Microprocessors and Microsystems 73 (2020) 102976

Contents lists available at ScienceDirect

Microprocessors and Microsystems

journal homepage: www.elsevier.com/locate/micpro

Optimization of harmonics with active power filter based on ADALINE

neural network
M. Sujith a,∗, S. Padma b
a
IFET College of Engineering, Villupuram, Tamil Nadu, India
b
Sona College of Technology, Salem, Tamil Nadu, India

a r t i c l e i n f o a b s t r a c t

Article history: Mostly the power quality issues in the distribution line system happen due to the presence of harmonics.
Received 23 September 2019 Especially, the nonlinear loads such as power electronic converters, high-speed semi-conducting switches,
Revised 20 December 2019
and solid state drives were the major causes for harmonics in distorted power system signals. Moreover,
Accepted 26 December 2019
the estimation of magnitude and the phase of this harmful harmonic interference are necessary. By tak-
Available online 27 December 2019
ing in to consideration of all the above factors, this paper develops an efficient technique for harmonic
Keywords: estimation and detection of the renewable wind energy resources and elimination of these harmonics
Harmonics estimation will also be done accordingly for getting desired output from wind energy. 8 bit inputs (4 + 4) are col-
Wind energy lected and used to generate the intended input set for ANN training. The proposed work develops an
Neural network Adaptive Linear Neural Network (ADALINE) for the estimation of harmonics which is the novelty of this
Non-linear load work. For making the harmonics content more negligible and to enhance the load power quality, an Ac-
Active power filter
tive Power Filter (APF) is used. The novel control design is developed with a Pulse Width Modulation
Adaptive linear neuron
(PWM) control. In addition, feed forward networks (trained by back propagation algorithm) works like a
hysteresis band comparator. An APF control design is developed with ADALINE network in which the load
and current along with voltage will be analyzed and then the controller will be calculating the control
signal by considering the reference compensation current. Afterwards, the power system is injected with
compensating current. The simulation is carried out with Matlab-Simulink laboratory prototype is de-
veloped with Xilinux 3E Spartan FPGA board to verify the proposed control designs. The proposed work
is compared with exiting method comprising Shunt Active Power Filters (SAPF) with ADALINE for the
performance perspectives. This method was found to be effective in terms of many parameters such as
load voltage, load current, voltage, reactive power, real power and especially THD value than those of the
existing works which are considered.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction situation, a progressive technique of signal processing was essen-

tial for accurate harmonics estimation parameters. However main-
The increased power electronic components caused distortion tenance of this accuracy remained as a tough task in the power
in the transmission line and made the load nonlinear. This caused system. Popular techniques for estimation of harmonics were Fast
voltage flicker, variation, and imbalance in three-phase lines re- Fourier transformation (FFT) of these signals. But applying FFT, the
ceived from wind energy. This resulted in power quality issues and phenomena of leakage, picket-fence, and aliasing effect reduced
harmonics in these signals. The presented harmonics will flow in the accuracy. To overcome these drawbacks, the Recursive Least
the electrical network which affects the process of the component, Square (RLS) and least square (LS) algorithms have been commonly
or else it resulted in damage. Thus, the calculation of system fre- applied in this scenario which would be effective in frequency es-
quency and the variation rates help in monitoring, protection, and timation.
control of electrical power system equipment [1,2]. Especially the The Artificial Neural Network had the ability to deliver an en-
estimation of magnitude and phase of harmful harmonics inference hanced methodology for deriving nonlinear models which pro-
was necessary for analysis and design of the equipment. In this vided more advantages over other conventional techniques. But the
data availability was a crucial issue for ANN in real time applica-
tions. Based on these network conditions, phase angle, power fac-
∗
Corresponding author. tor, and the inner harmonics, current and voltage signals would
E-mail address: sujith@ifet.ac.in (M. Sujith). vary. To overcome all these challenges & drawbacks, an ADALINE

https://doi.org/10.1016/j.micpro.2019.102976
0141-9331/© 2019 Elsevier B.V. All rights reserved.
2 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

