Anda di halaman 1dari 9

36

CHAPTER 3

ARTIFICIAL NEURAL NETWORK BASED OPEN LOOP CONTROL OF


MULTILEVEL INVERTER

3.1 INTRODUCTION

Power electronic circuits used for various applications inject harmonics


into the utility system. This will further reduce the quality of the power and hence
affect the performance of sensitive equipment. In this part of work, an attempt has
been made to minimize the harmonic content in the output voltage of a cascaded
multilevel inverter using selective harmonic elimination pulse width modulation
(SHEPWM) technique. The present work investigates the effectiveness of an
Artificial Neural Network (ANN) for calculating the switching angles for
multilevel inverter. The NN is trained to form a mapping between the inputs
(Modulation Index) and outputs (Switching angles) of a multilevel inverter. The
power circuit of a multilevel inverter is simulated in MATLAB SIMULINK in
open loop mode. Sinusoidal PWM technique is used to generate the switching
signals. The simulation results are presented and analyzed.

3.2 ARTIFICIAL NEURAL NETWORKS

An artificial neural network, usually called "neural network" (NN), is a


mathematical model or computational model that tries to simulate the structure
and/or functional aspects of biological neural networks. It consists of an
interconnected group of artificial neurons and processes information using a
connectionist approach to computation. In most cases, an ANN is an adaptive
system that changes its structure based on external or internal information that
flows through the network during the learning phase. Neural networks are non-
37

linear statistical data modeling tools. They can be used to model complex
relationships between inputs and outputs or to find patterns in data. A neural
network is an interconnected group of nodes, akin to the vast network of neurons in
the human brain.
3.2.1 Structure of feed forward network

The feed forward neural network is shown in figure 3.1. The feed
forward neural network was the first and arguably the simplest type of artificial
neural network devised. In this network, the information moves in only one
direction, forward, from the input nodes, through the hidden nodes and to the
output nodes.

Figure 3.1 Structure of feed forward neural network

3.2.2 Training of Neural Network

The expression for output voltage is obtained from the Fourier analysis
which is derived in the chapter 2 is shown equation 3.1.

(3.1)
The fundamental output voltage equation is given in expression (3.2).
(3.2)
38

The resulting harmonic equations are,

(3.3)

(3.4)

(3.5)

Where is modulation index (M).

The switching angles are found by solving the simultaneous equations


(3.3),(3.4) and (3.5) using the resultant theory, which are used for training the feed
forward neural network.

3.2.3 Introduction to fundamental switching frequency

Each H-bridge inverter unit has a conduction angle, which is calculated


to minimize the harmonic components. The conduction angles are the factors
determining the amplitude of the harmonic component. The output voltage of the
single phase cascaded nine level inverter is a stepped wave form. There are nine
levels in one cycle of the output voltage waveform including zero. The switching
angles θ1, θ2, θ3 and θ4 can be chosen such that the total harmonic distortion is
minimum. Normally, these angles are chosen so as to cancel the predominant
higher order harmonics. For nine-level inverters, the fifth and seventh order
harmonics can be reduced with the appropriate choice of switching angles. One
degree of freedom is used, so that, the shape of the output voltage waveform is
symmetry. To get symmetry, all the levels of the output voltage are maintained for
one half cycle.
Neural network coding in MATLAB is used to determine the switching
angles for Selective Harmonic Elimination PWM (SHEPWM) technique in
cascaded multilevel inverter. Such switching angles are defined by a set of
nonlinear equations to be solved. In the case of two possible solutions for an angle
θi, the criteria for selecting one of them can be the THD. The best angle values are
39

therefore the ones leading to the lowest THD. The flow chart representation of
neural network coding is as shown in figure.3.2.

Figure 3.2 Flow chart for proposed neural network coding

In many studies, ANNs have gained increasing popularity and have


demonstrated superior results compared to alternative methods. Indeed, ANNs are
able to map the underlying relationship between input and output data without prior
understanding of the process under investigation. This mapping is achieved by
adjusting their internal parameters called weights from data. This process is called
the learning or the training process. Their interest comes also from their
generalization capabilities, i.e., their ability to deliver estimated responses to inputs
that were not seen during training. Hence, the application of ANNs to complex
40

relationships and processes makes them highly attractive for different types of
modern problems.

Figure 3.3 Modulation index Vs corresponding switching angles

Neural network is used to learn the switching angles previously provided


by the resultant theory method. The number of inputs and outputs depend on the
considered process.

Figure 3.4 Neural network simulation report

In this application, the feed forward neural network has to map the
underlying relationship between the modulation index and the switching angles as
41

shown in figure 3.3. Figures 3.4 and 3.5 show the performance of the neural
network coding executed in MATLAB. The training performance of the neural
network is evaluated by mean squared error method.

Figure 3.5 Training performance of NN

The figure 3.6 shows that the relationship between modulation index and
its THD for their corresponding switching angles. The neural network is trained for
1000 epochs for the given set of inputs and targets.

Figure 3.6 Modulation index Vs THD

The switching angles for different modulation index and their


corresponding THD values are shown in Table 3.1. From this table, it has been
inferred that the modulation index increases, which results in reduction of THD
values.
42

Table 3.1 Switching angles for various modulation index and their
respective THD values

Switching Angles in
degree
Modulation
Index %THD
θ1 θ2 θ3 θ4

0.8 21.67 35.24 48.16 55.00 4.21


0.81 18.54 30.89 42.59 49.89 4.19
0.82 16.62 28.22 39.20 46.77 4.08
0.83 15.58 26.79 37.37 45.08 3.94
0.84 15.06 26.07 36.45 44.24 3.92
0.85 14.81 25.72 36.00 43.83 3.83
0.86 14.69 25.55 35.79 43.63 3.85
0.87 14.64 25.47 35.69 43.54 3.78
0.88 14.61 25.44 35.64 43.50 3.75
0.89 14.60 25.42 35.62 43.48 3.73
0.9 14.59 25.41 35.61 43.47 3.61
0.91 14.59 25.41 35.60 43.46 3.53

3.2.4 Simulation results

In view to confirm the validity of the obtained results using the harmonics
elimination technique with neural network, a test is carried out for modulation
index values ranging from 0.8 to 0.91. The switching angles for test data are
obtained from the neural network as its output. From that switching angle, the
desired levels of carrier signal are obtained. The triangular carrier signal is
compared with sinusoidal reference which results in the triggering pulses for
various switches.
43

Figure 3.7 Output voltage and current waveforms

These pulses are used to trigger switches in the nine level cascaded
multilevel inverter. The output voltage and current waveforms and its harmonics
spectrum are shown in Figures 3.7 and 3.8, that It is found from figure 3.8, the total
harmonic distortion is only 3.53% for the modulation index of 0.91.

Figure 3.8 Frequency spectrum of the output voltage

3.2.5 Conclusion

A complete analysis of the nine level cascaded multilevel inverter has been
presented in open loop mode. The neural network approach is based on the learning
and approximation of the relationship between the modulation index and the
switching angles with a feed forward network. The modulation index can be varied
44

from 0.775 to 1, but the test is carried out only for the range from 0.8 to 0.91 to
avoid reduction of level in output voltage. The THD is compared for open loop.
The result obtained from neural network has better response than the open loop
simulation without neural network. In this method of calculation, the THD is
reduced to 1/3rd of the SPWM based simulation without including NN based
calculation as discussed in chapter 2.

The resulting neural network implementation of the harmonic elimination


strategy uses very few computational costs, high performance and technical
advantages of the harmonic elimination strategy for the control of nine level
cascaded multilevel inverter.

Anda mungkin juga menyukai