3670 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO.

10, OCTOBER 2008

Optimal Predictive Control of Three-Phase NPC

Multilevel Converter for Power
Quality Applications
J. Dionsio Barros, Student Member, IEEE, and J. Fernando Silva, Senior Member, IEEE

AbstractThis paper presents the optimal control of the ac There are several multilevel converter topologies available,
currents, the dc voltage regulation, and the dc capacitor voltage with the neutral point clamped (NPC) [1], the flying capacitor
balancing in a three-level three-phase neutral point clamped mul- [2], and the cascaded H-bridge [3] the most studied and used
tilevel converter for use in power quality applications as an active
power filter. The ac output currents and the dc capacitor voltages [10], [14]. Among these, the three-phase three-level NPC con-
are sampled and predicted for the next sampling time using lin- verter is the most widely used despite some difficulty of neutral-
earized models and considering all the 27 output voltage vectors. point voltage balancing [7], [10], [13], [16].
A suitable quadratic weighed cost function is used to choose the Pulsewidth modulation (PWM) and space vector modulation
voltage vector that minimizes the ac current tracking errors, the (SVM) are the most common control techniques used in multi-
dc voltage steady-state error, and the input dc capacitor voltage
unbalancing. The obtained experimental results show that the out- level power converters [19], [20]; however, controlled out-
put ac currents track their references showing small ripple, a total puts can be affected by power semiconductor switching times.
harmonic distortion (THD) of less than 1%, harmonic contents Control methods for multilevel converters based on hysteresis
that are 46 dB below the fundamental, and almost no steady-state comparators, or sliding mode approaches, are robust to model
error (0.3%). The capacitor voltages are balanced within 0.05%, components and semiconductor switching times, and present a
and the balancing is assured even when redundant vectors are not
chosen. Near-perfect capacitor dc voltage balancing is obtained zero steady-state error; however, they need a variable switching
while reducing current harmonic distortion. Some experimental frequency that is higher than those of the SVM and the PWM
evidence of robustness concerning a parameter variation was to obtain similar performance [11], [19], [20].
also found, with the optimum controller withstanding parameter Optimal predictive controllers are based on the linear optimal
deviations from +100% to 50%. Compared to a robust sliding control system theory that aims to solve the minimization prob-
mode controller, the optimal controller can reduce the THD of the
ac currents or reduce the switching frequency at the same THD, lem of a cost function [5], [15], [21][27]. Therefore, predictive
being a suitable controller for power quality in medium-voltage controllers can be tailored to minimize system output errors, at
applications. a given sampling frequency, being suitable in the simultaneous
Index TermsActive power filters (APFs), multilevel converter, control of currents and voltages with coupled dynamics. Pre-
neutral point clamped (NPC), power quality, predictive control. dictive control is a very wide concept, and different approaches
are being applied to power converters with improved results
compared to standard modulation techniques [5], [21], [22],
I. I NTRODUCTION
[27][31].

M ULTILEVEL converters are designed to evenly inher-

ently share the total dc voltage between cascaded power
semiconductors [1][3]. Therefore, compared to a two-level
Predictive controllers can be real-time-implemented in
software-based platforms using a fixed sampling frequency and
microprocessors [20]. Powerful and fast microprocessors are
converter topology, they are the most suitable power converters readily available, allowing the design of high-performance real-
for high-voltage high-power applications [1][14] despite dc time controllers for power electronic converters to optimize
voltage balancing problems and complex modulation [7], [10], the onoff duty ratios of the power switches and achieve the
[11], [15][18]. desired high-quality performance and a very low total harmonic
distortion (THD) [24], [32].
The current control of the multilevel three-phase converters
is one of the most important and classical subjects in power
Manuscript received July 25, 2007; revised June 10, 2008. First published electronics and has been studied in the last decades [4], [5],
July 9, 2008; current version published October 1, 2008. This work was [11], [13][15], [17], [18], [21], [22], [27][31], [33]. Nowa-
supported in part by the Centro de Cincia e Tecnologia da Madeira, in part
by the Fundo Social Europeu, in part by the Programa Operacional Plurifundos days, it is still a very important topic of research, particularly
da Regio Autnoma da Madeira, and in part by the Fundao para a Cincia for applications to improve power quality, which demands the
e a Tecnologia-Fundo Europeu de Desenvolvimento Regional (FCT-FEDER) use of converters that are able to strongly reduce the ac current
Project POSC/EEA-ESE/60861/2004.
J. D. Barros is with the Department of Mathematics and Engineering, or voltage distortion and/or increase the power factor [33][37].
University of Madeira, 9000-390 Funchal, Portugal (e-mail: dbarros@uma.pt). Multilevel-converter-based active power filters (APFs) have
J. F. Silva is with the Department of Electrical and Computer Engineering, potential applications in medium-voltage networks to mitigate
Instituto Superior Tcnico, Technical University of Lisbon, 1049-001 Lisbon,
Portugal (e-mail: fernandos@alfa.ist.utl.pt). harmonics and/or to increase the power factor since the power
Digital Object Identifier 10.1109/TIE.2008.928156 factor can be continuously adjustable, not dependent on the

0278-0046/$25.00 2008 IEEE

BARROS AND SILVA: PREDICTIVE CONTROL OF NPC MULTILEVEL CONVERTER 3671

Fig. 1. NPC multilevel converter circuit, working as an active power filter.

