Page 1
International Journal of Computer Applications (0975 – 8887)
Volume 11– No.6, December 2010
17
Neural Network Controller for Asymmetric Cascaded
Multilevel Inverter
Bambang Sujanarko
Dept. of Elect. Eng., Universitas
Jember, currently toward Doctor in
Institut Teknologi Sepuluh Nopember
(ITS) Surabaya, Indonesia
ABSTRACT
Among the Artificial Intelligent (AI) technique, Neural Network
(NN) is emerging technology that is advancing rapidly and having
impact on many scientific and engineering applications. In this
paper, NN apply on Asymmetric Cascaded Multilevel Inverter
(ACMLI), in order to improve the controller easier to construct,
simpler and better performance. To verify these goals, the system
simulated based on Matlab Simulink after the NN controller build
using mfile program. The simulation results show that the goals
can realized.
Mochamad Ashari
Dept.of Elect. Eng., Institut Teknologi
Sepuluh Nopember (ITS) Surabaya,
Indonesia
Mauridhi Hery Purnomo
Dept.of Elect. Eng., Institut Teknologi
Sepuluh Nopember (ITS) Surabaya,
Indonesia
2. SYSTEM DESIGN
2.1 NN System
2.1.1 Biological and Artificial Neurons
The structure of biological neuron is shown in Fig. 1. Basically, it
is the processing element in the nervous system of the brain that
receives and combines signals from other similar neurons through
thousands of input paths called dendrites [4],[6]. Each input
signal flowing through dendrite passes through a synapse or
synaptic junction.
Keywords
Neural network, Asymmetric Cascaded Multilevel Inverter,
performance
1. INTRODUCTION
NN belongs to the area of AI like expert systems (ES), fuzzy logic
(FL) and genetic algorithm (GA), but it is a more generic form of
emulated human thinking that capability to memorize, store
knowledge, perceive, learn and take intelligent decision [4][12].
The NN structure consists of artificial neurons, which each order
interconnected. Because these capability to emulate human
thinking, hence NN will be used to overcome the problems in
many scientific and engineering applications.
In this paper, NN use to improve ACMLI controller especially to
make the controller easier to design, simpler, and better
performance if compared to other control of ACMLI, as
electronic circuits controller,
Programmable Gate Array (FPGA), or computer program within
certain language [1],[45],[8],[10,[13]. NN will be applied for
ACMLI within binary, trinary and sine quantization of DC voltage
progression. This ACMLI within these DC voltage progressions
selected, because its have better performance than others [2],[7].
microcontroller, Field
This application built using the Matlab toolbox, where the NN
structure easy construct using gensim(net) instruction in the m
file, and the system simulation so easy to construct using the
Simulink model. The parameter input for training this NN
determine from reference signals such as DC voltages, sine
reference, and trigger signals that determined from certain
algorithm as explained in the Ref. [3]. To verify the performance,
a comparative study used, among the proposed controller to the
other controllers.
Fig. 1. Structure of Biological neuron
The junction is an infinitesimal gap in the dendrite, which is filled
with neurotransmitter fluid that either accelerates or retards the
flow of the signal. These signals are accumulated in the nucleus
(or soma), nonlinearly modified at the output before flowing to
other neurons through the branches of axon. The adjustment of
the impedance or conductance of the synaptic gap by the
neurotransmitter fluid contributes to the memory or intelligence of
the brain.
Fig. 2. Structure of neuron (artificial neuron)
The model of a neuron (artificial neuron) is shown in Fig. 2
[4],[6]. It has opamp summerlike structure. The neuron has
Page 2
International Journal of Computer Applications (0975 – 8887)
Volume 11– No.6, December 2010
18
some input signal (p), which flows through a gain or weight (w),
which can be positive or negative, integer or noninteger. The
summing node accumulates all the inputweighted signals, adds to
the weighted bias signal (b) (if it’s used), and then passes to the
output through the nonlinear (or linear) activation or transfer
function (TF).
2.1.2 Transfer Function
Transfer Function is Function that maps a neuron’s (or layer’s)
net output n to its actual output. There are many types of transfer
functions [6], such as hardlim (hard limit), pureline (linear),
logsig (log sigmoid), satlins (saturation linear), etc. Fig. 3 shows
some of transfer function graph. Transfer function used depend on
the system to be finished by NN.
