International Journal of Scientific Engineering and Research (IJSER)

ISSN (Online): 2347-3878

Index Copernicus Value (2015): 62.86 | Impact Factor (2015): 3.791

D-STATCOM for Power Quality Improvement

Trained by ANN
Saravanan A1, Selvakumar G2, Maheswari A3, Muthulakshmi P4, Mohamed Haroon K. H 5
1, 2, 3, 4, 5Department of EEE, PSN College of Engineering and Technology, Tirunelveli, Tamil Nadu, India

Abstract: Power quality has become an important issue since many loads at distribution ends have become intolerant to various power
quality issues. The power quality issues such as voltage sag, voltage swell and harmonics which are certainly major concerning issues
in the present era. These issues can lead to failure of the many sensitive loads connected to the distribution system. To solve this
problem advanced custom power devices are used. One of such device is the Distribution Static Synchronous Compensation (D-
STATCOM). DSTATCOM provides to maintain the desired power factor, voltage and also suppress the harmonic content in
distribution network. The event of power quality is detected by trained artificial neural network then the appropriate event will be
mitigated using neural network aided D-STATCOM. The results show clearly the performance of the Artificial Neural Network based
D-STATCOM in mitigating the voltage sag, harmonics, reactive power compensation and power factor improvement.

Keywords: Distribution Static Compensator (D-STATCOM), Artificial Neural Network (ANN)