square of the medium-voltage value, and parallel resonance Assuming that a suitable control system balances the ca-
free, which are properties that are well beyond traditional pacitor voltages UC1 and UC2 , with UC1 UC2 Udc /2, the
capacitive power factor compensation [36], [37]. voltage umk between each leg and the neutral point of the
Therefore, Section II presents the multilevel-converter-based multilevel converter is given by
APF modeling. Section III deals with the optimal control of ac
currents, dc capacitor voltage regulation, and their balancing Udc
umk = k . (2)
in a three-level three-phase NPC multilevel converter APF to 2
choose the optimal vector to enforce the reference output ac
currents and the constant dc reference voltage, and to simul-
taneously balance the dc capacitor voltages. The performance A. Switched State-Space Multilevel APF Model
shown in Section IV of the proposed real-time optimal predic-
tive controller for the multilevel APF is compared to that of Applying the Kirchhoff laws to the multilevel converter APF
the robust sliding mode vector control, which uses the multi- (Fig. 1), the dynamic equations (3) of the ac line currents iL1 ,
level NPC converter redundant vectors to balance dc capacitor iL2 , and iL3 and the capacitor voltages UC1 and UC2 can be
voltages [11]. written as functions of the circuit parameters and k , i.e.,
diL1 R 0 0 11
12
II. M ODELING M ULTILEVEL APF dt L L L iL1
didtL2 0 R 0 21
iL2
22
di L L L
The NPC multilevel APF (Fig. 1) can be controlled to com- L3 = 0 0 R 31 iL3
32

dt L L L
pensate the power factor and the ac line iLk current harmonics dUC1 11 C121 C131 0 0 UC1
dt C1
introduced by nonlinear loads (the three-phase bridge rectifier dUC2
21
C222 C232 0 0 UC2
dt C2
in Fig. 1). The NPC multilevel ac side is shunt-connected to
the electrical network voltages UL1 , UL2 , and UL3 using three
L1 0 0 0
inductors L with loss resistor R to inject currents to cancel 0 L1 0 0 UL1
UL2
the harmonic component currents of the nonlinear load. The dc +
0 0 L1 0
UL3
voltage Udc must be regulated, and the capacitor voltages UC1 0 0 0 1
C1
and UC2 must be balanced. 1 idc
0 0 0 C2
The switching variables k represent the state of the mul-
tilevel converter active switches Skj , with k {1, 2, 3} and
iR1
j {1, 2, 3, 4}. Assuming ideal semiconductors, the three valid R 0 0 1 0 0
L i
combinations of the binary states of the switches Skj of each leg 0 R 0 0 1 0 R2
L iR3
k can be defined as follows: +
0 0 R L 0 0 1
didtR1


11 C121 C131 0 0 0
diR2


1 (Sk1 = 1 Sk2 = 1) (Sk3 = 0 Sk4 = 0) C1
dt
k = 0 (Sk1 = 0 Sk2 = 1) (Sk3 = 1 Sk4 = 0) C212 C222 C232 0 0 0 diR3
dt
1 (Sk1 = 0 Sk2 = 0) (Sk3 = 1 Sk4 = 1).
(1) (3)
3672 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 10, OCTOBER 2008

where

k (k + 1)
1k =
2
k (1 k )
2k =
2

1 
3

ki = 2ik ij . (4)
3 j=1
j=k

For ac networks with isolated neutral, these equations can be

simplified using coordinates.

B. Switched State-Space Multilevel System

Model in Coordinates
Fig. 2. Available voltage vectors at the output of the multilevel converter.
The relationship between the variables X123 (in system
coordinates) and the coordinates X is given by (C is the i values can be used to plot the 27 vectors of Fig. 2. (The
ClarkeConcordia transformation matrix) values of 1 , 2 , and 3 are inside the parentheses.)

X123 = CX
C. Switched State-Space Multilevel System

2
Model in dq Coordinates
X1 1 0 X
2
X2 = 2 21 3 2
X . (5) The relationship between the variables X in coordi-
3 2 2
X3 21 23 2 X0 nates and Xdq in dq coordinates is given by the Park transfor-
2
mation, i.e.,
Applying (5) to model (3), the coordinates multilevel
X = DXdq
model (6) is obtained, i.e.,     
X cos sin Xd
= . (8)
diL R 0 1 2
X sin cos Xq
dt L L L i
diL
0 R
1 2 L
= L L L iL

dt
1 The argument = t is the angular phase of the electrical
dUC1
dt C1 C11 0 0 UC1 network ac line voltage. Applying the Park transformation (8)
dUC2
C2 C22

0 0 UC2 to the multilevel converter model (6), the multilevel converter
dt
1
2

L 0 0 APF model (9) in dq coordinates is obtained, i.e.,

0 1
0 UL diLd R 1d 2d
+
0
L
1 UL dt L L L iLd
diLq 2q
0 1q
C1 R
idc dt = L L L iLq
dUC1
1
0 0 1q
C2 1d
0 0 UC1
R 0 1 0 iR dt C1 C1
L
dUC2
2d 2q

0 0 UC2
dt C2 C2
0 L R
0 1 iR
+ diR (6)
1 1 0 0 dt L1 0 0
C1 C1
0 ULd
diR L1 0
+

C2 C22 0 0 dt
0 1 ULq
2 0 C1
1 idc
where 0 0 C2
R 1 0 iRd
  L
2 i2 i3 R 0 1 iRq
i = i1
+ 1d L di (9)
3 2 2 0 dtRd

C1 C1q1 0
   diRq
2 3 3 C2d2 C2q2 0 0 dt
i = i2 i3 . (7)
3 2 2
where
    
This model is suitable to establish the optimal predictive id cos sin i
= , with i {1, 2}. (10)
current controller, and the 27 possible combinations of i and iq sin cos i
BARROS AND SILVA: PREDICTIVE CONTROL OF NPC MULTILEVEL CONVERTER 3673

The dynamic equation of the dc voltage Udc is directly where UCi (ts ) and ULx (ts ) are the sampled voltages. Assum-
obtained from (9) since Udc = UC1 + UC2 , i.e., ing ix (x {, } and i {1, 2}) as the control input during
  T , the dynamic behavior of the ac currents is approximately
dUdc 1d 2d 1q 2q
= + iLd + iLq
iLx (ts+1 ) iRx (ts+1 ) + e L T [iLx (ts ) + iRx (ts )]
R
dt C1 C2 C1 C2
 
