World Academy of Science, Engineering and Technology 33 2009

Artificial Neural Networks Application to

Improve Shunt Active Power Filter
Rachid.Dehini, Abdesselam.Bassou, Brahim.Ferdi

compensation modern power electronic technology, Active

Abstract—Active Power Filters (APFs) are today the most power filter (APF) have been considered as an effective
widely used systems to eliminate harmonics compensate power solution for this issue, it has been widely used.
factor and correct unbalanced problems in industrial power plants. One of the most popular active filters is the Shunt Active
We propose to improve the performances of conventional APFs by
using artificial neural networks (ANNs) for harmonics estimation.
Power Filter (SAPF) [2-6, 8]. SAPF have been researched and
This new method combines both the strategies for extracting the developed, that they have gradually been recognized as a
three-phase reference currents for active power filters and DC link workable solution to the problems created by non-linear loads.
voltage control method. The ANNs learning capabilities to The functioning of shunt active filter is to sense the load
adaptively choose the power system parameters for both to compute currents and extracts the harmonic component of the load
the reference currents and to recharge the capacitor value requested *
by VDC voltage in order to ensure suitable transit of powers to current to produce a reference current i c , a block diagram of
supply the inverter. To investigate the performance of this the system is illustrated in Fig. 1. The reference current
identification method, the study has been accomplished using
consists of the harmonic components of the load current
simulation with the MATLAB Simulink Power System Toolbox. The
simulation study results of the new (SAPF) identification technique which the active filter must supply. This reference current is
compared to other similar methods are found quite satisfactory by fed through a controller and then the switching signal is
assuring good filtering characteristics and high system stability. generated to switch the power switching devices of the active
filter such that the active filter will indeed produce the
Keywords—Artificial Neural Networks (ANN), p-q theory, harmonics required by the load. Finally, the AC supply will
(SAPF), Harmonics, Total Harmonic Distortion. only need to provide the fundamental component for the load,
resulting in a low harmonic sinusoidal supply.
I. INTRODUCTION Generally, the effectiveness of (SAPF) depends on three

D UE to proliferation of power electronic equipment and

nonlinear loads in power distribution systems, the
problem of harmonic contamination and treatment take on
design criteria: (i) design of power inverter; (ii) use of current
controller’s types (iii) methods used to obtain the reference
current. The presented work was oriented mostly on the latter
great significance .These harmonics interfere with sensitive criterion.
electronic equipment and cause undesired power losses in In order to determine harmonic and reactive component of
electrical equipment[1-8]. In order to solve and to regulate the load current, reference source current generation is needed.
permanent power quality problem introduce by this Current Thus, reference filter current can be obtained when it is
harmonics generated by nonlinear loads such as switching subtracted from total load current. For better filter
power factor correction converter, converter for variable performance, generation of reference source current should be
speed AC motor drives and HVDC systems, the passive filters done properly. For this purpose, several methods such as pq-
have been used; which are simple and low cost. However, the theory, dq-transformation, multiplication with sine function
use of passive filter has many disadvantages, such as large and Fourier transform have been introduced in literature [9-
size, tuning and risk of resonance problems. 14].
Lately, owing to the rapid improvement in power Recently, some methods based on artificial intelligence
semiconductor device technology that makes high-speed, have been applied In order to improve processing detecting
high-power switching devices such as power MOSFETs, time of harmonic current. The past decade has seen a dramatic
MCTs, IGBTs , IGCTS , IEGTs etc. usable for the harmonic increase in interest Artificial Neural Networks (ANNs) which
is characterized by its learning ability and high speed
Rachid.Dehini is with Department of the sciences and technology, , Bechar recognition but simple structure, the (ANNs) have been
University.A, B.P 417 BECHAR (08000), ALGERIA (corresponding author applied in many uses in the power electronic part of both
to provide phone: + 213-049-81-95-91; e-mail: dehinirachid@ yhoo.fr). machinery [16] and filters devices [17-24] where it have
Abdesselam.Bassou is with Department of the sciences and technology, ,
Bechar University.A, B.P 417 BECHAR (08000), ALGERIA Tel: + 213-049- justified their effectiveness. The results obtained with ANNs
90-24, Fax:.+049-81-52-44 (e-mail: [email protected]). are often better than those of traditional methods. Indeed, as a
Brahim.Ferdi is with Department of Electrical & Computer Engineering, result of their capacities to optimize simultaneously their
Oran University of the sciences and technology, Oran, ALGERIA (e-mail:
[email protected]).
weights and biases in an on-line training process, they are able