Fig. 3. Transfer functions
2.1.3 NN Type
Many NN models have been proposed, such as Perceptron,
Adaline and Madaline, Backpropagation (BP) Network, Radial
Basis Function Network (RBFN), Modular Neural Network
(MNN), Training Vector Quantization (LVQ) Network, Fuzzy
Neural Network (FNN), Kohonen’s SelfOrganizing Feature Map
(SOFM), Adaptive Resonance Theory (ART) Network, Real Time
Recurrent Network, Elman Network, Hopfield Network,
Boltzmann Machine, Recirculation Network, BrainStateInA
Box (BSB), BiDirectional Associative Memory (BAM) Network.
But generally, NN can be classified as feed forward and feedback
types. In the feed forward class, the signals flow only in the
forward direction, whereas in feedback types, the signals can flow
in forward as well as backward or lateral direction. A network can
be defined as static or dynamic, depending on whether it emulates
static or dynamic system. A NN is characterized by input–output
mapping property. Only a few topologies that are most used.
Currently, the back propagation network is most popular [4],[6].
There are many algorithms to determine parameters (weight and
bias) in neural network system. It is very difficult to know which
training algorithm will be the best for a given problem. It depends
on many factors, including the complexity of the problem, the
number of data points in the training set, the number of weights
and biases in the network, the error goal, and others [6],[12].
These algorithms are LevenbergMarquardt (LM), BFGS Quasi
Newton (BFG), Resilient Backpropagation (RP), Scaled
Conjugate Gradient (SCG),
Powell/Beale Restarts (CGB),
Gradient (CGF), PolakRibiére Conjugate Gradient (CGP), One
Step Secant (OSS), Variable Training Rate Backpropagation
(GDX), etc. Among these algorithms, LM algorithm is the fastest
for NN training. It is over four times faster than others [6].
Other classification is learning method, there are two kinds,
namely supervised learning, where there is a target output, and
unsupervised learning, a method, which no output as target.
Conjugate
FletcherPowell
Gradient with
Conjugate
2.1.4 NN Layer
To obtain correct expression among inputs and outputs, a network
can have several layers. Each layer has a weight matrix W, a bias
vector b, and an output vector a. Fig. 4 shows NN in multi layer
configuration. The configuration that similar with this NN will be
used in this paper.
Fig. 4. Multi layer NN
The layers of a multi layer network play different roles [6]. A
layer that produces the network output is called an output layer.
All other layers are called hidden layers. The threelayer network
shown earlier has one output layer (layer 3) and two hidden layers
(layer 1 and layer 2).
Page 3
International Journal of Computer Applications (0975 – 8887)
Volume 11– No.6, December 2010
19
2.2 ACMLI System
2.2.1 ACMLI structure
Recently the multilevel inverter (MLI) is the most popular dc to
ac converters for high voltage and high power in the power
industry. There are three wellknown topologies; diodeclamps,
flying capacitor, and cascaded multilevel inverter (CMLI)
[2],[5],[7]. Diode clamp and flying capacitor are MLI with
common dc sources. These inverters use capacitors in series to
divide the dc bus voltage in to a set of voltage levels. While
CMLI consists of several Hbridge inverter units and separated
DC sources. So CMLI is simply structure inverter because each
Hbridge inverter has the same configuration and can eliminates
the excessively large number of bulky transformers required by
conventional multi pulse inverters, clamping diodes required by
multilevel diode clamped inverters, and flying capacitors required
by multilevel flying capacitor inverters [2].
Fig. 5 shows a singlephase CMLI schematic that consist of
several HBridges. If in each HBridge the switches are S1j, S2j,
S3j, S4j, where j is sequence of HBridges, so there are four
switching combinations in each HBridge, but only three
possibilities of output voltage that occurred that is 0, Vdcj and 
Vdcj. Fig. 5 also shows that if sequence of HBridges (j) until N,
the voltage output of CMLI is obtained by summing of the output
voltages of Hbridge as equation (1).
S11
S21
S31
S41
Vo,1
Vdc,1
+

+

S12
S22
S32
S42
Vo,2
Vdc,2
+

+

S1N
S2N
S3N
S4N
Vo,N
Vdc,N
+

+
HBridge
Cells

Vo
Fig. 5. Singlephase cascaded multilevel inverter
) 1 ()( .....)()()(
, 2 ,
o
1 ,
o
tVtVtVtV
Noo
Based on equation (1) and three output voltages, maximum
number of switches of CMLI (s) show as equation (2). But not all
of voltage levels are the effective switching. These effective of
output voltages usually called voltage levels (n). The effective
switches only produced if the output voltages have different
amplitudes (3) and it’s depend on DC voltage variety.