1d 2d 1q 2q  
+ iRd + iRq 1x 2x ULx (ts )
C1 C2 C1 C2 + UC1 (ts ) + UC2 (ts )
 L L L
1 1
+ + idc . (11) L 
1 e L T .
R
C1 C2 (14)
R
This quasi-linear model of the dc voltage Udc is suitable to
design a linear predictive controller to regulate the dc voltage Assuming (R/L)T  1, the exponential function can be
of the APF. approximated by the first two terms of its Taylor series, i.e.,
R R R
e L T 1 + T, T  1. (15)
III. O PTIMAL P REDICTIVE C ONTROLLER D ESIGN L L
The ac current and the dc voltage optimal controllers are Using (15) in (14), the ac line currents (14) can be predicted
designed to choose the best output voltage vector that is able to at the next sampling time using the following:
minimize the ac line current iL and iL errors, the Udc voltage
regulation error, and the unbalancing of dc capacitor voltages iLx (ts+1 ) iLx (ts ) + iRx (ts ) iRx (ts+1 )
UC1 and UC2 . The obtained converter model (6) will be R R
solved to predict the state variable values at the next sampling T iLx (ts ) T iRx (ts )
L L
period for all the 27 available vectors.
1x 2x
+ T UC1 (ts ) + T UC2 (ts )
L L
A. Predictive Equations for AC Line Currents T
ULx (ts ). (16)
and DC Capacitor Voltages L
The solution of the ac line current iLx , with x {, }, is To predict the capacitor voltage difference UC1 (t) UC2 (t),
given by [23] the corresponding dynamic equations (6) must be solved. The
solution for the capacitor voltages is given by [23]
 ts+1
R
L T i i
iLx (ts+1 ) = iRx (ts+1 )+e iLx (ts )+iRx (ts ) UCi (ts+1 ) = UCi (ts ) + iL ( ) iR ( )
Ci Ci
ts
ts+1    i i idc ( )
R 1x 2x ULx ( ) iL ( ) iR ( ) + d (17)
+ eL UC1 ( )+ UC2 ( ) d Ci Ci Ci
L L L
ts
where UCi (ts ) are the sampled capacitor voltages at t = ts =
(12) kT . Assume T to be small enough to consider the dc
current idc (t) and nearly constant, and assume that the ac line
currents follow their references, i.e.,
where iLx (ts ) and iRx (ts ) are the currents iLx (t) and iRx (t) at
the sampling instant ts = kT , respectively, and iLx (ts+1 ) = idc ( ) idc (ts )
iLx (t = (k + 1)T ) is the current to be predicted for the next
iLx ( ) iLx (ts )
(k + 1) sampling interval T . The nonlinear current for the
next sampling interval iRx (ts+1 ) can be estimated from the iRx ( ) iRx (ts ), ts < t < ts+1 . (18)
sampled value of the previous period T of ac currents, i.e.,
iRx (ts+1 ) iRx (ts+1 T ), that is updated every period T . Substituting (18) into (17) and assuming ix (x {, } and
Assuming a sampling time T that is small enough, UC1 (ts ), i {1, 2}) as the control during T , the capacitor voltages are
UC2 (ts ), and ULx (ts ) can all be considered nearly constant i
during T (T must be at least ten times smaller than the UCi (ts+1 ) UCi (ts ) T iL (ts )
Ci
minimum period of voltages UC1 , UC2 , and ULx ), i.e.,
i i
T iR (ts ) T iL (ts )
Ci Ci
UCi ( ) UCi (ts )
i T
T iR (ts ) + idc (ts ). (19)
ULx ( ) ULx (ts ), ts < t < ts+1 (13) Ci Ci
3674 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 10, OCTOBER 2008

Considering that C1 C2 C, the capacitor voltage unbal-

ance is predicted to be

UC1 (ts+1 ) UC2 (ts+1 )

T
UC1 (ts ) UC2 (ts ) + (2 1 )
C
[iL (ts ) + iR (ts )] + (2 1 )
T
[iL (ts ) + iR (ts )] . (20)
C

B. Quadratic Cost Function Definition

If a three-leg NPC multilevel converter is operated as a cur-
rent source, three independent variables can be controlledtwo
ac line currents iL (t) and iL (t), and the capacitive unbal-
ance UC1 (t) UC2 (t). The main objective of the optimizing
controller must be the minimization of the ac line current
errors and the capacitor voltage difference using the following
suitable quadratic cost function of the tracking errors given in
(22) and (23):