1058
 World Academy of Science, Engineering and Technology 33 2009

to adapt themselves to any system. where A is the transformation matrix and is equal to:
In this paper, a detection method using artificial neural
network (ANN) is presented which can be utilized in both
harmonic current detection from distorted wave and DC link ⎡1 − 1 −1 ⎤
⎢ 2 2 ⎥
control voltage. This method can precisely obtain the [ A] = ⎢ (2)
reference current of each phase. The learning rate can be 3 − 3 ⎥
regulated in a wide range with little affection on the ⎢⎣0 2 2 ⎥⎦
performance with a simple structure and theory [17-24]. The
performances of the Neural Method are evaluated under
simulation and are compared with p-q theory.
This transformation is valid if and only if

3 phase
v (t ) + v (t ) + v (t ) = 0 and
a b c
also if the voltages are

AC balanced and sinusoidal. The instantaneous active and reactive

powers in the α − β coordinates are calculated with the
Supply i , i , i iia , iib , iic V
c
sa sb sc v sa , v sb , v sc
following expressions:
isa isb i sc
ica
Courrent icb Control p (t ) = v α (t )i α (t ) + v β (t )i β (t ) (3)

q (t ) = − v α (t )i β (t ) + v β (t )i α (t )
Detetction icc Circuit
(4)
S a S b Sc
iia Lf
iib Lf Vdc The values of p and q can be expressed From Eqs.(3) and (4)
iic Lf in terms of the dc components plus the ac components, that is:

iLa i Lb i Lc
p= p+~ p (5)
q = q + q~ (6)

Ls Ls Ls
where:

RL
p : is the dc component of the instantaneous power p, and is
related to the conventional fundamental active current.
LL
~
p : is the ac component of the instantaneous power p, it
Fig.1. Schematic diagram of shunt APF does not have average value, and is related to the harmonic
currents caused by the ac component of the instantaneous real
power.
II. REFERENCE SOURCE CURRENT GENERATION
The concept of instantaneous reactive power theory (p-q q : is the dc component of the imaginary instantaneous power
theory) method basically consists of a variable transformation q, and is related to the reactive power generated by the
from the a, b, c reference frame of the instantaneous power, fundamental components of voltages and currents.

voltage and current signals to the α − β reference frame q~ : is the ac component of the instantaneous imaginary power
[13]. The instantaneous values of voltages and currents in the
q, and is related to the harmonic currents caused by the ac
α − β coordinates can be obtained from the following component of instantaneous reactive power.
equations:
In order to compensate reactive power and current
harmonics generated by nonlinear loads, the reference signal
⎡v ⎤⎥ ⎡ ⎤ ⎡i ⎤
⎡ ⎤ ⎢ of the shunt active power filter must include the values of
⎢
v α
⎥ = [ A] ⎢ ⎥
a

, ⎢
iα ⎥ = [A] ⎢⎢ ⎥⎥
a

(1)
~
p and q~ . [5]In this case the reference currents required by the
⎢ ⎥ ⎢
v⎥ ⎢ ⎥
b i
⎢ ⎥b
⎣v β⎦
⎢⎣ ⎥ ⎣i β ⎦ ⎢
SAPF are calculated with the following expression:
v⎦ c ⎣i ⎥⎦
c

1059
 World Academy of Science, Engineering and Technology 33 2009

~
⎡
i α ⎤⎥ = 1 ⎡⎢vα vβ ⎤⎥ ⎡⎢ P ⎤⎥ ⎡ ⎤
*

⎢ c L
⎡ *
⎤ ⎢ 1 0 ⎥
⎥ ⎢~ ⎥
(7)
⎢ ⎥ ⎢ ⎢ i ca ⎥ ⎢ ⎥
i β ⎥⎦ vα + vβ ⎢⎣vβ −vα ⎥⎦ ⎢⎣q ⎥⎦
2 2
⎡
i α ⎤⎥
* *
⎢⎣ c L ⎢ * ⎥ 2 ⎢ 1 3⎥ ⎢
= −
c
⎢ i cb ⎥
3 ⎢ 2 2 ⎥ ⎢ * ⎥ (8)
The final compensating currents components in a, b, c ⎢ * ⎥ ⎢ ⎥ ⎢⎣ i β ⎥⎦
c
⎢⎣ i cc ⎥
⎦ ⎢ 1 3⎥
reference frame are the following:
⎢⎣ − 2
−
2 ⎥⎦

vsa , vsb , vsc vα , vβ

Park
Transforma tion
p )
p p reference
vα , vβ
P and q current
Transforma tion calculation
isa , isb , isc Park
iα , iβ q )
Transforma tion q q in
α −β
referance
i*ca ,i*cb ,i*cc current
calculation i *c α , i *c β
in
α −β

