Model Predictive Current Control for High-Power

Grid-Connected Converters with Output LCL Filter

Hernan Miranda, Remus Teodorescu
Department of Energy Technology
Aalborg University
Aalborg, Denmark
Email: [email protected]
Pedro Rodriguez
Department of Electrical Engineering
Universitat Politecnica de Catalunya
Barcelona, Spain
Email: [email protected]
Lars Helle
Vestas Wind Systems
Denmark
Email:[email protected]
AbstractA model predictive control strategy for a high-
power, grid connected 3-level neutral clamped point converter
is presented. Power losses constraints set a limit on commutation
losses so reduced switching frequency is required, thus producing
low frequency current harmonics. To reduce these harmonics an
LCL lter is used. The proposed control strategy allows control
of the active and reactive power fed into the grid, reduce the
switching frequency within acceptable operational margins and
keep balance of the DC-link capacitor voltages while avoiding
excitation of the lter resonance frequencies.
I. INTRODUCTION
Installation of wind power plants (WPP) have experienced
a dramatic growth during the last years and it is expected that
the total power installed will double in a few years [1]. Tied to
this growth the power delivered by each wind turbine generator
(WTG) has also increased. It is expected that connection with
the grid through a full-scale converter will be the preferred
technical solution since this allows decoupled control of active
and reactive power (even during faults), and it is easier to
control negative sequence components.
In order to keep power losses within acceptable margins,
high-power converters (around 5 MW) present a critical lim-
itation on the switching frequency. Available power semicon-
ductors are usually limited to operate below 600 HZ. For a
3-level neutral point connected (NPC) converter this means
that the switching frequency as seen in the line current is
1200 Hz. Such low switching frequency will produce side
bands harmonics that cannot be limited by a pure inductive line
lter. In order to comply with harmonic injection standards,
an inductive-capacitive-inductive (LCL) lter is preferred.
In recent years, some research has been developed in LCL
lter design [2]-[5] and current control of LCL connected
converters [2], [6] and [7]. In [2] and [6], linear controllers
are used to damp the lter resonances, while in [7] state
feedback is realized through a variable feedback gain, which
is calculated depending on the state position within the state
space plane, thus resulting in a piecewise linear controller.
However, this solutions are all oriented to high frequency
switching strategies (above 3 kHz).
The proposed strategy is based on model predictive control
as presented in [8] and [9]. The strategy makes use of a discrete
time model of the system to predict the future value of the
control state for different control inputs. Each predicted state is
introduced in a cost function containing the control objective,
and the input producing the state with the lowest cost is applied
during the following sampling instant. In this way, an on-line
optimization control algorithm is obtained. This is particularly
suitable for a 3-level NPC converter since it is possible to
predict the system behavior for each of the 27 voltage vectors
available.
The cost function to be optimized includes four differ-
ent control objectives. The rst one is fundamental current
tracking in order to deliver the required active and reactive
power into the grid. The second objective is to reduce the
switching frequency to within the limits imposed by the power
semiconductors (below 1200 Hz). The third objective is to
keep the DC-link capacitor voltages balanced, and the last
control objective is to avoid excitation of the LCL lter
resonance frequencies.
In section II the mathematical model required for each
control objective is presented. The cost function associated to
the control objectives is developed in section III. Active power
control regulation is explained in section IV and simulations
results for a 6MW, 3kV line voltage 3-level NPC converter
with 5kV DC-link voltage are presented.
II. SYSTEM MODELING
The system under study is shown in g. 1 and can be de-
composed in ve subsystems. Two of them: an external power
source injecting active power, and the grid voltage V
g
, can
be considered as external disturbances since they can not be
affected by internal variables and present a random behavior.
The remaining three subsystems are: the DC-link lter with
capacitances C1 and C2, an 3-level NPC converter and the
LCL grid lter with converter side resistance and inductance
R
1
and L
1
, capacitor lter C and grid side resistance and
inductance R
2
and L
2
. These subsystems are characterized by
variables that must comply with certain performance indexes
e.g. capacitor voltage balance (DC-link), switching frequency
(power converter) or output harmonic content (LCL lter). In
order to implement a control algorithm able to regulate these
and other variables of interest a model of each of the system
components must be developed.
2009 630 IreprInt oI CO 2009 IroceedIngs
Figure 1. System conguration.
A. DC-link model
According to g. 1 the currents owing into the DC-link
lter are I
+
in
and I