e2 (ts+1 ) e2 (ts+1 ) e2U C (ts+1 )
C(ts+1 ) = + + (21) Fig. 3. Feedback system of the multilevel converter APF with a predictive
U C
optimal controller.
where D. APF Current Reference Generation
and Power Flow Control
e (ts+1 ) = iLRef (ts+1 ) iL (ts+1 )
e (ts+1 ) = iLRef (ts+1 ) iL (ts+1 ) (22) The generation of the current references for the APF must
eU C (ts+1 ) = UC1 (ts+1 ) UC2 (ts+1 ). (23) assure that the ac line currents are sinusoidal, the power factor
is near unity, and the dc voltage is regulated.
iLRef (ts+1 ) and iLRef (ts+1 ) are the ac line current ref- A linear optimal predictive regulator is designed to generate
erences one sample time forward, i.e., ts+1 , to compensate the ac line current references iLdRef and iLqRef to cancel
for the processor calculation times. In cost function (21), the the mean value of the dc voltage error eUdc (t) = UdcRef (t)
errors are weighted by , , and U C for two purposes: Udc (t) during the time interval TUdc , i.e.,
1) they normalize the distinct errors, which have different units 
1
and ranges, and 2) they define the priority level of each error eUdc (t) = eUdc (t)dt
TUdc
variable. TUdc
To assure the adjacent level voltage transition in the multi- 
1
level output leg voltages, only adjacent vectors (needing just = UdcRef (t) Udc (t)dt = 0. (24)
TUdc
the switching of two semiconductors per leg) are analyzed. TUdc
The adjacent vectors must simultaneously verify |k (ts+1 )
k (ts )| < 2, with k {1, 2, 3}, in all legs of the multilevel To obtain a quasi-unity power factor, the ac line current
converter. must be in phase with the ac line voltage, forcing iLqRef = 0.
Using the dynamic equation of the dc voltage [see (11)] in (24),
assuming that the optimal predictive controller forces iLd =
C. Multilevel Converter Optimal Predictive Controller iLdRef and iLqRef = 0, the optimal predictive control law of
The block diagram of the multilevel converter with the opti- iLdRef is obtained, i.e.,
mal controller (Fig. 3) includes as inputs the states of the mul-
UdcRef C UdcRef (t) Udc (t)
tilevel converter switches (all the vectors to be tested) 1 (ts ), iLdRef (t) iRd (t)
2 (ts ), and 3 (ts ), the sampled capacitor voltages UC1 (ts ) 6UL TUdc
and UC2 (ts ), the ac line current references one sampling time 2
UdcRef L UdcRef L
forward iLRef (ts+1 ) and iLRef (ts+1 ), and the ac currents 2 iRq (t) + idc (t) (25)
6UL Rdc 6UL
iL1 (ts ), iL2 (ts ), iL3 (ts ), iR1 (ts ), iR2 (ts ), and iR3 (ts ), which
were read and transformed to coordinates iL (ts ), iL (ts ), where TUdc is the time interval used to compute the mean value
iR (ts ), and iR (ts ) using the ClarkeConcordia transforma- of the dc voltages Udc and UdcRef , the dc current idc , and
tion. The controller makes use of these inputs to compute the the nonlinear currents iRd and iRq . UL is the ac line voltage
optimal vector to apply it to the multilevel converter at the next amplitude, and is its fundamental angular frequency. The
sampling interval. time interval TUdc must be much larger than the period T of
BARROS AND SILVA: PREDICTIVE CONTROL OF NPC MULTILEVEL CONVERTER 3675

RL = 10 , Rdc = 10 k, RR = 0.1 , iac = 7 A, UL ob-

tained from 230/400 V through a 400/230-V transformer, fac =
50 Hz (T = 0.02 s and = 314.2 rad/s), TUdc = 2 s, =
0.09 A2 , = 0.09 A2 , and U C = 0.04 V2 .
A. Evaluation of NPC Multilevel
Converter Current Controllers
To evaluate the performance of the multilevel ac current
control and the capacitor voltage balancing with the optimal
predictive current controller, a dc source Udc (120 V) with
internal resistance of 0.2 is connected in the dc side of the
multilevel converter (Fig. 1), the nonlinear load is disconnected
from Fig. 1 (resulting in iL1 = i1 , iL2 = i2 , and iL3 = i3 ),
and the current references iLRef = iRef and iLRef = iRef
Fig. 4. Active power filter current reference generation. are obtained from DSP-based lookup tables. The results of
ac line currents (TUdc  T ) to maintain the ac line currents the multilevel currents and the capacitor voltages, obtained
sinusoidal (the time constant Rdc C is of no concern since Rdc with the optimal predictive controller, are compared with a
represents the safety discharging resistors of the capacitors; hysteresis-based sliding mode controller [11] since it also
TUdc  Rdc C). In (25), the switching variables 1d + 2d works at a fixed sampling frequency but at a variable switching
and 1q + 2q are given by the following and were similarly frequency.
obtained as in [38]: Simulation and experimental results of the ac currents i1 ,
i2 , and i3 , with T = 28 s and a steady-state operation
6UL [Fig. 5(c) and (d)], show that they are nearly sinusoidal with a
1d + 2d very small ripple factor (less than 3%) and a THD that is lower
UdcRef
than 1%. In comparison, the sliding mode controller shows a
2LUdcRef
1q + 2q . (26) more than 7% ripple factor [Fig. 5(a) and (b)] and a THD of
6UL Rdc around 6%.
In the feedback controller system of the APF, the dc voltages The experimental power spectral density of the ac current
Udc and the currents idc , iR1 , iR2 , and iR3 are sampled to i1 using the sliding mode controller [Fig. 6(a)] shows that the
generate the ac line current references (Fig. 4), which are main harmonics are nearly 32 dB below the 50-Hz fundamental.
applied to the optimal predictive controller of the multilevel The power spectral density shows switching frequencies of over
converter (Fig. 3) to force the ac line currents to follow its 3.5 kHz, with main spectral lines below 2 kHz. Using the op-
references. timal predictive controller [Fig. 6(b)], the experimental results
of the power spectral density of the ac current i1 show the main
harmonics to be nearly 46 dB below the 50-Hz fundamental,
IV. S IMULATION AND E XPERIMENTAL R ESULTS
an improvement of 14 dB compared to that of the sliding mode
The multilevel converter model and the optimal controller controller. The switching frequency of the optimal predictive
were implemented in the MATLAB/Simulink, as shown in controller is spreading over the frequency spectrum, which is
Figs. 3 and 4. The dynamic equations of the system model (3), advantageous in reducing the audible noise, with main spectral
the ac line currents iL1 (t), iL2 (t), and iL3 (t), and the ca- lines, compared to that of the sliding mode, of mainly below
pacitor voltages UC1 (t) and UC2 (t) were implemented in the 2.5 kHz. It has been found that the mean value of the switching
simulation program. Ideal switches were considered. The op- frequency of the NPC converter can be reduced by a factor
timal predictive controller was implemented in a MATLAB of 2, compared to that of the sliding mode, by increasing the
S-Function. To experimentally validate the simulations, a low- ripple factor of the optimum predictive controller algorithm to
power (3 kW) laboratorial prototype was built, according the levels of the sliding mode controller. This can be done by
to Figs. 1, 3, and 4, with insulated-gate bipolar transistors decreasing the sampling period of the microprocessor, which
(IGBTs; MG50Q2YS40), diodes (STTA5012TV1), capacitors, is further advantageous since it can decrease the cost of the
inductors, resistors, IGBT drives with optical isolation (IR2110 computing and sampling devices.
together with optocouplers HCPL-2200), sensors to read the The improvement of the optimal controller is due to the cho-
currents iL1 (t), iL2 (t), iR1 (t), iR2 (t), and idc (t) (Hall effect sen optimal vector that minimizes the errors of the ac currents
LEM LA25NP), and sensors to read voltages UC1 (t) and and the capacitor voltage imbalance. The optimal controller just
UC2 (t) (isolation amplifier AD210AN). To determine the op- chooses the vector, leading to a minimum current error at the
timal vector, the algorithm is implemented in a digital signal end of each sampling time. The sliding mode controller always
processor PowerPC-based board (DS1103), which is pro- chooses a vector to ensure fast reaching of the sliding mode
grammed in C language (T ranging from 28 s to 1 ms). regime, unnecessarily increasing the ripple factor, the THD, and
The following parameters were used: C1 = 4.4 mF, 200 V, the switching frequency.
C2 = 4.4 mF, 200 V (Udc max < 300 V to have a 25% safety The sliding mode controller is characterized by having a fast
margin), CL = 1 mF, L = 15.1 mH, LR = 1 mH, R = 0.1 , dynamic response to step ac current reference changes [11].
3676 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 10, OCTOBER 2008