)3(
)2(3
sn
s
N
2.2.2 DC Voltages Amplitude in ACMLI
CMLI can use equal DC voltages or unequal DC voltages
[2],[5],[7],[11], CMLI that used unequal DC voltages also named
ACMLI. Among the unequal DC voltages for ACMLI, binary
progressions and trinary are the most popular used. In the binary
progression the amplitude of DC voltage have ratio 1: 2: 4: 8 . . :
2N, where N is amount of HBridges in the ACMLI. Similar this
progression, trinary progression has amplitude of DC voltages
within ratio 1: 3: 9: 27. . : 3N.
If the highest voltage of HBridge equal to voltage rating of power
devices or Vdc1 =1 pu, then Vdc2 =1/2 pu,…. Vdcmin = VdcN =1/N2
pu. Trinary progression, also called orde3 [5],[14], have DC
voltages Vdc1 =1 pu, then Vdc2 =1/3 pu, …. Vdcmin = VdcN =1/N3 pu.
Using these DC voltages, the number of voltage levels and the
maximum output voltage binary and trinary DC voltage
progression on ACMLI are shown in (4), (5), (6) and (7).
1
2
n
N
Max
)
5
(
,....,
2 ,
1
,
2
)
V
1
2
j
(
1
,
1
N
j
V
V
V
dcmin
j
dc
dcmin
N
4
( )
n
)
7
(
,....,
1,2
,
3
)
V
2
1
V
1
3
(
3
,
N
j
V
V
dcmin
j
j
dc
dcmin
N
Max
N
(6)
The other progression is sine quantization. Each DC voltage in
this ACMLI can be determined by equation (8), while the
maximum output voltage calculate by equation (9). In this
equation the voltage of sine wave reference is Vm, the frequency
is f, the sequence number of HBridge is j and the number of H
Bridge is N. The process is produced by definition of quantization
that is conversion of a continuous range of values into a number
of discrete levels. Or in the ACMLI context, definition of sine
quantization progression is a sequence of DC voltages that
amplitude follows sine wave function value on discrete times.
)
sin(
t
V
V
)
8
(
6
) 9 (
...,
3
,
2
,
1
)
2
1
N+1
sin(
V
2
)
4
)
/
N+2
1
(
2
sin(
V
2
,
N
j
j
j
f
f
j
m
j
dc
...,3 , 2 , 1)
12
1
sin(2
1
max
Nj
N
jVV
N
2.2.3 Harmonic Minimization
In order to eliminate some harmonic distortion or minimize the
THD, ACMLI need controller to fire power electronics devices in
the optimum angle. One of these methods is THD minimization
[10], [14]. Basic principle of this method shows in fig. 6. To
produce output voltage similar to pure sine, the area in the upper
and lower of sine reference should be zero. By assuming Vj and
Vk are voltage amplitudes adjacent and the area between the sine
wave and the stepping wave are equal, the firing angle is equal to
equation (10), while the others angle are equal to equation (11).
) 10(
2
sin
1
Vk Vj
di
) 11()( sin)(sin
11
Vkdand Vjd
kj
Page 4
International Journal of Computer Applications (0975 – 8887)
Volume 11– No.6, December 2010
20
Fig. 5. Basic principle of THD minimization
2.3 System Configuration
Base on the sub section 2.1 and 2.2, so NN application for
ACMLI controller for singlephase electric systems as discuss in
this paper shown at Fig. 7. ACMLI in this system consist of five
HBridges inverters, within circuit detail as shown in Fig.1. IGBT
power electronic devices use as the switching components. Hence
in this ACMLI consist twenty IGBT with similar type, and five
separated DC sources within amplitude according to the DC
voltage progression.
ACMLINNLoad
DC
Sine
Ref
Fig. 7. System configuration
Input signals that entrance to NN have to normalized, and for
output signal that exit from NN have to denormalized. Amount
layers and neurons t in the each layer determined by optimization,
where the optimum condition occurs if the NN system have little
amount of neuron but have lowest error rate.
3. NN Controller SET UP
3.1 Training Signals
Training signals consists of sinus reference and trigger signals as
output. Sine signal is easy to determine. But to determinate the
trigger signal, some calculation or an algorithm as explained in
Ref. [3] must used. After all of Training signal obtained, the
Training signals must be save in the worksheet of Matlab
programming windows, within file name have according to NN
Training programs. An example of Training signal is shown in
Fig.8.
Training signals usually need to regulate, hence the NN training
due in the optimum data, time, and memories. These regulating
include sampling, limiting, resuming and matrix justification.
If in the ACMLI used N amount of HBridges, so there are 4N+1
of Training signals, where each HBridge need four signal for
each power electronics devices and one sine reference signal for
all HBridges inverter.