1. Introduction Static Compensator (D-STATCOM) is considered to be the

efficient method.
The electric power system is classified in to three division
viz. generation, transmission and distribution. The generation A shunt compensating device having ability to improve the
unit is required to produce adequate amount of power to meet power quality by voltage regulation, load compensation and
the customer’s demand for a reliable power supply system. power factor improvement. The use of various controlling
Transmission systems must transport bulk power generated devices like PI, PID, fuzzy logic, and sliding-mode are in
over long distances without overloading and damaging the practice now [2]. The most common type used in industrial
equipment and distribution systems should deliver the systems till date is Conventional PI based control. Though
electric power to each customer’s premises. Distribution the structure of PI is fixed, it cannot perform in an optimized
system is connected to the customer directly though it is manner under variable load or nonlinear load conditions [3].
located at the end of power, so the power quality mainly This raises a need for the development of new controllers
depends on distribution system. The reason behind this is with more improved performance. In the recent years, for the
that, the failures of electrical distribution network account for applications of power electronics Artificial-intelligence (AI)
about 90% of the average customer interruptions. In the techniques, particularly the NNs, are having a significant
earlier days, the major focus was on generation and impact on it. While maintaining the stability of the converter
transmission for power system reliability. But now a day’s system over a wide operating range, Neural-network- based
more attention for reliability assessment is focused only for controllers provide fast dynamic response and are considered
the distribution systems. Power Quality (PQ) related issues as a new tool to design control circuits for PQ devices [4].
are of most concern now-a-days. The use of power electronic
devices, viz. adjustable speed drives (ASD), converter and The ANN can perform better than conventional PI controller
inverter, choppers, programmable logic controllers (PLC), and it also provides very good basics to build nonlinear
energy-efficient lighting, etc. are lead to a complete change controllers. A lot of research works are going on for
of electric loads and its nature. These are the major causes DSTATCOM on combination with neural network. In this
and the major victims of power quality problems [1]. Due to paper design of ANN based controller for current control and
their use of this equipment, all these loads may cause voltage control of shunt active filter instead of PI controller
disturbances to the system such as transients, short duration is used.
variations (sags, swells, and interruption), long duration
variations (sustained interruptions, under voltages, over Neural networks are composed of numerous elements
operating in parallel, which is, inspired by biological nervous
voltages), voltage imbalance, waveform distortion (dc offset, systems. In general, the network function is determined
harmonics, inter harmonics, notching, and noise), voltage mostly by the connection between elements. To obtain a
fluctuations and power frequency variations and these are particular function neural network is trained by adjusting
generally referred as power quality problems. the value of weights between elements. The neural network is
trained offline using data obtained from conventional PI
For the reasons described above, there is a growing interest controller. A three leg VSI is used to inject or absorb the
in mitigating power quality disturbances by developing newer appropriate voltage through an LC filter and an injection
equipment. Even though, many techniques available for transformer is used to compensate load voltage from the
mitigating various power quality problems, the use of custom distorted supply voltage. The Levenberg - Marquardt Back
power devices like, Dynamic Voltage Regulator (DVR), propagation algorithm is used to implement the control
Unified Power Quality Controller (UPQC) and Distribution scheme of the VSI [5]. Initially the data from PI controller is
Volume 5 Issue 2, February 2017
www.ijser.in
Licensed Under Creative Commons Attribution CC BY
Paper ID: IJSER151233 37 of 42
 International Journal of Scientific Engineering and Research (IJSER)
ISSN (Online): 2347-3878
Index Copernicus Value (2015): 62.86 | Impact Factor (2015): 3.791
stored in workspace. These saved data are trained offline 3. Artificial Neural Network
using ANN.
Neural-networks is one of the new technology that is getting
2. DSTATCOM Model in Synchronous fashionable in the present era. This area of neural networking
Reference Frame is the "fuzziest" in terms of a definite set of rules to abide by.
There are many types of networks - ranging from simple
In distribution systems for more active power supply Boolean networks (perceptions), to complex self- organizing
capability implementation of shunt type of fact devices is networks (Kohonen networks), to networks modelling
preferred when compared to series type devices. Main thermodynamic properties (Boltzmann machines).
objectives of shunt compensator are [2]:
1)The source currents should be sinusoidal and contain only 3.1 Layers of ANN
positive sequence components.
2)The power factor or the voltage magnitude at PCC should The network consists of several "layers" of neurons, an input
be controlled. layer, hidden layers, and output layers. Input layers take the
input and give the data to the hidden layers (so-called hidden
To achieve this objective, the compensator should be able to because the user cannot see the inputs or outputs for those
inject negative, zero sequence fundamental frequency layers). These hidden layers do all the necessary computation
components in addition to harmonic currents of all and give the results to the output layer.
sequences. In DSTATCOM voltage source converter bridge
is connected to a common dc bus. The DC bus voltage is held
by the capacitor Cdc. Since there is no energy source
connected to the DC bus, power exchanged by the
DSTATCOM is zero if switches are ideal; the losses in the
reactors and capacitor are zero. Then the losses are to be
supplied by the active power drawn from the AC system. In
order to stabilize the capacitor voltage in steady state, a DC
bus voltage control is necessary. The power supplied to the
load varies over the line period of fundamental (supply)
frequency, the capacitor will charge and discharge resulting
in a voltage ripple. So that RMS current supplied from the
source to achieve unity factor at PCC is required to
minimize. The reference for the source current vector is first
computed and the desired compensator current is obtained as
the difference between the load and the source (reference
currents). To inject the desired currents, DSTATCOM must
have a high bandwidth closed loop current control loop. The Figure 1: Architecture of ANN [5]
synchronous reference theory is based on the transformation
of the stationary reference frame three phase variables (abc) The figure 1 illustrates the architecture of artificial neural
to synchronous reference frame variables (d, q, 0) whose network (ANN). The ANN controller consists of three
direct (d) and quadrature (q) axes rotate in space at the neuron layers, the input layer, the hidden layer and the output
synchronous speed ωe. layer. The input layer transmits the input signal to the hidden
layer. The latter begins the learning process and the output
The current transformation from abc to d-q-0 frame is layer continues the learning process and provides outputs.
defined as [5],
The designed ANN controller has 3 layers composed of 2
input layers, 10 hidden layers and 2 output layers. It has a 2
input layer obtained from the error voltage at the point of
common coupling (pcc) and the error voltage at the dc link.
The output of ANN controller is the reference variable for
the PWM generator. Therefore, the output of ANN with
The conversion of DQO current to ABC using Park varying amplitude and phase passes through a comparator
transformation: and is compared with a carrier signal. When the ANN
output’s magnitude is more than carrier signal’s magnitude,
the PWM circuit generates high output and when the ANN
output’s magnitude is less than carrier signal’s magnitude, the
PWM circuit produces low output.