structure for harmonics estimation of wind energy is developed. ANN to update these parameters continuously based on input sig-
In the ADALINE network, generally, weights are updated online by nal variation which delivered more accuracy and reliability of har-
using LMS (Least Mean Square), RLS and KF (Kalman Filter) recur- monics amplitudes. This method provided a practical estimation
sive algorithm. The output of the ADALINE will be compared with for every half cycle. This ESPRIT was good in accuracy but it lags
harmonic current and then it creates a modulating signal for gen- with computational time. [4] presented a Fast Transverse Recursive
erating the PWM pulse to the filter. Least Square (FT-RLS) which was used for harmonics estimation.
Harmonics reduction is the important process in power sys- This algorithm helped in calculation of amplitude, frequency, and
tem. In existing works, a passive filter was deployed to decrease phase in time-varying power signals. This approach was mainly
the harmonics distortion at the distribution line. But due to its developed for accuracy and fast estimation of harmonics signals
large structure with less durability, it caused resonance with series under a power system frequency. Initially, the choice of the co-
impedance. The recent trends have introduced an APF to improve variance matrix for the input signals was considered to be more
power quality which carries the advantages over passive filter like critical. Moreover, the computation time and estimation error rate
smaller, cheaper, more versatile, and less prone to failure. would be more under an improper choice of a covariance matrix.
The basic principle of APF was to generate the current compo- Although, the ANN-based estimators were very fast as compared
nent that would be removing the harmonics current produced. In with other earlier techniques. [5] reviewed the application of neu-
traditional methods, the control approach was based on frequency ral networks in wind energy and its applications were categorized
domain. But these frequency domains require large memory, com- into different topics.
putation power, and imprecise results under transient condition. To Based on the survey, the performance prediction of the wind
overcome all these limitations, time domain approach like d and q energy system can be effectively done using ANN. As far as the
transformation, p and q transformation, symmetrical components wind availability in India is concerned, India holds the record of
transformation, are introduced. being the largest wind power capacity in the world right after
Moreover, ADALINE based technique is utilized to extricate basic other giants like China, U.S and Germany [6]. We all know that
sinusoidal from a distorted load current waveform that makes it a wind is one of the cleanest renewable resources available glob-
simpler and faster method for current extraction. ally. Wind power has also captured the global researcher’s atten-
The main motive of this research work is to estimate and elim- tion due to its latest developments.
inate the harmonics in the power system. The main subject of the Hence, future research on neural networks is also discussed in
research work is optimizing the harmonics based on ADALINE net- this work like wind energy optimization. Perfect data selection was
work. The harmonics are optimized by controlling the entire de- required for the ANN set up. The review analyzed that the ANN
sign using active power control. The proposed work contains an approach had performed better than the conventional approach.
APF which is connected with the distribution system for harmon- [7] worked with a single-phase standalone wind energy translation
ics reduction and ADALINE, a version of ANN is introduced as a system that utilized a two winding Self-Excited Induction Genera-
new harmonic detection technique. The active power control de- tor (SEIG). The controller utilized in this work consists of an In-
pends on neural network techniques. Initially, the current signals telligent Neural Network-Based Control (INNBC) algorithm which
from the distribution system is fed to the ADALINE, based on that controls a Battery Energy Storage System (BESS) and two leg Volt-
the PWM generates the switching signal to the active filter. The age Source Converter (VSC) is contained in it. An excellent dynamic
versatile neural network calculates the fundamental and harmonic and steady-state response of the system was achieved by the de-
components from non-linear load current signal. Then the active gree of load current that was modified using the proposed algo-
filter output current will be fed to the distribution line. In between rithm. Constant voltage and frequency were maintained in all types
that a hysteresis current, controller will gather the signals, inject of load conditions which were proved by the experimental and
the compensating signal to make the line current as sinusoidal. The simulation test respectively. [8] came up with a new strategy to es-
major contribution of this research work is: timate harmonics distortion based on ADALINE network in which
the decomposition depended on the LMS training algorithm and
• To estimate the magnitude and phase of the harmonics with
the Fourier series analysis of current signals. The output of ADA-
the method of FFT-Fast Fourier Transform using ADALINE for
LINE was checked with the line current to generate the modulat-
renewable energy resources.
ing signal from PWM for active line conditioner. From this work,
• To propose a new control technique by using an Active
the switching power supply was improved and an Insulated Gate
power filter for harmonics reduction.
Bipolar Transistor (IGBT) locked up problem was solved. More-
Other sections of this paper are organized as follows: over, ZVS-PWM - Zero Voltage Switching - Pulse Width Modula-
Section 2 presents recent techniques related to harmonics estima- tion had better efficiency. [9] presented a Bilinear Recursive Least
tion and reduction. Section 3 provides the design procedure and Square (BRLS) for estimating the frequency, phases, and amplitudes
t mathematical equation of the proposed work. Section 4 com- based on time variance of power signals. The stationary harmon-
prises performance analysis and its comparative results with the ics signal and the dynamic harmonics signals were analyzed in this
proposed techniques are related. The conclusion is provided in paper.
Section 5. To prove the performance a variable frequency drive panel was
used to control the induction motor in a large paper industry.
2. Literature review [10] uses fractional order repetitive control method under a fixed
sampling rate, which will handle the variable frequency of all pe-
In this section, the different harmonic detection methods with riodic signals. A Lagrange-interpolation-based Fractional Delay (FD)
harmonics compensators are discussed. filter was utilized for certain fractional delay items. This technique
offers a fast online tuning of fractional delay. This work provides a
2.1. Harmonics estimation fast update of the coefficient.

[3] asserted the harmonics estimation based on trained ANN 2.2. Harmonics elimination
with a time-dependent power supply signals. This work con-
tributes better estimation accuracy by ESPRIT - Estimation of Sig- [11] presented a harmonic detection method based on Artifi-
nal Parameters via Rotational Invariance Technique which helps cial Neural Network (ANN) with Hybrid Active Power Filter (HAPF)
 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 3