Fig.2. Block diagram for the instantaneous active and reactive power

The (SAPF) control strategy involves not only the

production of currents whether to eliminate the undesired
harmonics or to compensate reactive power, but also to (PI)
Controller
recharge the capacitor value requested by VDC voltage in
order to ensure suitable transit of powers to supply the
inverter[8-12]. The storage capacity C absorbs the power
fluctuations caused by the compensation of the reactive
power, the presence of harmonics, and the active power
control and also by the losses of the converter. The average Fig.3. Control of DC Voltage
voltage across the capacitor terminals must be kept at a
constant value. The regulation of this voltage is made by To realize these objectives, a controller as shown in
absorbing or providing active power on the electrical network. Figure.3 is added to regulate the capacitor dc voltage of the
The correction of this voltage must be done by adding the (SAPF). In this circuit, the actual dc capacitor voltage is
fundamental active current in the reference current of (SPAF) detected and compared with the reference value, and the error
[30].
is amplified then is added to the ~
p , the output of high-pass
L
filter in Figure. 2. Therefore, active power allowed into the
capacitor is been changed and the dc voltage is controlled.

1060
 World Academy of Science, Engineering and Technology 33 2009

III. NEURAL NETWORKS FOR REFERENCE SOURCE hidden layers with 12 neurons each, and one output layer with
CURRENT AND DC VOLTAGE CONTROL 3 neurons. The logarithmic activation function is the base of
In this work, the p-q theory is modeled, as depicted in Fig. the two hidden layers neurons, and linear activation function
4, by an artificial neural network (ANN) made up of two for the output layer neurons.

Hidden Hidden Output Weights and bias

layer 1 layer 2 layer
adaptation

i *ca , i *cb , i *cc

vsa,vsb,vsc,vdc
Mean square error
. . computation
. .
i sa , i sb , i sc . .
. .
. .

i *ca , i *cb , i *cc

In1 W1 In1 W1
. .
. .
. .
. .
Inn
Wn ∑ Out Inn
Wn ∑ Out

b b
Neuron with logarithmic activation Neuron with linear
function activation function

Fig.4. Neural network for (p-q theory) modelling

The ANN in Fig. 4 has seven inputs artificial neural network (ANN) was examined through
v ,v ,v ,v simulations. The system model was implanted in Matlab /
( sa sb sc dc; i sa , i sb , i sc ) and three outputs( i *ca , i *cb , i *cc ), Simulink environment. The (SAPF) was designed to
as observed in the p-q theory. The model of the neurons of the compensate harmonics caused by nonlinear loads. The system
hidden layers is represented in Fig.4, where each neuron has n model parameters are shown in Table І.
inputs. This parameter varies in function of the chosen hidden
TABLE І SYSTEM PARAMETERS
layer, where n equals 7 if the neuron belongs to hidden layer
Parameters
1, and n equals 12 if the neuron belongs to hidden layer 2. For Supply phase voltage U 220 V
the neurons of the output layer, n equals 12. Supply frequency fs 50 Hz
The adaptation of the weights (W) and bias (b) in the ANN, Filter inductor Lf 0.7 mH
Dc link capacitor Cf 0.768474 mF
is based, first, on the computation of the mean square error Vdc 850V
(MSE) between the outputs of the PQ technique and those of Smoothing inductor Lsmooth 70 μH
the ANN, and secondly, on the execution of ‘Levenberg-
Marquardt backpropagation’ algorithm [18-24]. A three-phase diode rectifier with an RL load was used as a
harmonic producing load. The load value is (resistance was
IV. SIMULATION RESULTS 10/3 Ω and the inductance 60 mH. or Load apparent power
The performance of the proposed detection method using SL=82VA).