Fig. 8. Training signals example
3.2 NN Training
ACMLI within five HBridges used in this paper. Each HBridges
have four power electronics devices that need to trigger. In this
paper the triggering done using in the adding scheme only, so in
the inverter only happen 2N+1 output voltage levels.
Training of NN is done using mfile program in Matlab, with the
urgent instruction of the program as shown in Fig.9. This program
have five step, fill the Training signal, setting the NN structures,
Training with certain parameter, simulation and build NN in the
simulink models.
%Training signal
x=(T6');
y=[T1'; T2'; T3'; T4'; T5'] ;
P=[x(:,:)];
T=y;
%NN setting
net = newff(minmax(P),{5 20},{'purelin', 'logsig', 'logsig' ,
'logsig', 'logsig', 'logsig', 'logsig', 'logsig', 'logsig', 'logsig',
'logsig', 'logsig', 'logsig', 'logsig', 'logsig', 'logsig', 'logsig',
'logsig', 'logsig'});
net.trainParam.lr=0.1;
net.trainParam.epochs= 150;
%NN Training
net.trainParam.lr = 0.6;
….……
net = train(net,P,T);
……
%NN simulation
Y = sim(net,P);
……
%Build NN Simulink
gensim(net);
Fig. 9. mfile of NN Training
Training signal souce is T1, T2, T3, T4, T5 and T6 matrix, where
T6 is an input signal, while others are the output signals. After
this fill these signals, NN structure then setting. In this program
setting is done in the 1520, that means one input within satlins
Page 5
International Journal of Computer Applications (0975 – 8887)
Volume 11– No.6, December 2010
21
transfer function, five hidden layer within satlins transfer
functions and 20 output signals within purelin transfer functions.
The weights and bias of this NN then calculate using Training
step. In this Training, some parameters Training can be set, as
Training rate, maximum epoch, error rate goals, etc.
Simulation step, optional to used, but build NN simulink step is
an urgent step in this program. With using this step, a
sophisticated NN can automatically construct base on gensim(net)
instruction. Performance of training process in the graph format is
shown in Fig. 10. This figure indicate that training have small
error.
Fig. 10. Performance of training process
4. RESULTS AND DISCUSSIONS
Fig. 11 shows ACMLI simulation circuit using NN controller
based on Matlab Simulink. This NN circuit obtains from NN
training. In the NN circuit weight and bias are :
Hidden layer :
Weight :
9.7278
66.4382
56.3512
26.5248
130.5043
Output layer :
Weight :
[3.9873979925178689;
2.4516448217640328;
9.2898968607227328;
6.2620493906409553;
19.049422965101748]
……….
[13.528947259153412;
17.45542716175429;
63.93938459593258;51.47495749
3540209;104.62756360582517]
By using this simulation circuit, output voltage and frequency
spectrum of ACMLI are shown in Fig.12. If this figure compare to
Bias :
46.8819
86.9086
193.7695
111.0933
145.6643
Bias :
1.0378
2.6857
0.0098
………
8.7541
7.4550
5.2099
ACMLI within conventional controller as shown in Fig. 13, hence
we can say that NN controller can applied to ACMLI controller
within comparable performance.
Fig. 11. ACMLI simulation circuit using NN controller
Fig. 12. Output voltage and frequency spectrum of ACMLI usin
NN controller
But these result have different condition, ACMLI within
conventional controller simulate in the sine frequency 50 Hz,
while within NN controller, sine frequency for simulation is 0.005
Hz. These condition done because NN circuit require more time to
solve calculation than conventional control. This problem can
answer if the NN configuration implemented using Field
Programmable Gate Array (FPGA) or others. The advantages of
NN controller for ACMLI such as, it can design more easily, more
flexible, quicker and better performance (depend on capability of
designer) than conventional controller.
5. CONCLUTIONS
This paper show that NN controller can applied for ACMLI
within more easily, more flexible, quicker and better performance
than conventional controller. From simulation this goals can
realized, as indicated by low error rate of training, and high
performance of simulation, although the NN controller still need
implemented on high speed computing component to compensate
its speed.
Page 6
International Journal of Computer Applications (0975 – 8887)
Volume 11– No.6, December 2010
22
Fig. 13. Output voltage and frequency spectrum of ACMLI usin
conventional controller
6. REFERENCES
[1] B. Cheng, F. Huang, ―Multilevel Inverter Control Scheme‖,
US Patent 6,867,987 B2, March, 2005.
