive strategy on a passive load are presented in Fig. 10. The algorithm was tested
in this case with circuit parameters from Table I, with the
exception that the EMF amplitude (active component of the
load) was set to zero. The method was simulated with (14) as
a quality function, with n = 0.001. The system presented a
similar behavior than with the active load. The tracking error e
and switching frequency fs were practically the same. The main
difference can be observed in the load voltage [Fig. 10(b)]. The
waveform of that signal in this case has a lower amplitude or
fundamental component, as a result of the absence of the active
component of the load. Thus, the method adjusts that variable
to maintain a load current close to the reference signal.

2702

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 5, OCTOBER 2007

Fig. 12. Experimental results with fs = 720 Hz, load current on phase a.
(a) PWM. (b) Predictive.

Fig. 11. Flow diagram of the implemented control algorithm.

VI. E XPERIMENTAL R ESULTS

The experimental prototype of the three-level NPC inverter
was built based on insulated gate bipolar transistors (IGBT)
APT25GP120BDF. The system includes an Altera Flex10K10
field-programmable gate array (FPGA) for the control and protection logic and Altera EPM7064 FPGAs for the inverter leg
logic and deadtime. The switching commands are transmitted
via an Agilent HFBR2521/1521 optic-fiber system.
The control strategy was implemented based on a dSPACE
DS1104 rapid prototyping system and MATLAB/Simulink 7.0
(R14) installed on a host personal computer. The sampling
period used with the predictive strategy was Ts = 100 s or
a 10-kHz sampling frequency. The predictive algorithm implemented with the control platform based on the dSPACE DS1104
is explained in a flow diagram presented in Fig. 11. The control
loop begins sampling the required signals. Then, the algorithm
estimates the active component of the load by means of (8)
and initializes the value of gop , which is a variable that will
contain the value of the lower quality function evaluated by
the algorithm so far. Then, the strategy enters a loop where,
for each possible switching state, the quality function (10) is
evaluated considering current and voltage predictions obtained
from (6) and (9), respectively. If, for a given switching state, the
evaluated quality function g happens to be less than gop , that
lower value is stored as gop , and the switching state number is

Fig. 13. Experimental results with fs = 720 Hz, load voltage on phase a.
(a) PWM. (b) Predictive.

stored as jop . The loop ends when all 27 switching states have
been evaluated. The state that produces the optimal value of g
(minimal) is identified by the variable jop and will be applied
to the converter during the next sampling interval, starting the
control algorithm again.
The PWM method was implemented with carrier frequencies
of 1440 and 400 Hz. The total dc-link voltage was maintained
at 533 V by a dc source that also maintained voltage balance
during initial tests. A passive load was connected to the inverter,
with the same parameters and characteristics of the load used in
Section V-C in simulations. A sinusoidal reference current of
10-A amplitude and 50-Hz frequency was applied.
The predictive strategy was tested using the quality function presented in (14) with n = 0.001. The PWM method
was implemented with a carrier signal of frequency fc =
1440 Hz. Both implementations presented an average switching frequency per semiconductor fs = 720 Hz. Results can
be observed in Fig. 12 for load current on phase a and in
Fig. 13 for load voltage. The resemblance of the signals for
both methods is clearly observed and matches the results
presented in simulations for the same situation in Fig. 10.

VARGAS et al.: PREDICTIVE CONTROL OF A THREE-PHASE NPC INVERTER

2703

TABLE II
COMPARATIVE PERFORMANCE OF PWM AND PREDICTIVE METHODS

Fig. 14. Experimental results with fs = 200 Hz, load current on phase a.
(a) PWM. (b) Predictive.

Fig. 15. Experimental results with fs = 200 Hz, load voltage on phase a.
(a) PWM. (b) Predictive.