Fig. 5. Sinusoidal ac currents i1 , i2 , and i3 in a steady-state operation. i1 has a displacement of 2 divisions, and i3 has a displacement of 2 divisions (vertical,
12 A/div; horizontal, 10 ms/div). (a) Sliding mode controller simulations. (b) Sliding mode controller experiments. (c) Optimal controller simulations. (d) Optimal
controller experiments.

Fig. 6. Power spectral density of the ac current i1 in a steady-state operation (vertical, 10 dB/div; horizontal, 500 Hz/div). (a) Sliding mode controller experiments.
(b) Optimal controller experiments.

Applying a positive step (3.5 to 7 A) in the ac current references dynamic response and also indicates that after the rise time,
(Fig. 7, at 35 ms) and a negative step (7 to 3.5 A) at 75 ms the ac current i2 follows its reference, i.e., i2Ref , without a
to compare the performance of the dynamic response for these stationary error [Fig. 7(d)].
two controllers, the results show that they have similar dynamic The optimal predictive controller also chooses the opti-
responses with no overshoot. The zoom of current i2 [Fig. 7(c) mal vector that minimizes the capacitor voltage errors, and
and (d)] at the negative step time (75 ms) shows the detailed these voltages are balanced [Fig. 8(b)], with a mean error
BARROS AND SILVA: PREDICTIVE CONTROL OF NPC MULTILEVEL CONVERTER 3677

Fig. 7. Sinusoidal ac currents i1 , i2 , and i3 with an amplitude step. i1 has a displacement of 2 divisions, and i3 has a displacement of 2 divisions [vertical,
12 (A/div)/6 (A/Div); horizontal, 10 (ms/Div)/2 (ms/Div)]. (a) Sliding mode controller experiments. (b) Optimal controller experiments. (c) Sliding mode controller
experiments (i2 zoom). (d) Optimal controller experiments (i2 zoom).

Fig. 8. Capacitor voltages UC1 and UC2 with the ac current experiment step variations (vertical, 10 V/div; horizontal, 20 ms/div). (a) Sliding mode controller
experiments. (b) Optimal controller experiments.

of lower than 0.05%. The ac measurements of the capacitor

voltages show, based on the zooming results of Fig. 8(b),
with a vertical scale of 0.4 V/div, that the two capacitor volt-
ages are really overlapped. Near-perfect capacitor dc voltage
balancing is obtained while reducing the current harmonic
distortion.
The capacitor voltages of the sliding mode controller are also
balanced, but they present a ripple error of about 6% [Fig. 8(a)].
The step variations of the ac current amplitude do not affect
the balancing of the capacitor voltages (Fig. 8) in these two
controllers.
To investigate the influence of the 6% ripple of the sliding
mode equalized capacitor voltages, in the ac current ripple
and the THD, simulations were done using two dc voltage
sources to enforce UC1 = UC2 = Udc /2 instead of the capaci- Fig. 9. Simulation results of ac currents i1 , i2 , and i3 using the sliding
tors. The sliding mode controlled ac currents (Fig. 9) presented mode controller and two dc voltage sources to enforce UC1 = UC2 = Udc /2
almost the same ripple factor (7\%) and THD (6%) as in the (vertical, 12 A/div; horizontal, 10 ms/div).
3678 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 10, OCTOBER 2008

Fig. 10. Sinusoidal ac currents i1 , i2 , and i3 with nonnominal parameters. i1 has a displacement of 2 divisions, and i3 has a displacement of 2 divisions
(vertical, 12 A/div; horizontal, 10 ms/div). (a) Double inductance 2 L and half resistance R/2 test. (b) Half electrical network voltages UL /2 test.

Fig. 11. Capacitor voltages UC1 and UC2 with nonnominal parameters (vertical, 10 V/div; horizontal, 20 ms/div). (a) Double capacitance 2 C1 and
2 C2 test. (b) Half capacitance C1 /2 and C2 /2 test.