for reducing the issues in enabling the design of current controller. power compensation, but some systems were subjected to errors
This method and design is utilized in ADALINE network. By this and time-consuming processes.
approach, the harmonic current can be satisfied based on neu-
ral network utilized. But the conventional Pi controller based P-Q 3. Proposed work
theory regulated the capacitor DC (Direct Current) voltage of APF.
When this was compared with the ANN, the PI controller incor- The research work demonstrates that the estimation and
porated with P-Q theory had low settling time and rising time. elimination of harmonics is an important issue in the distribu-
[12] analyzed APF that suppressed the harmonics current in the tion system. For illustration, outages in power, variation in volt-
distributed network. Two different control strategies were provided age, sags, swells, harmonics, transients, flicker and noise, are
namely, independent current control strategy and harmonics elim- most governing power quality issues. Due to overloading and
ination. In this p and q methods control each phase individually. loose connections, the voltage variation occurs. Harmonics are
The main objective of this proposed work was to develop an al- initiated due to utilization of dynamic loads that draws a non-
gorithm based on Synchronous Reference Frame (SRF) based con- sinusoidal source. These conflicts result in interruptions, reduced
troller. From this technique, it was possible to control each phase lifespan, and disaster for electrical equipment. The most of the
of the 3 PH (Phase) 4 wire method. But in this work, negative researchers investigated and executed the various strategies such
fundamental and zero sequence compensation were not consid- as Recursive least square (RLS), differential evolution (DE) algo-
ered. [13] presented a comparative study on involved harmon- rithm and so on are employed to eliminate the harmonics. The
ics characteristics using reference generation methodologies with error value is estimated by determining the difference between
shut power filter which was active. This work considered instan- desired and filter output signal by employing RLS method, which
taneous power theories (p and q theories) and ADALINE network is less than least mean square. But this method exhibits com-
control techniques based on Least Mean Square - Equalizer (LMSE) plexity and unstable. The DE method is employed for contin-
and Recursive Inverse Algorithm (RIV) were taken and considered uous optimization function, but it takes more execution time.
for their comparison. The harmonics estimation was calculated The algorithm complexity is increased due to high requirement
based on the online mode and individual mode so that APF could of samples for mitigation the harmonics. To overcome these
realize compensation. The author had concluded that a neural net- contests, optimal ANN controller is developed to eliminate the
work could effectively estimate the harmonics individually. The harmonics.
paper had also concluded that implementation of ADALINE was In this section, different methods involved in the proposed
simple enough for practical applications considered by them. work will be explained. The general concept of this work is to con-
[14] proposed a reference compensation current extraction scheme nect an APF to the distribution system, in order to minimize the
for APF. The main motive was to minimize the harmonics level harmonics of the wind energy and estimate the harmonics present
of 5% at the point of common coupling. The PI controller effec- in it. In this proposed work, an ADALINE based network is also in-
tively minimized the DC link voltage regulation and these gains cluded for accurately estimating the harmonics which will be pro-
were optimized by Particle Swarm Optimization (PSO) in a more viding perfect compensation to the power lines. The combination
enhanced manner. In this, the APF contained the voltage source, of using the APF with the ADALINE makes this proposed method
inverter, pulse width modulation technique. [15] proposed a Model a novel one. As a result, the power received from the wind en-
Reference Adaptive Sliding Mode Control (MRASMC) using a Radi- ergy is getting converted into three-phase signals with the aid of
cal Basic Function (RBF) for controlling the single-phase ACF. Permanent Magnet Synchronous Generator (PMSG). In the above
This RBF with NN helped to estimate the nonlinear functions process, the signal of the line current is obtained from the distri-
to rectify modeling error at the APF system. The main advantage bution system which is in turn passed on to ADALINE. All the req-
of this problem was that the asymptotic stability and the line of uisites are shown in the Fig. 1 with the help of a frame work. In
the system had the ability to adjust the weight of the neural net- this framework, wind acts as the driving force for harmonics elim-
work. Thus the tracking performance was improved by the slid- ination which is going to be executed after appropriate estimation
ing mode controller at the DC side. [16] proposed derivate loss of the same.
controller based on a nonlinear load, which helped to control the
SAPF. In this work, some base frames were introduced by replac-
ing it with stationary frame. Both the current dynamic inner loop
and the voltage dynamic outer loop were utilized in this work. But
when compared with SAPF, Shunt Hybrid Power Filter (SHPF) min-
imized the active parts power rating to the better part.
So, this SAPF was not up to mark for meeting the day-to-
day requirements. [17] presented a comparative analysis on the
weight updating adaptive algorithm based on fuzzy logic based
variable step size, Least Mean Square - LMS and LMS based
ADALINE for current harmonics detection. These techniques were
compared with Distribution Static Synchronous Compensator
(DSTATCOM). [18] proposed an enhanced self-charging algorithm
by including a step size error cancellation in SAPF. The step size
cancellation had some additional features to the self-charging algo-
rithm with dynamic and steady operations. [19] reviewed a prob-
lem based on reference signal generation for SAPF to eliminate
Total Harmonics Distortion (THD) in the distribution line. The
problems related to the adaptive control methods like steady-state
error, speed, and stability were addressed in this paper. From the
results, this paper had delivered the control method SAPF with
ADALINE that performs well in dynamic conditions. This technique
provided satisfying results in harmonics reduction and reactive Fig. 1. Overall frame work of the proposed work.
4 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

Fig. 2. Proposed ANN Controller Block.

Fig. 5. Schematic diagram of the devised work.

signal disintegration will aid in defining the input for the neural
network that is being used here. The weight training will be car-
ried out by LMS algorithm. Then these fundamental components
outputs are compared with distorted line current. This kind of
process will be helping in generating the modulating signal. This
modulation switching pattern is deployed for producing the PWM
switching pattern for the power switch and then, the active fil-
ters output current is injected into the power line. The hysteresis
current controller will be helping to support the harmonics detec-
tion process by comparing the DC side voltage control and current
signals. The ANN uses the fundamental and the harmonic compo-
nents from the non-linear load current signal for further calcula-
tion. For regulation of voltage (DC) of inverter constant involved,
the outputs of the proportional and integral controller are used.
Afterwards, comparisons are carried out using the output value to
generate the output compensating current is produced by the in-
verter. Thus, the differences between these above values are given
as an input to hysteresis band comparator. At the same time, hys-
teresis current controller will be gathering these signals and it in-
jects the compensating signal to make the line current as sinu-
soidal.
Fig. 3. Flow of algorithm.
3.1. Active power filter with neural network