Nevertheless, a mean absolute error of e = 0.184 A was measured for the PWM strategy. The predictive method presented a
mean absolute error of e = 0.165 A.
The second step was to increase the frequency reduction
weighing factor to n = 0.16. The predictive method presented
a switching frequency of fs = 200 Hz. The PWM method was
adjusted to match the switching frequency, with a carrier signal
of frequency fc = 400 Hz. Results on the load current for
both methods can be observed in Fig. 14. The load voltage
signals for PWM and predictive methods can be observed in
Fig. 15. Comparing Figs. 14 and 15 with Figs. 12 and 13, it
is possible to verify the reduction in the switching frequency,
as well as an increase in the reference tracking error for
both methods. Analysis of the mean absolute error, however,
reveals a significant difference in the performance of both
methods. The PWM strategy presented a mean absolute error of
e = 0.406 A, whereas the predictive method achieved a value
of e = 0.283 A, both working at fs = 200 Hz.
Table II presents a review summarizing the most relevant
characteristics and results for both methods, including average switching frequency per IGBT fs , mean absolute tracking

Fig. 16. Experimental test regarding voltage balance in the dc-link capacitors
applying the predictive strategy.

error e, and sampling frequency required to apply the method.

The theoretical maximum switching frequency that each
method can reach will depend basically on the sampling frequency. For the PWM method, the theoretical maximum fs
is equal to the sampling frequency used, whereas for the
predictive strategy, the theoretical maximum fs is equal to
half the sampling frequency. Those values limit the switching
frequency. The real fs presented by each method will depend
on specific aspects of the method, such as the n parameter of
the predictive strategy, and the behavior of the reference signal.
Finally, one of the most interesting aspects of the predictive
method is the simplicity to implement voltage balance in the dc
link. This feature was tested disconnecting the middle point of
the dc link from the source and applying the predictive control
method with the following quality function:


g = |i (k + 1) i (k + 1)| + i (k + 1) i (k + 1)
+ dc |V c1 (k + 1) V c2 (k + 1)| .

(15)

The dc weighing factor was set at dc = 0.1. The method

succeeded in maintaining voltage balance, using the same
reference signal and parameters used in previous experimental
implementations. To show the capabilities of the method, the
voltage balance section of the quality function was disabled,
setting dc = 0 at time t = 0.67 s, as presented in Fig. 16. The
method then will not consider the voltage unbalance within
g. As expected, both voltages in the dc link quickly began to
separate until the circuit protection stopped the system when
the unbalance reached 40 V.
Summarizing, the predictive current control method was
implemented, confirming observations made in simulations.
The strategy succeeded in maintaining voltage balance in the
dc link and reducing the switching frequency. Working at the
same switching frequency, the presented method achieved better reference tracking than the carrier-based method. However,
the proposed method requires a greater sampling frequency
or data acquisition frequency. The previous fact should not
be a problem, considering the new technologies available in

2704

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 54, NO. 5, OCTOBER 2007

digital signal processors (DSPs). It is important to mention

that the sampling instant is always located in a fixed position
within the sampling period, making easy the acquisition of
measurement data, avoiding problems with switching the power
devices. The dSPACE system used to obtain the results had no
problem running the algorithm at the sampling time selected,
i.e., Ts = 100 s. In fact, it took only 52 s to execute the
entire algorithm, including voltage balance and reduction of
the switching frequency. The algorithm was also implemented
on a DSP Texas Instruments TMS320F2812, using the same
sampling frequency and achieving similar results in terms of
processing times.
One of the aspects that must be mentioned is the simplicity
to implement the voltage balance strategy with the presented
method. There is no need to consider long lookup tables or
additional control blocks.
VII. C ONCLUSION
The predictive current control method presented does not
require any kind of linear controller or modulation technique.
It effectively controls the load current and compares well with
established control methods, such as PWM, achieves a comparable dynamic response and reference tracking, and works
at lower switching frequencies. If both methods are compared
at the same switching frequency, as exposed in Table II, the
predictive strategy presents lower tracking error. In addition,
the proposed method presents no interaction between both
components of the load current.
One of the remarkable aspects of the method is the use of
costs assigned to each objective to achieve reference tracking,
balance in the dc link, and reduction in the switching frequency.
The simplicity of the theory makes it easy to understand and
implement. The strategy allows the designer to adjust the parameters to fit his requirements in terms of switching frequency,
voltage balance, and reference tracking.
The method can be easily implemented taking advantage of
the present technologies available in DSPs. The higher sampling frequencies required should not be a problem nowadays.
This control strategy uses, in a very convenient way, the discrete
nature of power converters and microprocessors used in their
control.
These results show that predictive control is a very powerful
tool, with a conceptually different approach, which opens new
possibilities in the control of power converters.
R EFERENCES
[1] A. Nabae, I. Takahashi, and H. Akagi, A new neutral-point-clamped
PWM inverter, IEEE Trans. Ind. Appl., vol. IA-17, no. 5, pp. 518523,
Sep./Oct. 1981.
[2] B. Wu, High-power converters and AC motor drives, in Proc. PESC,
Recife, Brazil, 2005, pp. 23602364. Tutorial Session.
[3] H.-P. Krug, T. Kume, and M. Swamy, Neutral-point clamped three-level
general purpose inverterFeatures, benefits and applications, in Proc.
Conf. Rec. PESC, Aachen, Germany, 2004, pp. 323328.
[4] M. K. Buschmann and J. K. Steinke, Robust and reliable medium voltage
PWM inverter with motor friendly output, in Proc. 7th Eur. Conf. Power
Electron. Appl., Trondheim, Norway, 1997, pp. 35023507.
[5] T. Bruckner, S. Bernet, and H. Guldner, The active NPC converter and
its loss-balancing control, IEEE Trans. Ind. Electron., vol. 52, no. 3,
pp. 855868, Jun. 2005.