in
(superindex + for the positive bar, and
for the negative), and the currents going out of the DC-link
are I
+
out
and I

out
. According to these denitions the capacitor
voltages V
C1
and V
C2
are dened in discrete-time as in (1)
and (2).
V
C1
[k+1] = V
C1
[k] +
t
C1

I
+
in
I
+
out

(1)
V
C2
[k+1] = V
C2
[k] +
t
C2

in
+I

out

(2)
where t is the sampling period.
B. NPC converter model
Power converter output variables are three phase currents
i
1a
, i
1b
and i
1c
and phase voltages from the DC-link medium
point N to outputs A, B and C. These currents and voltages
can be represented by two phase vectors in a complex
frame. Using (3) the vectors

i
1
and v
1
are obtained.
x =
2
3

1 1/2 1/2
0

3/2

3/2

x
a
x
b
x
c

(3)
A total of 27 different switching combinations are available
in a 3-level NPC converter, providing 19 voltage vectors at
the converter output. A vector diagram of the output voltage
vectors is shown in g. 2. Switching combinations producing
redundant voltage vectors will be used for DC-link capacitor
voltage balance.
C. LCL lter model
A per phase model of the output lter is shown in g. 3. A
state space variable model will be used with space variables
being the converter side current

i
1
, the grid side current

i
2
and the lter capacitor voltage v
c
. From electric equations the
Figure 2. Voltage vectors available in a 3-level NPC converter
Figure 3. LCL lter per phase equivalent circuit
lter is described by the state space model (4).

i
1
dt
d

i
2
dt
dv
c
dt

R
1
L
1
0
1
L
1
0
R
2
L
2
1
L
2
1
C

1
C
0

i
1

i
2
v
c

1
L
1
0
0
1
L
2

v
1
v
2

(4)
2009 631 IreprInt oI CO 2009 IroceedIngs
Figure 4. Filter admittance frequency response.
The system (4) has as inputs the voltage v
1
corresponding
to the voltage vector applied at the converter terminals and
the grid voltage vector v
g
which is considered as an external
disturbance.
From the model (4) it is possible to obtain the frequency
response between the input voltage v
1
to the input and output
currents

i
1
and

i
2
, as well as from the input current to the
output current.
A shown in g. 4 in both frequency responses from v
1
to

i
1
and

i
2
a resonant behavior is present. This means that any
harmonic component close to the resonant frequency will be
greatly amplied. The same resonant characteristic is observed
between the input current

i
1
and the output current

i
2
III. MODEL PREDICTIVE CURRENT CONTROL
The control algorithm is based on a discrete-time equivalent
model of (4). This discrete-time model is able to predict
the future (time k + 1) state of the system based on actual
measurements (at time k) and the converter voltage to be
applied. For each voltage vector available in a 3-level NPC
converter, the future values of converter and grid side currents
are predicted. These predicted values will be used as inputs for
a cost function G and the voltage vector producing the lowest
value of the cost function is selected to be applied during the
next sampling instant.
A. DC-link capacitors voltage balancing
According to [10] it is possible to obtain the current owing
out of the capacitors if the output currents i
1a
, i
1b
and i
1c
,
and the switching state of the converter are known. From the
predictive model it is possible to obtain the future values of
the current vector

i
1
and thus, the three phase currents. From
these currents and switching state of the converter, which is
different for each voltage vector available bur always known,
it is possible to predict the currents owing out of the DC-link
I
+
out
and I