1061
 World Academy of Science, Engineering and Technology 33 2009

(a)
200
0
-200
0 0.05 0.1 0.15 0.2 0.25 0.3

(b)
200
0
-200
0 0.05 0.1 0.15 0.2 0.25 0.3

(c)
200
0
-200
0 0.05 0.1 0.15 0.2 0.25 0.3

(d)
500
0
-500
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (sec)

Fig.5. (a) Simulated phase-a load current waveforms, (b) Simulated phase-a reference current waveforms , (c) Simulated phase-a the supply
current waveforms , (d) Simulated the supply voltage waveforms with a (p-q theory) method

(a)
200

-200
0 0.05 0.1 0.15 0.2 0.25 0.3
(b)
200

-200
0 0.05 0.1 0.15 0.2 0.25 0.3
(c)
200

-200
0 0.05 0.1 0.15
(d) 0.2 0.25 0.3
500

-500
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (sec)

Fig.6. (a) Simulated phase-a load current waveforms, (b) Simulated phase-a reference current waveforms , (c) Simulated phase-a the supply
current waveforms , (d) Simulated the supply voltage waveforms with a (ANN) method

1062
 World Academy of Science, Engineering and Technology 33 2009

TABLE ІІ HARMONIC SUPPLY CURRENT PHASE-A-COMPONENT

Harmonic Supply Current Components

Isa(n)/Isa(1) [%]

n load p-q (ANN)

theory method
5 19.59 0.37 0.28
7 13.56 0.43 0.37
11 8.06 0.17 0.04
13 6.48 0.27 0.24
17 4.38 0.41 0.06
19 3.63 0.21 0.17
23 2.51 0.11 0.06
25 2.08 0.15 0.12
29 1.43 0.21 0.04
31 1.18 0.05 0.08
35 0.82 0.07 0.04
37 0.70 0.03 0.05
41 0.56 0.16 0.03
43 0.51 0.01 0.05
47 0.46 0.11 0.04 Fig.7. Harmonic spectrum of supply current Phase ‘a’
49 0.44 0.05 0.05
THD 26.91 1.05 0.74

900

800

700
Vdc (p-q theory)
600 Vdc (ANN methode)
Vdc(V)

500

400

300

200

100

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time(sec)

Fig.8.DC link voltage regulation comparison between (p-q theory) and (ANN)

TABLE .ІІІ PERFORMANCE INDICES, SYSTEM WITH THE TWO used .Because of the (ANNs) capacities to optimize
METHODS
simultaneously their weights and biases in an on-line training
(ANN) Case (p-q theory) process; this approach improves the (SAPF) performance. The
Case filtering result can be seen in Figure 5and 6. The deformations
Rise Time 0.1488 0.08598 have now been reduced and the harmonic distortion calculated
overshoot 0% 0%
Settling Time(s) 0.15 0.42 up to 2.5 kHz (THD2.5kHz) has been weakened. Although the
Steady-state Error 0.3280% 0.3301% filtering performance especially with the low order harmonics
has been improved, this can be seen in Table ІІ, where the
In simulations the two different identification methods were THD calculated up to 2.5 kHz remains less than the case of