The involved nonlinear load will be creating many power and

characteristic issues like voltage flicker, harmonics, and voltage sag,
swell. For reducing the harmonics in the distribution line, a filter
will be added for compensation.
In this proposed work, an APF will be generating the compensa-
tion current. The main motive is to obtain a source current with-
out any harmonics. The perfect compensation current is injected
by the APF will enable to resemble the load current as a non-active
component. This proposed filter is developed with inverter circuit
on the DC side. In this paper, an active or a dynamic power fil-
Fig. 4. Artificial neuron model.
ter consisting a hysteresis band comparator is added to eliminate
the harmonics in the nonlinear loads. Most importantly, the major
aim of deploying the controller is to maintain the compensation
current. In this process, the 3ph inverters’ switching strategy will
Main parts such as PMSG, PWM generator, a band comparator, keep the current under hysteresis band. Meanwhile, the load cur-
active power filter, ADALINE based network and DC regulator will rent will be calculated and then it will be analyzed with non-active
be working together for improving the some such as current, volt- components and reference currents.
age, etc. if any towards achieving the productive harmonics elimi- The design of the proposed ANN block diagram is depicted in
nation the Fig. 2. The number of neurons in the input layer is 4 and the
This framework will use the fundamental components from line output layer is 3. The number of data considered for the training
current signals based on Fourier series. This kind of new current state is 200. As shown in Fig. 3, the proposed Adaline controller is
 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 5

Fig. 6. Generator design.

Table 1
Simulation specifications.

Parameters Values

Source Direct Driven PMSG

V = 440V, f = 50Hz, Ls = 0,6mH
APF DC link Capacitor = 2100μF
DC link voltage = 680V
Filter inductance = 1,2e-3mH
Resistance = 0,5Ω
ANN Input Layer = 4, Hidden Layer = 20, Output Layer = 3
Load 3-Phase Bridge rectifier
Parameters Values
Source Direct Driven PMSG
V = 440V, f = 50Hz, Ls = 0,6mH

Fig. 7. Hysteresis current controller design.

employed to control the current in APF and used to generate the

reference signals for the current controller block. The Adaline neu-
ral network is trained by the Levenberg-Marquardt Algorithm. The
devised work is to train the neural network for maintaining con-
stant dc-link voltage and for the estimation of reference signals.
For upholding the dc-link capacitor voltage, the reference voltage is
matched with the actual Vdc value and error is taken as input data
and the loss component of current (Idc ) is given as the target data
to the network. For generating the reference currents, the load cur-
Fig. 9. Time window when voltage was active.
rent (ILabc ) and Idc are considered as input data and the predicted
reference currents are considered as target to the network. Based
on all these happenings, the proposed control will deliver with ex-
cellent filtering dynamic response. In addition, for any kind of load
current, this compensation current can be adapted quickly. From is estimated, correspondingly the phase angle and voltage magni-
the instantaneous power block, the real (P) and reactive (Q) power tude vectors are measured, the measured value is given as input to

Fig. 8. Hysteresis controller pulse.

6 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

function. The neural architecture depends on three layers: the in-

put layer, hidden layer, and an output layer.
Thus, the feedback architecture computes the input data in a
parallel way rather than computer sequential algorithm. To sup-
ply the output target, the network can be trained based on the
appropriate input. The back propagation algorithm is most widely
utilized in this scenario. Under this methodology, the random value
will be assigned initially and afterwards, the current output will
be compared with the initial output pattern. Until the error rate
is minimized, the utilized algorithm will keep on adjusting the
weights.
Firstly, input-output data is obtained from the electric power
generated by using the wind as the medium which is then fed to
Fig. 10. Time window when current was active.
in-built layered ANN for computing the errors. Then, the decision
is made after computing the errors in total after which decision
will be required whether to update the weights or not.
After this decision making operation, if there is no error, the
weights would be updated accordingly. If there is error, again the
training will be initiated for re-computing the errors encountered
until now. Finally, the testing will be done and then the algorithm
ends.
Typical construction of an artificial neural network is depicted
in the Fig. 4. This ANN network is utilized for the training the
dataset efficiently because of its in-built layers for smooth con-
duct of the further processes. Generally, there are three layers pre-
sented in the ANN such as the Input layer, Hidden layer, and an
output Layer which had been discussed already. The i(1) to i(n) is
given as input data on to the ANN. The weight value of this value
is reckoned in the layer which is hidden and it gets summed up.
Then, the weight value is altered based on the learning rate and
the partial derivatives of Loss Function. After the transfer function,
the output layer will give rise to output. Below Fig. 5 shows the
Simulation pictorial of the devised work which is being utilized
for our proceedings. This simulation comprises the generator and
its subsidiary components such as controllers, pilot operated open
Fig. 11. THD value for the source voltage waveforms by FFT.

Table 2
the estimator of ANN block. Fig. 3 shows the working flow of the
ANN Parameters For Training.
algorithm that is preceded here in this work.
Parameters Values

No of Training data 100

3.2. Control of compensation current No of Testing data 50
No of neurons- input layer 4
In the feed-forward neural network control principle, the neu- No of neurons -hidden layer 20
No of neurons- output layer 3
ral network contains many strongly connected elements. Thus, the Algorithm Levenberg–Marquardt (LM)
input data of i(l), i(2), i(3), …, i(n) will go through weights which Function Mean Squared error
will be collected in a node and get characterized as a circle. Before Maximum Epochs 50
their addition, these weights will be modifying the input signals Learning rate 0.05
Performance goal 1e-6
in either way by magnifying or devitalizing. If the process is com-
Accuracy 98.19
pleted, then the data goes through the output through the transfer

Fig. 12. Three phase waveform for source voltage.

 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 7

Fig. 13. THD value for the source current waveforms by FFT.

Fig. 14. Three phase waveform for source current before APF.

Table 3
ANN Measurement.