out
. Then, using (1) and (2) the future voltage of
the DC-link capacitors is predicted.
Figure 5. Frequency response of the resonance damping lters.
The cost function associated with the voltage unbalance is
dened as in (5).
G
unb
= (V
C1
[k+1] V
C2
[k+1])
2
(5)
B. Switching frequency reduction
High power converters need to work with a reduced switch-
ing frequency in order to have lower switching losses. Because
of this, a term has been added to the cost function G that gives
a higher costs to states producing more commutations of the
power semiconductors.
The total number of commutations is counted for the
last fundamental period (usually 20ms). Commutations taking
place within the last fundamental period are added to the
counter thus increasing the cost, while switchings that occur
more than a period ago are discarded. The total number of
switch commutations in the last fundamental period N
sw
is
used as part of the cost function (6).
G
SW
= N
SW
[k+1]
2
(6)
C. Current reference tracking and lter resonance damping
The main objective of the power converter is to be able to
inject the required currents into the grid and in this way deliver
a certain amount of active and reactive power. For this purpose
a term on the cost function has been added that assigns a cost
to the difference between the output current reference vector
(in coordinates)

i
2

and the predicted current vector

i
2
[k+1]. The associated cost to the tracking error G
i
2
is given
by (7)
G
i
2
=

i
2

i
2
[k+1]

2
(7)
As explained in section II-C, an LCL lter presents an
intrinsic resonant behavior. The rst resonance is produced
at frequency
1
and is related to parallel impedance between
the lter capacitor C and the grid side inductance L
2
. The
second resonance occurs at frequency
2
due to the parallel
2009 632 IreprInt oI CO 2009 IroceedIngs
Figure 6. Control diagram.
connection of L
1
, C and L
2
. The resonant frequencies are
given by (8) and (9)

1
=
1

CL
2
(8)

2
=
1

C
L
1
L
2
L
1
+L
2
(9)
To avoid exciting the resonances it is necessary to provide
a control algorithm that does not put signicant energy into
these two frequencies. Two achieve this it is possible to add
a cost term to the cost function G considering the distortion
generated by the resonance in

i
1
and to modify the cost G
i
2
in order to consider the energy put in frequencies close to the
resonance.
Based on [11], where frequency dependant weight functions
have been used to obtain a pulse-width-modulation like current
spectrum; to avoid any resonance in the range of
1
it is
possible to add a term G
i
1
. This term has a frequency response
with high gain near the resonance frequency, thus adding a
high cost if the current

i
1
has harmonic components close to
the lter resonance frequency. To do this a rst-order, band-
pass lter tuned at
1
is used to lter the predicted value
of the converter side current

i
1
[k+1]. In this way the cost
of voltages producing harmonic content nearby the resonant
frequency, will be high enough to avoid their selection for the
next sampling period.
On the other hand, it is necessary to include a stop-band
lter in the range of the fundamental frequency, since the
control strategy is supposed to put energy there, to avoid any
extra cost in the reference tracking of

i
2
.
The cost functions associated to the resonance near
1
is
given by (10)
G
i
1
= W
1

i
2
1

+i
2
1

(10)
where W
1
is the lter shaping the frequency components of

i
1
.
A similar approach is used to damp the resonance in
2
.
A rst-order pass-band lter is used in G
i
2
to ensure that
no harmonics are present near the resonance
2
. The lter
is centered exactly at the resonance frequency. Also a pass-
band lter has been added to the predicted current and to
the reference signal at the fundamental frequency to enhance
current tracking. With all this the cost function G
i
2
is given
by (11).
G
i
2
=