[2] B. S. Suh, G. Sinha, M. D. Manjrekar, T. A. Lipo,
―Multilevel Power Conversion – An Overview of Topologies
and Modulation Strategies‖, IEEEOPTIM Conference
Record, pp. 1124, vol. 2, 1998.
[3] Bambang Sujanarko, Mochamad Ashari and Mauridhi Hery
Purnomo, ―Universal Algorithm Control for Asymmetric
Cascaded Multilevel Inverter‖, International Journal of
Computer Applications (0975 – 8887), Vol.8, No.15,
November 2010.
[4] Bimal K. Bose, ―Neural Network Applications in Power
Electronics and Motor Drives – An Introduction and
Prespective‖, IEEE Trans. On Industrial Electronics., Vol.
54, No. 1, pp. 1433, Pebruary 2007.
[5] E. Babaei, S.H. Hosseini, G.B. Gharehpetian, M. Tarafdar
Haquea, M. Sabahi, ―Reduction of dc voltage sources and
switches in asymmetrical multilevel converters using a novel
topology‖, Electric Power Systems Research, 77, 2007, pp.
1073–1085
[6] Howard Demuth, Mark Beale,Martin Hagan. 2008, ―Neural
Network Toolbox™ 6 User’s Guide‖, The MathWorks, Inc.
3 Apple Hill Drive Natick, MA 017602098
[7] J.S. Lai and F.Z. Peng, ―Multilevel converters—A new breed
of power converters‖, IEEE Transactions on Industry
Applications, vol.32, pp. 509–51, May/June, 1996.
[8] Kuhn, H. Ruger, N.E. Mertens, A., ―Control Strategy for
Multilevel Inverter with Nonideal DC Sources‖, Power
Electronics Specialists Conference (PESC), Hanover, 2007.
[9] L. M. Tolbert, John N. Chiasson, Zhong Du, and Keith J.
McKenzie, ―Elimination of Harmonics in a Multilevel
Converter With Nonequal DC Sources‖, IEEE Transactions
On Industry Applications, Vol. 41, No. 1, January/February
2005, pp. 7582.
[10] M. G. Hosseini Aghdam, S. H. Fathi, G. B. Gharehpetian, ―A
Complete Solution of Harmonics Elimination Problem in a
MultiLevel Inverter with Unequal DC Sources‖, Journal of
Electrical Systems, 34, 2007, pp.259271
[11] M. H. Rashid, ―Power Electronics Handbook‖, Academic
Press, 2001, Canada
[12] R. Taleb, A. Meroufel, 2009, ―Control of Asymmetrical
Multilevel Inverter Using Artificial Neural Network‖,
Electronics and Electrical Engineering, No. 8(96), 2009, pp.
9398.
[13] S. J. Park, F. S. Kang, S. E. Cho, C.J. Moond, H. K. Nam,
―A novel switching strategy for improving modularity and
manufacturability of cascadedtransformerbased multilevel
inverters‖, Electric Power Systems Research, 74. 2005, pp.
409–416
[14] S. Krishna, ―Harmonic Elimination by Selection of
Switching Angles and DC Voltages in Cascaded Multilevel
Inverters‖, Fifteenth National Power Systems Conference
(NPSC), IIT Bombay, December 2008.
AUTHOR PROFILE
Bambang Sujanarko received the B.Sc. from Universitas Gadjah
Mada, Yogyakarta Indonesia and Master from Universitas Jember,
Indonesia. He is senior lecture of Departement Electrical
Universitas Jember and currently toward his Ph.D in Institut
Teknologi Sepeluh Nopember (ITS) Surabaya, Indonesia. His
research interests included power electronics and renewable
energy systems, hybrid power systems, artificial intelligent, and
instrumentation.
Mochamad Ashari received the Bachelor degree in electrical
engineering from the Institut Teknologi Sepuluh Nopember(ITS)
Surabaya, Indonesia, in 1989 and Master and Ph.D. from Curtin
University of Technology, Perth, Australia. He has been with ITS
since 1990 as a Lecturer in the Department of Electrical
Engineering. He is a Professor and head of Electrical Engineering
ITS. His research interests include power electronics and inverter
applications, power system modeling, simulation, and analysis of
hybrid power systems.
Mauridhi Hery Purnomo received the B.S. degree from Institut
Teknologi Sepuluh Nopember (ITS) Surabaya, Indonesia and
Master ad Ph.D from Osaka City University, Osaka, Japan. He is
a Professor in the Department of Electrical Engineering, ITS.
Since 2007, he was vice director on ITS postgraduate program.
He has been engaged in research and teaching in the areas of
intelligent system and pattern recognition, power system
simulations, and computer programming.