Frequency (Hz) Order of Harmonics Test Signal ANN Measurement Relatively Error (%)

49 3th Mag (p.u) 0.06 0.0627 0.0027

Angle (p.u) 180 179.9901 -0.0099
5th Mag (p.u) 0.08 0.0781 -0.0019
Angle (p.u) 30 29.9701 -0.0299
49.5 3th Mag (p.u) 0,26 0.2501 -0.0099
Angle (p.u) 120 119.2571 -0.7429
5th Mag (p.u) 0.2 0.1902 -0.0098
Angle (p.u) 60 59.9312 -0.0688
50 3th Mag (p.u) 0.45 0.4421 -0.0079
Angle (p.u) 90 89.9286 -0.0714
5th Mag (p.u) 0.39 0.3852 -0.0048
Angle (p.u) 90 89.7551 -0.2449
8 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

Fig. 15. (a) Training (b) Validation (c) Testing, and (d) Total response.

and closure doors, 3 phase lines to make use of the wind power
to give rise to voltage, current, etc. for the purpose of the harmon-
ics elimination via the simulation. Below Fig. 6 illustrates the de-
sign of the Generator that is being deployed in our work which
makes use of the wind energy to produce the required power. In
this work, we are just using the generator for the purpose of test-
ing the outputs raised form it for the efficient elimination of the
harmonics in the wind. In later sections, we will be discussing the
various components along with its function as and when required.

3.3. Adaptive linear neural network

The Fourier analysis will expand the periodic waveform by the

sum of sine and cosine frequency components. In this process, Gra-
dient descent- an iterative mitigation technique is being added for
the betterment. The route of the steepest ascent of the error func-
tion will be always indicated with the error functions’ gradient.
Initially, it will be starting with a random weight vector and then
Fig. 16. Performance of proposed ANN control.
follows the negative gradient as shown below,

wt+1 = tw + tw (1)

  follows:
tw = −Efficiency ∂ E (w ) /∂ w at tw
T
(2)
 
It helps in repeating the gradient descent procedure for number
tw = α tw−1 − (1 − α ) ∗ efficieny ∂ E (w )T /∂ w at tw for 0
of random initial states because of the local minimum, which is as ≤ α≥ 1 (3)
 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 9

Fig. 17. Number of Epochs Vs. training state parameters.

 
Thus, it will be able to increase the efficiency by minimizing h ( 2 ) a1 ( 2 )
the error as a result of all the above procedures that had been fol- o= h ( 2 ) a2 ( 2 ) (13)
lowed. h ( 2 ) an ( 2 )
The neural network will process with its input layer, an output
layer, and hidden layer. The weight training process will be car- ∂o  
= diag h (2 )(ka (2 ) ) = h(2 ) (14)
ried out by the hidden layer. Consider a two layer neural network, ∂ a (2 )
namely an output layer and hidden layer, we get The network has d inputs, m hidden neurons and n output neu-

a j (1 ) = w ji (1 )ix = jw (1 )x (4) rons, therefore the transfer function is given by,
  
i 
m 
d
  ko = h ( 2 ) kw j (2 )h(1 ) jwi (1 )ix (15)
a1
j−1 i−1
a (1 ) = a2 = X (1 )w (5)
am In the matrix form, we get
O = h (2 ) (w (2 ) h (1 ) (w (1 ) x ) (16)
With W ji (1 ) = W ji (1 ) as the weight matrix of the hidden layer
and the jth row contains the weights of neuron, Thereby determining the output “o” function of the input x is
also represented as forward sweep in the back propagation algo-
Z j = h(1 )( ja (1 ) ), ∂ jz /∂ ja (1 ) = h (1 )( ja (1 ) ) (6)
rithm.
 
h ( 1 ) a1 ( 1 )
3.4. Hysteresis band controller
z= h ( 1 ) a2 ( 1 ) (7)
h ( 1 ) am ( 1 )
The feed-forward neural network works as a hysteresis band
∂z   comparator under the PWM control. This network is developed
= diag h (1 )( ja (1 ) ) = h(1 ) (8)
∂ a (1 ) with 2 hidden layers along with 14 neurons and one output layer
along with one neuron. The activation functions are linear in the
The output layer is given as,
layer of the output and log sigmoid in the hidden layer. The pat-

ka ( 2 ) = kw j (2 )z j = kw (2 ) (9) terns are trained by the back propagation algorithm. The output of
j the comparator will be based on their evolution and inputs respec-
  tively. The network weight issues can be fixed under these hap-
a1 ( 2 )
penings. Thus, there is a need for the comparison of the network
a (2 ) = a2 ( 2 ) = zw ( 2 ) (10) output with the real electrical system output. After which, the hys-
an ( 2 )
teresis band comparator will be delivering the output pulses to the
ko = h ( 2 ) ( ka ( 2 ) ) (11) inverter,
The algorithm for this controller is,
∂ ko/∂ ka (2 ) = h (2 )(ka (2 )) (12) mi,re f (t ) = mI,re f (t )sin(ωt ) (17)
10 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

Fig. 18. (a) Three phase waveform for load voltage (b) DC voltage for the controller (C) Reference current and injected current.

Maintain the same rate, where m = a, b, c three phases, I is the

Upper band ui = mi,re f (t ) + i (18)
load current, and the Vs is the dc link voltage of the inverter. Even
Lower band li = mi,re f (t ) + i (19) then, the frequency in the PWM method will not be constant so
that the output will be giving rise to non-optimum harmonics.
Where,
i = hysteresis band limit, Compute Peak detector pdc = (LMS(vsa + vsb + vsc ) ∗ π /2 )
If mi > ui , mov = − vs /2 (20) ∗ (Isa + Isb + Isc ) (22)

If mi < li , mov = vs /2 (21) Let, Vmax(abc ) = LMS (1.414 ∗ vs ) // maximum voltage source (23)
M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 11

Fig. 18. Continued

Fig. 19. Load current output with an Active Power Filter (APF).
12 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

Fig. 20. Load current output without an active power filter.