W
2
i
2
[k+1] W
r
i

2
+

W
2
i
2

[k+1] W
r
i

2
(11)
To avoid any ambiguity, i.e. low cost for signals with energy
near
2
for G
i
1
and high cost for G
i
2
, the same digital lter
used for

i
2
is used in W
1
. In this way the nal lter used for

i
1
has one peak at
1
and another peak in the vicinity of
2
and a very low cost at the fundamental frequency. In g. 5
the frequency response of the digital lters W
1
, W
2
and W
r
is shown.
Finally, the cost function to be evaluated for each of the 27
voltage vectors is given by (12).
G = K
unb
G
unb
+K
SW
G
SW
+K
i
1
G
i
1
+K
i
2
G
i
2
(12)
Where K
unb
, K
SW
, K
i
1
and K
i
2
are weight gains to change
the relevance of each term within the cost function. The
voltage vector producing the lowest cost is selected to be
applied during the next sampling instant.
IV. ACTIVE POWER CONTROL
DC-link voltage regulation is performed by an outer control
loop. A proportional-integral (PI) controller produces an active
current reference signal i
2d
which is proportional to the output
active power. Based in the transfer function (13) the control
loop is designed. A low bandwidth closed loop is preferred to
avoid resonances and low frequency harmonics due to voltage
ripple.
G(s) =
3
2
|V
g
|
V
DC

1
(C
1
+C
2
)s
(13)
The active current reference and an external reactive current
reference signals are rotated to the plane. The resulting
i
2
and i
2
are used as references for the model predictive
controller. A block diagram of the complete control system
is shown in g. 6. Where the active power controller C(s)
provides the active current reference signal which together
with an arbitrary reactive current reference signal are rotated
to coordinates. The current reference vector in stationary
coordinates is given to the model predictive control algorithm,
represented by the MPC block. The algorithm produces the
NPC converter ring pulses corresponding to the optimum
voltage vector.
2009 633 IreprInt oI CO 2009 IroceedIngs
(a) Active power.
(b) Reactive power.
Figure 7. Output power regulation.
V. SIMULATION RESULTS
Simulations were developed considering a 6 MW 3-level
NPC converter connected to a 3kV grid. Output lter param-
eters are given in table I. The gird side inductance considers
a 6.11% grid impedance and and 9.77% transformer series
inductance. The series impedance 2 50(L
1
+L
2
) is around
32% of the base impedance. The LCL lter has resonance
frequencies in
1
= 2 181 and
2
= 2 330 [rad/s]. The
selected values produce resonances of 47.5 and 49.3 dB for