3.2 Neural Network Training

The main aim of training of ANN is to finds the suitable

weight values which may cause the desired output. For fast

Volume 5 Issue 2, February 2017

www.ijser.in
Licensed Under Creative Commons Attribution CC BY
Paper ID: IJSER151233 38 of 42
 International Journal of Scientific Engineering and Research (IJSER)
ISSN (Online): 2347-3878
Index Copernicus Value (2015): 62.86 | Impact Factor (2015): 3.791
convergence, back propagation algorithm is used. Select The ANN training is stopped, if the number of maximum
ANN topology with number of layers, nodes and initialize epoch to train is reached or the time is beyond the maximum
with random weights. The input to the neural network is the limit or the performance is minimized to the goal or the
error voltage at the point of common coupling and error performance gradient falls below minimum gradient or initial
voltage at the DC link. Calculate the difference with the blending factor (mu) beyond maximum mu or the
obtained output and actual output which is nothing but an performance of the validation has increased more than the
error. The error is determined and a portion of it is Maximum validation failures. In this work the best validation
propagated backward through the network [6]. performance is 0.00030248 at epoch 61. Now the ANN is
trained offline to control the DSTATCOM. After the
3.3 Levenberg-Marquardt Algorithm completion of offline training the created neural network
fitting toolbox is placed in the replacement of PI controller.
The back- propagation (BP) algorithm is used to train the
most commonly used multilayer perceptron (MLP) model of Then the voltage at the point of common coupling is
neural network which is too slow for practical problems. compared with the reference voltage and dc link voltage is
Some high performance algorithms are available to train compared with the reference voltage is given as the input
MLP models but it is 10 to 100 times faster than BP layer of the ANN. The designed ANN controller has 3 layers
algorithm. These algorithms are based on numerical composed of 2 input layers, 10 hidden layers and 2 output
optimization techniques like Conjugate gradient, quasi layers. After the training is complete the output produced is
Newton and Levenberg Marquardt algorithms (LMA), etc. real and reactive. The aim of the control scheme is to
Out of these, in Levenberg Marquardt algorithm, training maintain constant voltage magnitude at the point where a
process converges quickly. LMA combines the excellent sensitive load is connected, under system disturbances. The
local convergence properties of Gauss-Newton method hence main function of a DSTATCOM control system is to detect
found to be the fastest training algorithm for moderate size the disturbances occurring in the system and compute the
neural networks. missing voltage to generate gate pulses using Discrete PWM
generator, then the IGBT converter converts the input DC
4. Proposed Control Scheme voltage to a sinusoidal AC voltage through an LC filter and
injection transformer. The reactive power depends on the 0 to
4.1 Block Diagram 1’s. If the voltage magnitude is below 1 then it represents the
voltage sag. If the voltage magnitude is above 1 then it
represents voltage swell.

The compensating current injected by the DSTATCOM

system is stopped, only after the absence of the disturbance.
In this work, the inverse park’s transformation is used to
calculate the missing voltage. The dqo to abc transformation
is transformed the dq rotating coordinate system to 3Ф
stationary coordinate system.
Figure 2: Block Diagram
The dq coordinate which is compared with the reference
value, will calculate the disturbance in the dq coordinate and
these can be transformed again to abc coordinate. The Phase
Locked Loop (PLL) measures the system frequency and
gives the phase synchronous angle θ for the dq coordinate
system. The comprehensive result for the above mentioned
controllers are presented to investigate the performance of
controller in the proposed system.

4.2 Simulation

The SIMULINK model of proposed DSTATCOM is shown

in fig. And it is simulated in MATLAB SIMULINK to test
the performance of the control system.

Figure 3: ANN Controller

First, in order to create the neural network fitting tool box

training of neural network is done with number of input and
output data that are stored in the work space which is taken
from the conventional PI controller. The levenberg-
marquardt back propagation algorithm is used for the training
purpose.

Volume 5 Issue 2, February 2017

www.ijser.in
Licensed Under Creative Commons Attribution CC BY
Paper ID: IJSER151233 39 of 42
 International Journal of Scientific Engineering and Research (IJSER)
ISSN (Online): 2347-3878
Index Copernicus Value (2015): 62.86 | Impact Factor (2015): 3.791

Figure 4: Simulation model of test system to measure

voltage sag, detection by Neural Network and mitigation by
DSTATCOM.
Figure 8: Source voltage

Figure 5: Source Voltage

Figure 9: Compensated source voltage and current

Figure 6: Compensated source voltage and current

Figure 10: Simulation model of test system to measure

voltage sag, detection by Neural Network and mitigation by
DSTATCOM

Figure 7: Simulation model of test system to measure

voltage sag, detection by Neural Network and mitigation by Figure 11: Source voltage
DSTATCOM
Volume 5 Issue 2, February 2017
www.ijser.in
Licensed Under Creative Commons Attribution CC BY
Paper ID: IJSER151233 40 of 42
 International Journal of Scientific Engineering and Research (IJSER)
ISSN (Online): 2347-3878
Index Copernicus Value (2015): 62.86 | Impact Factor (2015): 3.791

Table 1: THD Analysis

Source Voltage Source Current
Without DSTATCOM 0.11% 27.24%
With ANN controlled
0.10% 10.72%
DSTATCOM

5. Conclusion
Considering the neural network controlled DSTATCOM, it is
observed that the power quality of the interconnected power
system is improved. The results can be tested for both linear
Figure 12: Source current and nonlinear loads in MATLAB/SIMULINK. The
performance of ANN controlled DSTATCOM is more
effective rather than other conventional controllers. By
adding neural network control to DSTATCOM it reduces the
harmonic content, improve the power factor and also reduce
the voltage sag and maintain the quality of power supply.