leakage scenario. Likewise, for the harmonic estimation, the gen-

Let Vmax = LMS (1.414 ∗ (vsa + vsb + vsc ) ) (24) eral equation for time window for the FFT is given by,
∞  
F F T (ω ) = f (t ) e(−kωt ) dt (29)
Let Ica = pdc ∗ Vmax(abc ) V (25) −∞
max
Where, f (t) be the signal which is band limited necessarily, FFT
Let Icb = pdc ∗ Vmax(abc ) V (26) (ω) be the entity defined by the well sampled information, e is
max indicating the exponential operation, and k is just a constant that
is being introduced for better correlation.
Let Icc = pdc ∗ Vmax(abc ) V (27) The designated data resolution [20] for FFT can be found by us-
max
ing the below devised relation as follow:

I∗ abc = (Ica − Isa ) + (Icb − Isb ) + (Icc − Isc ) (28) fe

S = β (30)
N
In this processes, the source voltage and current are considered
as the input. In order to find the peak detector value, the least
mean square of all the three-phase voltage is calculated and to 4. Performance analysis
measure the maximum voltage source, the gain value will be mul-
tiplied with the voltage source. The current value will be calculated The performance or fulfillment of the proposed system with
by raising the peak detector value along with the individual voltage different parameters like voltage, current, load, and total distor-
values. Finally, the current in all three phases will be calculated by tion occurring in terms of harmonics were analyzed in this sec-
adding and subtracting the individual phase current values. Based tion. Moreover, to show the improvement of the outcome of the
on these values, the signal will be generated accordingly. proceeded work, a comparison is made with the proposed work
Figs. 7 and 8 illustrate the design of the hysteresis current con- and the existing PI controller. Table 1 shows the specifications of
troller deployed in this work and the pulse generated during the matlab simulations.
hysteresis operation respectively towards the effort of the harmon- Figs. 9 and 10 represents active time of simulation profile for
ics elimination in this work. voltage and current. The above Fig. 11 depicts the THD value for
Any Harmonic estimation will be always subjected to the time the corresponding source voltage waveforms that is being used
windows [20] for lessening the severity caused due to the leak- throughout the performance analysis of our work. This value in-
age of spectra which might be caused or being at the verge of the dicates the various inputs utilized and standards available in the
 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 13

Fig. 21. (a) Total Harmonics Distortion before APF (b) Total Harmonics Distortion after APF.
14 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

exported place of the FFT inbuilt which had been taken from the
scope and arrays that had been generated earlier.
Fig. 12 represents the source voltage waveform of the proposed
circuit in our work, in which a three-phase source is taken as an
input in order to get a valid output. From the waveform, it is clear
that the voltage sources are more stable than being unstable in the
worst cases.
Fig. 13 depicts the THD value for the corresponding source cur-
rent waveforms that is being used throughout the performance
analysis of this proposed work. This value helps to realize the var-
ious inputs utilized and standards available in the exported place
of the FFT inbuilt which had been taken from the scope and arrays
that had been generated earlier.
Fig. 14 represents the three phases of the current waveform at
the load side which is depicted above. As we already know that,
due to the power electronics components the distribution line will
be more disturbed. The graph also resembles in the same way
which depicts that the nonlinear load creates distortion in all these
three phase lines.

4.1. ADALINE network performance

Fig. 15 represents the training, testing, validation, and total

response of the data for neural network strategy. The R value
represents the association between targets and its correspond-
ing outputs. If the value of R equals to 1, then there will be a
linear association between the targets and its corresponding out-
puts. The Fig. 15 shows the performance of ANN training and vali-
dation process.
Fig. 16 depicts the fulfillment of the devised control technique. Fig. 22. (a) An analysis of THD with comparison with existing techniques (b) Com-
parative analysis based on Total Harmonics Distortion.
From the above graph, we can infer that all the testing, training,
and validation values were above the best or benchmarked values.
The performance or fulfillment of the training state with re-
spect to a gradient, Mu, and validation are explained with help The load currents’ THD is monitored to analyze the fulfill-
of the above Fig. 17. This shows how much iteration has been ment or performance of devised controller in this work. The below
carried out in this process in the considered number of epochs Fig. 19 illustrates the total harmonics level which is measured only
50. Table 2 shows the network parameters considered for test- after connecting the filter.
ing and validation of neural network. Table 3 shows the accu- By this way, the selected signals represent the voltage source,
racy of the measurement done by the ANN for the harmonics current active power filter, and dc voltage which is being used for
components. the controller. Then, the fundamental frequency is taken as 50 HZ
- normal value and distortion rate in the total harmonics are also
calculated simultaneously. The THD obtained after using the filter
4.2. After applying filter is 3.84%, which is considered to be lower harmonics when it is un-
der a nonlinear load. Thus the performance or fulfillment of the ac-
The main contribution is to estimate the number of harmonics tive power filter based on ADALINE which performs well in terms
present in the power system which is being shown by the above of harmonics elimination. Fig. 20 shows the total harmonics level
Fig. 14 that illustrated the current outputs from the source as well which is measured before connecting the filter.
as the load measurement. After applying the filter to the distribu- Fig. 20 shows the total harmonics level which is measured be-
tion system, the harmonics present in the line are eliminated and fore connecting the active power filter (APF). It can be seen that
then it attains a pure sine waveform. Fig. 18 shows how the har- the three phases of output load current varies with each other and
monics content are being eliminated by our proposed work. as a result, fluctuation occurs in the current.
The Fig. 18(a) shows the behavior of the involved load voltage,
current and source current after applying the devised filter.
4.4. Total harmonics distortion
The Fig. 18(b) is a graph plotted for DC link capacitor voltage. It
shows the behavior of the involved load voltage after applying the
The distortion in terms of the total harmonics is an important
devised filter.
parameter which helps to analyze the number of harmonics pre-
The Fig. 18(c) represents the reference current and injected cur-
sented in the voltage or current waveform. Any repetitive wave-
rent in the line respectively.