conditions of Fig. 5(a), showing that small levels of the capaci- TABLE I
THD OF PREDICTIVE CONTROLLED AC CURRENTS
tor voltage ripple [Fig. 8(a)] do not significantly affect the ripple AS A F UNCTION OF S AMPLING T IME T
and the THD of ac currents that are controlled by the sliding
mode method.
The sliding mode controller uses a robust control law, im-
plying a parameter-independent dynamic behavior. The optimal
predictive controller, which chooses the optimal vector of the
multilevel converter, is a model-based controller depending on
system parameters [the capacitors C1 and C2 , the inductors L, The ac line voltage during the experiments presented a THD
with loss resistor R, the electrical network UL , voltages (Fig. 1), of about 5%, and the THD of ac currents with predictive control
and the sampling time T ]. To investigate the robustness of is lower than 1%. This illustrates that the optimal predictive
the optimal predictive controller, with industrial component controller tolerates common disturbances of the electrical net-
tolerances, the parameters of the predictive laws were increased work voltage.
100% or decreased 50% in two limit situations: 1) 2 C1 , 2 Table I contains the evolution of the THD of the predictive
C2 , 2 L, R/2, and 2 UL ; and 2) C1 /2, C2 /2, L/2, 2 R, controlled ac currents as the sampling time T increases from
and UL /2 (comparatively to the nominal values of C1 , C2 , L, 28 s to 1 ms. The THD of ac currents increased from 1% to
R, and UL ). The experimental results (Fig. 10) show that, even 10%. During these experiments, the capacitor voltages of the
with these predictive parameter strong mismatches, the ac cur- multilevel converter were always balanced.
rents follow their references, with mostly the same performance The sliding mode controller balances the capacitor volt-
obtained when using the nominal parameters (Fig. 5), and the ages using only the redundant vectors {(2, 15), (4, 17), (5, 18),
capacitor voltages are also balanced (Figs. 8 and 11). This (10, 23), (11, 24), (13, 26)} [11]. Thus, if these vectors are
indicates that this optimal predictive controller also presents not selected, during enough time, from the current errors, the
some robustness to industrial parameter tolerances. capacitor voltage can become unbalanced. Fig. 12(a) shows a
BARROS AND SILVA: PREDICTIVE CONTROL OF NPC MULTILEVEL CONVERTER 3679

Fig. 12. Sinusoidal ac currents i1 and i2 (vertical, 12 A/div) together with capacitor voltages UC1 and UC2 (vertical, 10 V/div) with a dc voltage having 80%
of its rated value. The time division is 10 ms/div. (a) Sliding mode controller experiments. (b) Optimal controller experiments.

Fig. 13. Experimental results of the nonlinear currents iR1 , iR2 , and iR3 , and the ac line currents iL1 , iL2 , and iL3 , of the active power filter. iR1 and iL1 have
a displacement of 2 divisions, and iR3 and iL3 have a displacement of 2 divisions (vertical, 12 A/div; horizontal, 10 ms/div). (a) Nonlinear current experiments.
(b) AC line current experiments.

situation where the sliding mode is not able to balance the

capacitor voltage. In this case, a simple resistive load (4 ) is
connected as the electrical network (Fig. 1), and the dc voltage
is at 80% of its rated value. The application of nonredundant
vectors with large voltage values results in an unbalanced
capacitor voltage [Fig. 12(a)]. This particular situation can lead
to failure of the power semiconductors. The optimal predictive
controller that can use all the vectors to balance the capacitor
voltages completely solves this problem [Fig. 12(b)]. This
shows that it is not enough to use only the redundant vectors
to balance the capacitor voltages to assure the balancing of
capacitor voltages in all operating conditions.

B. Nonlinear Current Compensation and Power Factor

Correction With the Multilevel APF
These experiments assess the performance of the multilevel Fig. 14. Experimental results of the multilevel output currents i1 , i2 , and i3
APF. The shunt-connected multilevel APF (Fig. 1) is controlled of the active power filter. i1 has a displacement of 2 divisions, and i3 has a
to compensate the current harmonics and regulate the power displacement of 2 divisions (vertical, 12 A/div; horizontal, 10 ms/div).
factor (to be nearly unity). The current references of the APF
are generated to regulate the mean value of the dc voltage The ac line current references iLRef and iLRef are the in-
UdcRef = 240 V, the power factor, and the ac line currents. puts of the optimal predictive controller (Fig. 3), which chooses
3680 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 10, OCTOBER 2008