References
[1] Bhism Singh and Sabha Raj Arya “Design and control
of a DSTATCOM for power quality improvement using
cross correlation function approach” (2012),
Figure 13: Load current International Journal of Engineering, Science and
Technology Vol. 4, No. 1, 2012, pp. 74-86.
[2] N. Hingorani, “Introducing Custom Power,” IEEE
Spectrum. Vol. 32, no. 6 pp. 41-48,1995
[3] H. Fujita and H. Akagi, ―The unified power quality
conditioner: The integration of series active filter and
shunt active filters, ‖ in Proc. IEEE/ Power Eng. Soc.
Power Electronics Specialist Conf., Jun. 1996, pp. 491–
501.
[4] Vadirajacharya G. Kinhal, “Performance Investigation
of Neural-Network-Based Unified Power-Quality
Conditioner,” IEEE Trans. Power delivery, VOL. 26,
NO. 1, JAN. 2011.
[5] M.Bhanu Siva, M. R. P Reddy, Ch.Rambabu “Power
Quality Improvement of Three-phase four-wire
DSTATCOM with Fuzzy logic Controller” (2011)
International Journal of Computer Science and
Information Technologies, Vol. 2 (5) , 2011, 2273 –
2276.
Figure 14: Simulation model of test system to measure [6] Venkata Suresh and Dr. B. Venkata Prasanth “Design of
voltage sag, detection by Neural Network and mitigation by Neural Network Controlled DSTATCOM Using SRF
DSTATCOM Algorithm to Improve Power quality” (2013),
International Journal of Engineering Research &
Figure 14 shows the performance of ANN Aided Technology (IJERT), Vol. 2 Issue 8, ISSN: 2278-0181.
DSTATCOM with a load having 0.9 p.f lagging. It is [7] Shaik Khaja Gareeb Nawaz, Shaik Hameed “Mitigation
observed that supply voltages and currents are in phase when of Power Quality Problems by Using D-STATCOM”
DSTATCOM is connected. Therefore, designed (2013), International Journal of Recent Advances in
DSTATCOM is able to improve P.F of supply to unity. The Engineering & Technology (IJRAET), ISSN: 2347-
power factor level is improved in load side or utility side by 2812, Vol.1, Iss.3.
using the DSTATCOM. The directly source signal is
supplying through the load to affect the sensitive loads and Author Profile
linear loads. These above problems are overcome by using
DSTATCOM to improve the power factor in load side level. Saravanan A received the Diploma Certificate in
Similarly, it is also reducing the Total Harmonic Distortion Electrical and Electronics Engineering from Board of
(THD) [7]. Technical Education Tamil Nadu in 2007, B.E. degree
in Electrical and Electronics Engineering and M.E.
degree in Power Electronics and Drives from Anna
University Chennai in 2012 and 2015, respectively. During 2008-
Volume 5 Issue 2, February 2017
www.ijser.in
Licensed Under Creative Commons Attribution CC BY
Paper ID: IJSER151233 41 of 42
 International Journal of Scientific Engineering and Research (IJSER)
ISSN (Online): 2347-3878
Index Copernicus Value (2015): 62.86 | Impact Factor (2015): 3.791
2009, he stayed in Technical Assistant Trainee, VSSC, ISRO to
study Integration and Testing of Satellite Launch Vehicle. He now
worked as an Assistant Professor in Electrical and Electronics
Engineering department, PSN College of Engg. & Tech.,
Tirunelveli, Tamil Nadu.

Selvakumar G studying final year B.E. degree in

Electrical and Electronics Engineering in PSN College
of Engineering and Technology, Tirunelveli affiliated
by Anna University Chennai.

Maheswari A studying final year B.E. degree in

Electrical and Electronics Engineering in PSN College
of Engineering and Technology, Tirunelveli affiliated
by Anna University Chennai.

Muthulakshmi P studying final year B.E. degree in

Electrical and Electronics Engineering in PSN College
of Engineering and Technology, Tirunelveli affiliated
by Anna University Chennai.

Mohamed Haroon K. studying final year B.E.

degree in Electrical and Electronics Engineering in
PSN College of Engineering and Technology,
Tirunelveli affiliated by Anna University Chennai.

Volume 5 Issue 2, February 2017

www.ijser.in
Licensed Under Creative Commons Attribution CC BY
Paper ID: IJSER151233 42 of 42