4.3. Simulation result for the proposed controller Table 4

Steady-state analysis of the proposed filter.

The steady state analysis based on pro-active power filter is Steady state analysis
given in Table 4. The analysis is done on the basis of setting time Without APF With APF
and overshoot. As a result, before adding the filter circuit, the set-
Setting Time 7.4915% 0.503%
ting time and overshoot seemed to be high. But, after the APF in- Overshoot 2.8% 1.928%
troduction, the output value is found to be significantly reduced.
 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 15

proposed and PI controller respectively. This seemed to be a clear

proof of the proposed filter performing well in the harmonics elim-
ination.

4.5. Comparative analysis of the proposed work

The overall performance of the novel controller is validated by

comparing the proposed controller with the existing PI controller
deployed in the work of [21]. In this exiting work, the induction
motor is considered to be a nonlinear load. The comparative graphs
show the total harmonics value before and after adding the fil-
ter. In the considered existing method, the THD value obtained
was 6.2% but in the proposed work, the THD value is found to be
reduced to 3.8%. Even though, the initial value of the harmonics
before adding filter in the proposed work was higher than the ex-
isting work.
Fig. 23. Laboratory Prototype.
It can be inferred that the proposed works’ performance is
found to be much more superior to the existing PI controller. Be-
form can be represented in mathematical form as pure sine waves’ fore adding the Filter, the existing PI controller showed THD value
series in the harmonic analysis that we make. The multiplying fac- as 46% and the proposed controller exhibits THD value as 30.32%.
tors of underlying frequency is contained in the sine waves are Before the filtering case shown in Fig. 22, the value of harmonics
called Harmonics. is found to be reduced when compared with the existing PI con-
Thus the Fig. 21(a) and (b) illustrate the harmonics presented troller.
in the distribution is considered, before connecting with the filter Table 5 shown above indicates the overall comparison of the
and after connecting to the filter. In this proposed work, the filter THD for all three phases with the existing LMS and RLS in this
used here is an APF based one. Before being connected to the filter, work [19]. While comparing with the existing work, the pro-
the THD value is noted as 30.32% with the proposed controller and posed control method of APF with ADALINE network performs
46% with the PI controller. At that moment, after being connected well. Table 6 shows the current compensation in source and load
to the APF, the harmonics level drops to 3.80% and 6.2% with the side.

Fig. 24. (a) Procured load and source current before compensation with injected current (b) Procured source voltage and current after compensation (c): Filter performance
based on compensation of neutral current.
16 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976

Table 5 Declaration of Competing Interest

Comparative analysis based on Total Harmonics Distor-
tion.
An APF control design is developed with ADALINE network
Phase 1 Phase 2 Phase 3 in which the load and current along with voltage will be an-
Proposed ADALINE 3.84% 4.00% 3.80% alyzed and then the controller will be calculating the control
LMS 14.37% 18.94% 15.74% signal by considering the reference compensation current. After-
RLS 6.27% 7.71% 6.76% wards, the power system is injected with compensating current.
The simulation is carried out with Matlab-Simulink to analyze
the proposed control designs efficiency. The proposed work con-
Table 6
summation is compared with conventional PI controller method
Before and after (Voltage, current) compensation. comprising Shunt Active Power Filters (SAPF) with ADALINE for the
performance perspectives. This method was found to be effective
Current
in terms of many parameters such as load voltage, load current,
Source side (A) Load Side (A) voltage, reactive power, real power and especially THD value than
After compensation 100 50 those of the existing works which are considered.
Before compensation 50 33