1
and
2
, any harmonic in the resonant frequencies will be
amplied by a factor of 237 and 292, respectively.
The Matlab-Simulink simulation includes a model of the
three level converter, LCL lter and control algorithms. The
system was tested by injecting active power into the DC-
link, going from 0 to rated power in 1.5 s. The response of
the control system is shown in g. 7(a). Although the MPC
controller has an intrinsic non-linear characteristic, the power
response is clearly linear because it is dominated by the slow
Table I
OUTPUT FILTER PARAMETERS
Parameter Value p.u.
L
1
481 H 0.1119
R
1
4.95 m 0.0036
L
2
884 H 0.2057
R
2
4.22 m 0.0031
C 528 F 4.464
DC-link voltage control loop.
A reactive power distortion can be observed in g. 7(b) dur-
ing the period the active power is changing. This distortion is
due to the change in the LCL lter reactive power consumption
under different operating points. In time t = 10 s, a change
on reactive power demand from 0 to nominal reactive power
is demanded. The change on reactive current also affects the
response in active power, although it is not so relevant due to
the presence of the outer active power control loop.
Filter performance is tested by comparison of converter and
gride side currents. In g. 8(a) three phase currents into the
LCL lter are shown. The total harmonic distortion (THD)
of the currents delivered by the NPC converter is 13.49%. As
shown in g. 8(b), sinusoidal currents with almost no harmonic
content are injected into the grid. The THD of grid side
currents is 3.74% thus demonstrating the proper performance
of the LCL lter.
Filter resonance damping is analyzed by looking at the
current spectrum of both converter and grid side currents.
Fig. 9(a) shows the converter side spectrum and the frequency
response of the digital lter used to damp the resonance. It is
clear that the control algorithm does not inject harmonics in
the frequencies where the cost is high.
The same behavior is observed in the grid side current
spectrum in g. 9(b) where harmonics in the vicinity digital
lter picks are avoided.
In both converter and grid side current spectrums a low
frequency distortion is appreciated. This distortion is generated
by the DC-link voltage control loop, which continuously
changes the operation point to keep track of the voltage
reference signal.
If no cost is associated with commutation losses i.e. K
SW
=
0, the switching frequency is around 2500 Hz with a standard
deviation of 210 Hz. The switching frequency reduction algo-
rithm effectively reduces the commutation losses by operating
the converter with 1070 Hz of switching frequency with a
standard deviation of 106 Hz.
VI. CONCLUSION
A current control algorithm for high-power grid-connected
converts has been presented. On-line optimization of a cost
function is performed in every sampling period allowing
control of several variables and performance indexes.
The control algorithm also includes a cost on power semi-
conductors commutations. In this way lower switching fre-
quency, close to 1100 Hz, is obtained thus reducing commu-
tation losses. Harmonic content injected into the grid because
2009 634 IreprInt oI CO 2009 IroceedIngs
(a) Converter side current.
(b) Grid side current.
Figure 8. Three phase currents.
of the low switching frequency limitations are reduced by the
use of an LCL lter. Filter resonances excitation is avoided by
the use of frequency dependant cost functions, so no resonant
currents are injected into the grid or the lter capacitor while
fundamental current is controlled to keep track of active and
reactive power reference signals.
REFERENCES
[1] Makarov, Y.V. and Shuai Lu and McManus, B. and Pease, J., The future
impact of wind on bpa power system load following and regulation
requirements, May 2008, pp. 16.
[2] M. Liserre, F. Blaabjerg, and S. Hansen, Design and control of an
lcl-lter-based three-phase active rectier, Industry Applications, IEEE
Transactions on, vol. 41, no. 5, pp. 12811291, Sept.-Oct. 2005.
[3] H. Karshenas and H. Sagha, Basic criteria in designing lcl lters for
grid connected converters, vol. 3, July 2006, pp. 19962000.
[4] M. Borage, K.V. Nagesh, M.S. Bhatia, S. Tiwari, Design of lcl-t
resonant converter including the effect of transformer winding capaci-
tance, Industrial Electronics, IEEE Transactions on, Accepted for future
publication.
[5] K. Jalili, S. Bernet, Design of lcl-lters of active front end two level-
voltage source converters, Industrial Electronics, IEEE Transanctions
on, Accepted for future publication.
(a) Converter side current spectrum.
(b) Grid side current spectrum.
Figure 9. Current spectrum.
[6] A. Papavasiliou, S. Papathanassiou, S. Manias, and G. Demetriadis,
Current control of a voltage source inverter connected to the grid via
lcl lter, June 2007, pp. 23792384.
[7] S. Mariethoz and M. Morari, Explicit model-predictive control of a
pwm inverter with an lcl lter, Industrial Electronics, IEEE Transac-
tions on, vol. 56, no. 2, pp. 389399, Feb. 2009.
[8] P. Cortes, M. Kazmierkowski, R. Kennel, D. Quevedo, and J. Rodriguez,
Predictive control in power electronics and drives, Industrial Electron-
ics, IEEE Transactions on, vol. 55, no. 12, pp. 43124324, Dec. 2008.
[9] J. Rodriguez, J. Pontt, C. A. Silva, P. Correa, P. Lezana, P. Cortes, and
U. Ammann, Predictive current control of a voltage source inverter,
Industrial Electronics, IEEE Transactions on, vol. 54, no. 1, pp. 495
503, Feb. 2007.
[10] N. Celanovic, D. Boroyevich, A comprehensive study of neutral-point
voltage balancing problem in three-level neutral-point-clamped voltage
source pwm inverters, IEEE Trans. on Power Electron., vol. 15, no. 2,
pp. 242249, March 2000.
[11] P. Cortes, J. Rodriguez, D. Quevedo, and C. Silva, Predictive current
control strategy with imposed load current spectrum, Power Electronics,
IEEE Transactions on, vol. 23, no. 2, pp. 612618, March 2008.
2009 635 IreprInt oI CO 2009 IroceedIngs