[6] R. Tallam, R. Naik, and T. Nondahl, A carrier-based PWM scheme for

neutral-point voltage balancing in three-level inverters, IEEE Trans. Ind.
Appl., vol. 41, no. 6, pp. 17341743, Nov./Dec. 2005.
[7] A. Bendre, G. Venkataramanan, D. Rosene, and V. Srinivasan, Modeling
and design of a neutral-point voltage regulator for a three-level diodeclamped inverter using multiple-carrier modulation, IEEE Trans. Ind.
Electron., vol. 53, no. 3, pp. 718726, Jun. 2006.
[8] H. du Toit Mouton, Natural balancing of three-level neutral-pointclamped PWM inverters, IEEE Trans. Ind. Electron., vol. 49, no. 5,
pp. 10171025, Oct. 2002.
[9] G. Holmes and T. Lipo, Pulse Width Modulation for Power Converters,
Principles and Practice, ser. IEEE Press Series on Power Engineering.
New York: Wiley-Interscience, 2003.
[10] M. P. Kazmierkowski, R. Krishnan, and F. Blaabjerg, Control in Power
Electronics. New York: Academic, 2002.
[11] J. Holtz, Pulsewidth modulation for electronic power conversion, Proc.
IEEE, vol. 82, no. 8, pp. 11941214, Aug. 1994.
[12] J. E. Espinoza, J. R. Espinoza, and L. A. Moran, A systematic controllerdesign approach for neutral-point-clamped three-level inverters, IEEE
Trans. Ind. Electron., vol. 52, no. 6, pp. 15891599, Dec. 2005.
[13] E. Camacho and C. Bordons, Model Predictive Control. New York:
Springer-Verlag, 1999.
[14] H. Abu-Rub, J. Guzinski, Z. Krzeminski, and H. A. Toliyat, Predictive
current control of voltage-source inverters, IEEE Trans. Ind. Electron.,
vol. 51, no. 3, pp. 585593, Jun. 2004.
[15] R. Kennel and A. Linder, Predictive control of inverter supplied electrical
drives, in Proc. Conf. Record PESC, Galway, Ireland, 2000, pp. 761766.
[16] H.-T. Moon, H.-S. Kim, and M.-J. Youn, A discrete-time predictive
current control for PMSM, IEEE Trans. Power Electron., vol. 18, no. 1,
pp. 464472, Jan. 2003.
[17] P. Wipasuramonton, Z. Q. Zhu, and D. Howe, Predictive current control
with current-error correction for PM brushless AC drives, IEEE Trans.
Ind. Appl., vol. 42, no. 4, pp. 10711079, Jul./Aug. 2006.
[18] P. Mattavelli, G. Spiazzi, and P. Tenti, Predictive digital control of
power factor preregulators with input voltage estimation using disturbance
observers, IEEE Trans. Power Electron., vol. 20, no. 1, pp. 140147,
Jan. 2005.
[19] P. Malesani, P. Mattavelli, and S. Buso, Robust dead-beat current control
for PWM rectifier and active filters, IEEE Trans. Ind. Appl., vol. 35, no. 3,
pp. 613620, May/Jun. 1999.
[20] B.-M. Han, B.-Y. Bae, and S. J. Ovaska, Reference signal generator for
active power filters using improved adaptive predictive filter, IEEE Trans.
Ind. Electron., vol. 52, no. 2, pp. 576584, Apr. 2005.
[21] S. Mller, U. Ammann, and S. Rees, New time-discrete modulation
scheme for matrix converters, IEEE Trans. Ind. Electron., vol. 52, no. 6,
pp. 16071615, Dec. 2005.
[22] J. Rodrguez, J. Pontt, C. Silva, P. Correa, P. Lezana, P. Corts, and
U. Ammann, Predictive current control of a voltage source inverter,
IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 495503, Feb. 2007.