V. C ONCLUSION
This paper has presented an optimal predictive control
method for the NPC multilevel converter, working as an APF
for power quality applications, and has compared the current
controller results with those of a robust sliding mode controller.
Experimental results obtained show that the robustness prop-
erty of the sliding mode, which is particularly useful in the
presence of unknown disturbances, does not optimize the ripple
or the THD of the ac currents.
The proposed optimal predictive controllers, which predict in
real time the state space voltages and currents of the multilevel
converter and compute a quadratic cost function to choose the
optimal vector, present ac currents that track their references
with a smaller ripple and no steady-state error.
The capacitor voltages are correctly balanced and almost
indistinguishable, having a much smaller ripple than that of the
sliding mode controller. The predictive optimal vector selection
Fig. 15. Experimental results of the ac line currents iL1 and iL2 (vertical, can reduce the mean value of the switching frequency and
12 A/div) and the ac line voltages UL1 and UL2 (vertical, 40 V/div) of the lower the cost of the microprocessor and sampling devices
active power filter. iL1 and UL1 have a displacement of 2 divisions, and iL2
and UL2 have a displacement of 2 divisions (horizontal, 10 ms/div). comparatively to those of the sliding mode controller for the
same level of the ac current ripple factor. The optimal predictive
controller also presents some robustness to industrial compo-
nent tolerances, even with parameter deviations from +100%
to 50%.
The balancing of the capacitor voltage using redundant vec-
tors is not assured in all operating conditions. Since all vectors
are analyzed for the capacitor voltage balancing, the predic-
tive controller solves the problem of balancing the capacitor
voltages.
Since the optimal predictive controller achieves THD < 1%
and harmonic contents that are 46 dB below the fundamental,
this controller is a valuable addition for applications to enhance
power quality. The application of this controller as an APF
reduces the THD of the ac line current from 36% to 1% with
a near-unity power factor (0.997), significantly improving the
Fig. 16. Experimental results of the dc voltage Udc of the active power filter power quality.
(vertical, 40 V/div; horizontal, 20 ms/div).
R EFERENCES
the optimal vector to minimize the ac line current errors and [1] A. Nabae and I. Takahashi, A new neutral-point-clamped PWM inverter,
balance the capacitor voltages. IEEE Trans. Ind. Appl., vol. IA-17, no. 5, pp. 518523, Sep./Oct. 1981.
The experimental results of the nonlinear currents iR1 , iR2 , [2] T. A. Meynard and H. Foch, Multi-level choppers for high voltage appli-
cations, EPE J., vol. 2, no. 1, pp. 4550, Mar. 1992.
and iR3 [Fig. 13(a)] show that the THD is about 36%, and [3] M. Marchesoni, M. Mazzucchelli, and S. Tenconi, A non conventional
the power factor is 0.91. The multilevel APF injects currents power converter for plasma stabilization, in Proc. PESC, Apr. 1988,
(Fig. 14) to cancel the harmonic components of the nonlinear pp. 122129.
[4] Y. Xiong, D. Chen, X. Yang, C. Hu, and Z. Zhang, Analysis and ex-
currents. The THD of ac line currents is reduced from 36% to perimentation of a new three-phase multilevel current-source inverter, in
nearly 1% [Fig. 13(b)]. Proc. IEEE 35th Annu. PESC, Aachen, Germany, Jun. 2004, pp. 548551.
Fig. 15 simultaneously shows the experimental results of the [5] P. Correa, M. Pacas, and J. Rodrguez, Predictive torque control for
inverter-fed induction machines, IEEE Trans. Ind. Electron., vol. 54,
ac line currents iL1 and iL2 with the ac line voltages UL1 no. 2, pp. 10731079, Apr. 2007.
and UL2 . The results indicate that the ac line voltage and the ac [6] Z. Du, L. M. Tolbert, and J. N. Chiasson, Reduced switching frequency
line currents are in phase, showing that the power factor is near computed PWM method for multilevel converter control, in Proc. IEEE
36th Annu. PESC, Recife, Brazil, Jun. 2005, pp. 25602564.
unity. The power factor was improved from 0.910 (without the [7] A. R. Being, G. Narayanan, and V. T. Ranganathan, Modified SVPWM
APF) to 0.997 (using the APF). algorithm for three level VSI with synchronized and symmetrical wave-
The dc voltage Udc (Fig. 16) shows that the dc voltage forms, IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 486494, Feb. 2007.
[8] J. Rodrguez, S. Kouro, J. Rebolledo, and J. Pontt, A reduced switching
follows its references (UdcRef = 240 V) with almost no steady- frequency modulation algorithm for high power multilevel inverters, in
state error (0.3%). Proc. IEEE 36th Annu. PESC, Recife, Brazil, Jun. 2005, pp. 867872.
The optimization of the multilevel current control improves [9] A. Cataliotti, F. Genduso, A. Raciti, and G. R. Galluzzo, Generalized
PWM-VSI control algorithm based on a universal duty-cycle expression:
the performance of the APF by reducing the THD of ac line Theoretical analysis, simulation results, and experimental validations,
currents. IEEE Trans. Ind. Electron., vol. 54, no. 3, pp. 15691580, Jun. 2007.
BARROS AND SILVA: PREDICTIVE CONTROL OF NPC MULTILEVEL CONVERTER 3681