References

4.6. Laboratory prototype [1] Y. Xia, D.P. Mandic, A widely linear least mean phase algorithm for adaptive
frequency estimation of unbalanced power systems, Int. J. Electr. Power Energy
Syst. 54 (2014) 367–375.
A laboratory prototype is executed to validate the ANN based
[2] H. Xue, M. Wang, R. Yang, Y. Zhang, Power system frequency estimation
shunt active filter using FPGA Embedded board -Xilinux Spartan 3E method in the presence of harmonics, IEEE Trans. Instrum. Meas. 65 (2016)
depicted in Fig. 23. The pulse width modulation is used to switch 56–69.
[3] S.K. Jain, S. Singh, ESPRIT assisted artificial neural network for harmonics de-
the voltage source inverter to controls the supply current followed
tection of time-varying signals, in power and energy society general meeting,
by derived reference source current. The input source voltage and 2012, IEEE (2012) 1–7.
DC (Vdc ) voltage are predicted using voltage transducer and input [4] S.K. Singh, A.K. Goswami, N. Sinha, Harmonic parameter estimation of a power
currents are predicted using current transformers (CTs). The fore- signal using FT-RLS algorithm, in harmonics and quality of power (ICHQP), in:
2014 IEEE 16th International Conference on, 2014, pp. 157–161.
cast signals are carried to the ADC card of FPGA embedded con- [5] R. Ata, Artificial neural networks applications in wind energy systems: a re-
troller. The unit vector template is generated with help of input view, Renew. Sustain. Energy Rev. 49 (2015) 534–562.
source voltage. The actual value of DC voltage is compared with [6] B.K. Sahu, M. Hiloidhari, D. Baruah, Global trend in wind power with special
focus on the top five wind power producing countries, Renew. Sustain. Energy
the reference voltage to generate the error signal and it is handled Rev. 19 (2013) 348–359.
through proposed controller. In addition, reference source current [7] U.K. Kalla, B. Singh, S.S. Murthy, Intelligent neural network-based controller for
is generated by the product of the input source current and unit single-phase wind energy conversion system using two winding self-excited
induction generator, IEEE Trans. Ind. Inf. 12 (2016) 1986–1997.
vector template. The switching pulses are spawned by using pro- [8] H. Laib, A.E. Chaghi, P. Wira, A neural and fuzzy logic based control scheme
posed control algorithm based pulse width modulation and refer- for a shunt active power filter, in: in International Conference on Electrical En-
ence current extraction. These switching pulses are used to per- gineering and Control Applications, 2016, pp. 201–211.
[9] S.K. Singh, A.K. Goswami, N. Sinha, Power system harmonic parameter estima-
ceive the Shunt active filter.
tion using bilinear recursive least square (BRLS) algorithm, Int. J. Electr. Power
Fig. 24 shows the experimental results to verify the effective- Energy Syst. 67 (2015) 1–10.
ness of the proposed ANN controller. Also, the developed control [10] Z.-X. Zou, K. Zhou, Z. Wang, M. Cheng, Frequency-adaptive fractional-order
repetitive control of shunt active power filters, IEEE Trans. Indust. Electron. 62
strategy is verified by the neutral current compensation. After ap-
(2015) 1659–1668.
plying the filter, the neutral current drops to zero shown in Fig. [11] Y. Han, L. Xu, M.M. Khan, C. Chen, G. Yao, L.-D. Zhou, Robust deadbeat con-
24(c). trol scheme for a hybrid APF with resetting filter and ADALINE-based har-
Table 6 shows the current compensation in source and load monic estimation algorithm, IEEE Trans. Indust. Electron. 58 (2010) 3893–
3904.
side. [12] L.B.G. Campanhol, S.A.O.da Silva, A. Goedtel, Application of shunt active power
filter for harmonic reduction and reactive power compensation in three-phase
four-wire systems, IET Power Electronics 7 (2014) 2825–2836.
[13] S. Sharma, V. Verma, Comparative analysis of ADALINE based shunt active
power filter, in: 2017 IEEE International WIE Conference on Electrical and
Computer Engineering (WIECON-ECE), 2017, pp. 39–42.
5. Conclusion [14] V. Gali, N. Gupta, R. Gupta, Enhanced particle swarm optimization technique
for interleaved inverter tied shunt active power filter, in: In Soft Computing for
To achieve a better harmonics elimination and accurate esti- Problem Solving, Springer, 2019, pp. 571–583.
[15] Y. Fang, J. Fei, K. Ma, Model reference adaptive sliding mode control using
mation, an APF control method has been presented in this work. RBF neural network for active power filter, Int. J. Electr. Power Energy Syst.
To increase the accuracy in the performance, a PWM control 73 (2015) 249–258.
with ADALINE network was developed in this proposed work. The [16] S. Ye, Y. Zhang, L. Xie, H. Lu, Shunt active power filter based on proportional in-
tegral and multi vector resonant controllers for compensating nonlinear loads,
method for obtaining the reference current and the fundamental J. Electric. Comp. Eng. 2018 (2018).
active currents are also described. A hysteresis band comparator [17] G.S. Chawda, A.G. Shaik, Performance evaluation of ADALINE controlled Dstat-
was developed with a feed-forward well-trained neural network com for multifarious load in weak AC grid, in: In 2019 IEEE PES GTD
Grand International Conference and Exposition Asia (GTD Asia, 2019, pp. 356–
with the devised back propagation algorithm. Thus the proposed
361.
work effectively eliminates and accurately calculates the harmon- [18] M.A.A.M. Zainuri, M.A.M. Radzi, A.C. Soh, N. Mariun, N.A. Rahim, DC-link ca-
ics. The results obtained from Matlab Simulink shows voltage and pacitor voltage control for single-phase shunt active power filter with step size
error cancellation in self-charging algorithm, IET Power Electronics 9 (2016)
current. As compared with the traditional controller technique, the
323–335.
proposed controller has reduced total harmonic distortions more [19] R. Martinek, J. Vanus, P. Bilik, T. Stratil, J. Zidek, An efficient control method of
effectively for dynamic loading conditions. Experimental and sim- shunt active power filter using adaline, IFAC-PapersOnLine 49 (2016) 352–357.
ulation results verify the effectiveness of the control strategy and [20] F.J. Harris, On the use of windows for harmonic analysis with the discrete
Fourier transform, Proc. IEEE 66 (1978) 51–83.
it is more operative than the conventional controller. (Tables 1 and [21] A. Sivakumar, N.M. Selvan, Reduction of source current harmonics in ANN con-
4, Figs. 9 and 10). trolled induction motor, Alexandria Eng. J. (2017).
 M. Sujith and S. Padma / Microprocessors and Microsystems 73 (2020) 102976 17

M. Sujith was born in Namakkal, Tamil Nadu in 1987. S. Padma was born in Salem, Tamil Nadu in 1969. She
He received the B.E. degree in Electrical and Electronics received the B.E. from Government College of Engineer-
Engineering from K.S.R. College of Engineering, Tiruchen- ing, Salem, in 1990 and M.E. degree from Annamalai Uni-
gode, in 2008 and the M.E. degree in Applied Elec- veristy, Chidambaram, in 2002. She received her Ph.D. in
tronics from the Annai Mathammal Sheela Engineering 2011 from Anna University, Chennai. She is working as
College, Namakkal, in 2012. He is working as Associate Professor in the Department of Electrical and Electron-
Professor in the Department of Electrical and Electron- ics Engineering at Sona College of Technology, Salem, In-
ics Engineering at I.F.E.T College of Engineering, Villupu- dia from 2011. Her research interests include power elec-
ram. His main research interests include power electronic tronics converters, drives, energy storage, power quality
converters, compensators, power quality issues, and ac- issues, and active power filters.
tive power filters.