Ren Vargas (S05) received the Engineer and

M.Sc. degrees in electronics engineering (with
honors) from the Universidad Tcnica Federico
Santa Mara, Valparaso, Chile, in 2005. He is currently working toward the Ph.D. degree at the same
university.
He worked at the Institute of Power Electronics
and Control Engineering, University of Stuttgart,
Stuttgart, Germany, during a scientific stay in 2006.
His main research interests include matrix converters and new control techniques applied to power
converters.

Patricio Corts (S05) received the Engineer

and M.Sc. degrees in electronic engineering from
the Universidad Tcnica Federico Santa Mara
(UTFSM), Valparaso, Chile, in 2004. He is currently working toward the Ph.D. degree at the same
university.
In 2003, he joined the Department of Electronics
Engineering, UTFSM, as a Research Assistant. His
main research interests are power electronics and
adjustable speed drives.

VARGAS et al.: PREDICTIVE CONTROL OF A THREE-PHASE NPC INVERTER

Ulrich Ammann (M06) received the Dipl.-Ing.

degree in electrical engineering from the University of Stuttgart, Stuttgart, Germany, in 2002. He is
currently working toward the Ph.D. degree on the
field of discrete-time modulation schemes, including
predictive techniques, at the same university.
In 2002, he joined the Institute of Power Electronics and Control Engineering, University of Stuttgart,
as a Research Assistant. His fields of interest cover
electric drives, high-power current sources, and automotive power electronics.

Jos Rodrguez (M81SM94) received the Engineer degree in electrical engineering from the Universidad Tcnica Federico Santa Mara, Valparaso,
Chile, in 1977 and the Dr.-Ing. degree in electrical engineering from the University of Erlangen,
Erlangen, Germany, in 1985.
Since 1977, he has been with the Universidad
Tcnica Federico Santa Mara, where he is currently
a Professor and the President. During his sabbatical
leave in 1996, he was responsible for the Mining
Division, Siemens Corporation, Santiago, Chile. He
has a large consulting experience in the mining industry, particularly in the
application of large drives like cycloconverter-fed synchronous motors for SAG
mills, high-power conveyors, controlled drives for shovels, and power quality
issues. He has authored or coauthored more than 130 refereed journals and
conference papers and contributed to one chapter in the Power Electronics
Handbook (Academic Press, 2006). His research interests are mainly in the area
of power electronics and electrical drives. In the last years, his main research
interests are in multilevel inverters and new converter topologies.

2705

Jorge Pontt (M00SM04) received the Engineer and M.Sc. degrees in electrical engineering
from the Universidad Tcnica Federico Santa Mara
(UTFSM), Valparaso, Chile, in 1977.
Since 1977, he has been with UTFSM, where he is
currently a Professor in the Department of Electronics Engineering, the Director of the Laboratory for
Reliability and Power Quality, and the Director of the
Nucleus for Industrial Electronics and Mechatronics.
He has authored more than 90 international refereed
journals and conference papers. He is the coauthor
of the software Harmonix used in harmonic studies in electrical systems
and also of patent applications concerning innovative instrumentation systems
employed in high-power converters and large grinding mill drives. He is a
Consultant to the mining industry, particularly in the design and application of
power electronics, drives, instrumentation systems, and power quality issues,
with management of more than 80 consulting and R&D projects. He has had
scientific stays at the Technische Hochschule Darmstadt, Darmstadt, Germany
(19791980), the University of Wuppertal, Wuppertal, Germany (1990), and
the University of Karlsruhe, Karlsruhe, Germany (20002001).