[10] L. Lin, Y. Zou, Z. Wang, and H. Jin, Modeling and control of neutral- [31] P. Zanchetta, D. B. Gerry, V. G. Monopoli, J. C. Clare, and P. W. Wheeler,
point voltage balancing problem in three-level NPC PWM inverters, in Predictive current control for multilevel active rectifiers with reduced
Proc. IEEE 36th Annu. PESC, Recife, Brazil, Jun. 2005, pp. 861866. switching frequency, IEEE Trans. Ind. Electron., vol. 55, no. 1, pp. 163
[11] J. F. Silva, N. Rodrigues, and J. Costa, Space vector alpha-beta sliding 172, Jan. 2008.
mode current controllers for three-phase multilevel inverters, in Proc. [32] R. Vargas, P. Cortes, U. Ammann, J. Rodriguez, and J. Pontt, Predictive
IEEE 31st Annu. PESC, Jun. 2000, vol. 1, pp. 133138. control of a three-phase neutral-point-clamped inverter, IEEE Trans. Ind.
[12] J. Rodriguez, J.-S. Lai, and F. Z. Peng, Multilevel inverters: A survey Electron., vol. 54, no. 5, pp. 26972705, Oct. 2007.
of topologies, controls, and applications, IEEE Trans. Ind. Electron., [33] B. Singh, V. Verma, and J. Solanki, Neutral network-based selective
vol. 49, no. 4, pp. 724738, Aug. 2002. compensation of current quality problems in distribution system, IEEE
[13] S. Ceballos, J. Pou, E. Robles, I. Gabiola, J. Zaragoza, J. L. Villate, and Trans. Ind. Electron., vol. 54, no. 1, pp. 5360, Feb. 2007.
D. Boroyevich, Three-level converter topologies with switch breakdown [34] S. Gri, R. Cardoner, R. Costa-Castell, and E. Fossas, Digital repet-
fault-tolerance capability, IEEE Trans. Ind. Electron., vol. 55, no. 3, itive control of a three-phase four-wire shunt active filter, IEEE Trans.
pp. 982995, Mar. 2008. Ind. Electron., vol. 54, no. 3, pp. 14951503, Jun. 2007.
[14] J. Rodrguez, S. Bernet, B. Wu, J. O. Pontt, and S. Kouro, Multi- [35] Y. W. Li, F. Blaabjerg, D. M. Vilathgamuwa, and P. C. Loh, Design and
level voltage-source-converter topologies for industrial medium-voltage comparison of high performance stationary-frame controllers for DVR
drives, IEEE Trans. Ind. Electron., vol. 54, no. 6, pp. 29302945, implementation, IEEE Trans. Power Electron., vol. 22, no. 2, pp. 602
Dec. 2007. 612, Mar. 2007.
[15] G. S. Perantzakis, F. H. Xepapas, S. A. Papathanassiou, and S. N. Manias, [36] L. Encarnao, J. F. Silva, and V. Soares, Reactive power compensation
A predictive current control technique for three-level NPC voltage source using sliding mode controlled three-phase multilevel converters, in Proc.
inverters, in Proc. IEEE 36th Annu. PESC, Recife, Brazil, Jun. 2005, IEEE Power Eng. Soc. ICHQP, Cascais, Portugal, Oct. 2006, pp. 17.
pp. 12411246. [37] K.-K. Shyu, M.-J. Yang, Y.-M. Chen, and Y.-F. Lin, Model reference
[16] O. Ojo and S. Konduru, A discontinuous carrier-based PWM modulation adaptive control design for shunt active-power-filter system, IEEE Trans.
method for the control of neutral point voltage of three-phase three-level Ind. Electron., vol. 55, no. 1, pp. 97106, Jan. 2008.
diode clamped converters, in Proc. IEEE 36th Annu. PESC, Recife, [38] J. F. Silva, Sliding-mode control of boost-type unity-power-factor PWM
Brazil, Jun. 2005, pp. 16521658. rectifiers, IEEE Trans. Ind. Electron., vol. 46, no. 3, pp. 594603,
[17] D. Xu and B. Wu, Multilevel current source inverters with phase-shifted Jun. 1999.
trapezoidal PWM, in Proc. IEEE 36th Annu. PESC, Recife, Brazil,
Jun. 2005, pp. 25402546.
[18] B. P. McGrath, D. G. Holmes, and T. Lipo, Optimized space vec-
tor switching sequences for multilevel inverters, IEEE Trans. Power J. Dionsio Barros (S04) received the Dipl.Ing. de-
Electron., vol. 18, no. 6, pp. 12931301, Nov. 2003. gree in systems and computer engineering from the
[19] D. G. Holmes and T. A. Lipo, Pulse Width Modulation for Power University of Madeira, Funchal, Portugal, in 1998,
Converters. Hoboken, NJ: Wiley, 2003. and the M.Sc. degree in electrical and computer
[20] M. P. Kazmierkowski and L. Malesani, Current control techniques for engineering and the Ph.D. degree in electrical and
three-phase voltage-source PWM converters: A survey, IEEE Trans. Ind. computer engineering from the Instituto Superior
Electron., vol. 45, no. 5, pp. 691703, Oct. 1998. Tcnico, Technical University of Lisbon, Lisbon,
[21] J. Rodrguez, J. Pontt, C. A. Silva, P. Correa, P. Lezana, P. Corts, and Portugal, in 2002 and 2008, respectively.
U. Ammann, Predictive current control of voltage source inverter, IEEE He is currently an Assistant with the Depart-
Trans. Ind. Electron., vol. 54, no. 1, pp. 495503, Feb. 2007. ment of Mathematics and Engineering, University
[22] S. A. Larrinaga, M. A. R. Vidal, E. Oyarbide, and J. R. T. Apraiz, Predic- of Madeira, and a Researcher with the Center for
tive control strategy for DC/AC converters based on direct power control, Innovation in Electrical and Energy Engineering. His main interests include
IEEE Trans. Ind. Electron., vol. 54, no. 3, pp. 12611271, Jun. 2007. modeling, simulation, and control of multilevel converters applied to power
[23] H. Kwakernaak and R. Sivan, Linear Optimal Control Systems. quality.
New York: Wiley, 1972.
[24] J. D. Barros and J. F. Silva, Optimal predictive control of three-
phase NPC multilevel inverter: Comparison to robust sliding mode con- J. Fernando Silva (M92SM00) was born in
troller, in Proc. IEEE 38th Annu. PESC, Orlando, FL, Jun. 1721, 2007, Mono, Portugal, in 1956. He received the
pp. 20612067. Dipl. Ing. degree in electrical engineering, the Ph.D.
[25] D. Barros and J. F. Silva, Sliding mode assisted predictive pseudo- degree in electrical and computer engineering (EEC),
optimal control for three-phase three-level converters, in Proc. 10th Int. and the Habilitation degree in EEC from the In-
Conf. OPTIM, May 2006, vol. 2, pp. 18. stituto Superior Tcnico, Technical University of
[26] D. Barros, J. F. Silva, and S. F. Pinto, Real-time optimal controller Lisbon, Lisbon, Portugal, in 1980, 1990, and 2002,
for four legged multilevel converters, in Proc. 1st Int. CEE, Coimbra, respectively.
Portugal, Oct. 1012, 2005. He is currently the Vice-President of the Depart-
[27] P. Corts, J. Rodrguez, D. E. Quevedo, and C. Silva, Predictive current ment of EEC (DEEC) and an Associate Professor
control strategy with imposed load current spectrum, IEEE Trans. Power of power electronics in the Energy Group of DEEC,
Electron., vol. 23, no. 2, pp. 612618, Mar. 2008. Technical University of Lisbon. He teaches power electronics, control of
[28] J. F. Silva, Controlo preditivo por simulao para rectificadores com switching power converter systems, and power quality. As the Leader of the
eliminao de falhas de comutao, Ph.D. dissertation, Electr. Comput. Power Electronics and Power Quality group of the Center for Innovation
Eng., IST, TULisbon, Lisboa, Portugal, 1989. in Electrical and Energy Engineering, his main research interests include
[29] J. F. Silva, B. Borges, and A. A. Anunciada, Improving control strategies modeling, simulation, topologies, and advanced control of power electronics
for HF resonant link converters: The current mode predictive modulator, systems and power quality. He supervised seven Ph.D. theses and is currently
in Proc. IEEE PESC, Boston, MA, Jun. 1991, pp. 268275. cosupervising four Ph.D. students. He has authored two books (in Portuguese)
[30] J. F. Silva, Detailed model predictive control for a 3 phase and more than 150 papers in international journals, books, and conferences with
thyristor current source, in Proc. IMACS-MCTS, Moudni, Borne, and reviewers.
Tzafestas, Eds., 1991, vol. I, pp. 236241. Dr. Silva is a member of the Ordem dos Engenheiros, Portugal.