1063
 World Academy of Science, Engineering and Technology 33 2009

(p-q theory) approach. [9] Hui.L, Guohai.L and Yue. S., ‘A Novel harmonics Detection Method
Based on Wavelet Algorithm for Active Power Filter’,Proceedings of the
Figure 8 represents the controlled voltage in the borders of 6th World Congress on Intelligent Control and Automation, June 21 -
the condenser. We compared between the pro-posed approach 23, 2006, Dalian, China.
(ANN) and the case of the PI controller which is incorporated [10] Khadkikar.V.,Chandra.A., Singh. B.N., ‘Generalised single-phase p-q
theory for active power filtering: simulation and DSP-based
in (p-q theory) as shown in Table ІІІ It seems clearly that the
experimental investigation’, Published in IET Power Electronics
PI controller in (p-q theory) is characterized by a very low Received on 2nd September 2007 Revised on 5th April 2008 doi:
Rising Time and Settling Time (Tr is equal to 0.08598 s, Ts is 10.1049/iet-pel:20070375.
equal to 0.15 s) compared to the (ANN) case [11] O. Abdelkhalek, C. Benachaiba, T. Benslimane, M. Haidas, 'A Novel
Theory of Reference Harmonic Current Identification Based on the Per
(Tr is equal to 0.1488 s, Ts is equal to 0.42 s). The former Unit System Used for The Active Filters', Istanbul University – Journal
case presents acceptable results at the level of DC voltage of Electrical & Electronics Engineering (JEEE), 2008 Vol. 8, No. 2, pp.
control. 747-757.
[12] O. Abdelkhalek, C. Benachaiba, T. Benslimane, M. Haidas, ' A Novel
Theory of Reference Reactive Current Identification Based on the Per
V.CONCLUSION Unit System Used for The Active Filters ', Istanbul University – Journal
of Electrical & Electronics Engineering (JEEE), 2008 Vol. 8, No. 2, pp.
The work presented in this paper makes a significant 759-767.
contribution to identification and control strategies in order to [13] A. TAHRI, A. DRAOU “Instantaneous Active and Reactive Power
improve the (SAPF) performance. The novel approach is Measuring Method in Three Phase Power System” Leonardo Electronic
Journal of Practices and Technologies, Issue 6, January-June 2005, p.
based on intelligent neural techniques, has been proposed. The
17-28.
performance of the proposed (ANN) was verified through [14] S. A. Soliman, K. El-Naggar, and Al-Kandari,"Kalman filtering based
simulation studies with Matlab and confronted with classical algorithm for low frequency power systems sub-harmonics
technical. The complete (SAPF) structure has been implanted identification," International Journal of Power Energy Systems, vol.17,
1997.
to compensate harmonics caused by nonlinear loads. [15] Matthew.A. G.,‘A Comparative Analysis Of Proportional-Integral
At this level, comparative studies between the neural compensated And Sliding mode Compensated Shunt Active Power
approach and; one of the most conventional techniques used Filters’, A Thesis Submitted to the Faculty of Mississippi State
University in Partial Fulfillment of the Requirements for the Degree of
to extract the harmonic component of the load current to Master of Science in Electrical Engineering in the Department of
produce reference currents; (p-q theory) have been Electrical and Computer Engineering ,December 2004,Copyright 2005
accomplished. The achieved results can be asserted that all the by Gray, Matthew Alan , 10-18
[16] M. Ahmed ‘Sliding Mode Control For Switched Mode Power Supplies’,
identifying objectives of the harmonic currents could be Thesis for the degree of Doctor of Science (Technology) to be presented
satisfied by the approach based on neural networks. However, with due permission for public examination and criticism in the
the (p-q theory) merit is that the latter contains integrated (PI) auditorium 1382 at Lappeenranta University of Technology,
controller, added to regulate the capacitor dc voltage of the Lappeenranta, Finland on the 14th of December 2004,at noon.
Page(s):32-34.
(SAPF). [17] C. Benachaiba, B. Ferdi, ‘A Comparative Study of IEC 61000 Part 3-2
& 3-4 and IEEE 519-1992 Standards in Low Voltage Applications’,
International Review of Electrical Engineering (IREE) December 2007,
pp. 771-776.
REFERENCES [18] Yongtao,Dai1 Wenjin, Dai2 “Harmonic and Reactive Power
[1] Youssef ,K., “Industrial power quality problems Electricity Compensation with Artificial Neural Network Technology “Proceedings
Distribution,” IEE Conf. Pub1 No. 482, vo1.2. June 2001, 18-21. of the 7th World Congress on Intelligent Control and Automation June
[2] Ping.W.,Zhixiong.Z., Houquan.C., ‘A DSP-based active power filter for 25 - 27, 2008, Chongqing, China.
three-phase power distribution systems’, 2009International Conference [19] M. Rukonuzzaman ,M.Nakaoka “Adaptive Neural Network Based
on Computer Engineering and Technology 978-0-7695-3521-0/09 Harmonic Current Compensation in Active Power Filter “ 0-7803-7044-
$25.00 © 2009 IEEE DOI 10.1109/ICCET.2009.140 9/01/$10.00 02001 IEEE,pp 228 1- 2286.
[3] C. Benachaiba, S. Dib and O. Abdelkhlek, ‘Genetic Algorithm-Based [20] O. Bouhali, M. Berkouk, B. Francois, C. Saudemont, S. Labiod
Self-Learning Fuzzy PI Controller for Shunt Active filter’, International “Solving Harmonics Elimination Problem in Voltage-controlled Three
Journal of Applied Engineering Research V. 1/ N. 2 (IJAER) 2006, pp. Phase Inverter using Artificial Neural Networks” J. Electrical Systems 1-
203-216. 1 (2005): 47-61.
[4] O. Abdelkhalek, C. Benachaiba, M.Haidas, T. Benslimane, 'A new [21] N. K.Nguyen, D. Ould Abdeslam, P.Wira, D.Flieller, J. Mercklé
technique applied to a fuzzy regulator to control the shunt active filter “Artificial Neural Networks for Harmonic Currents Identification in
DC bus voltage', journal "Information Technology and Control (ITC) Active Power Filtering Schemes” 9978-1-4244-1766-7/08-2008 IEEE,
2008" , September 2008 Vol.37 No.3 (ISSN 1392-124X). pp 2696-2701.
[5] Hugo.A, Ramos.C., Aurelio.M., Gary W. C., ‘Real-Time Shunt Active [22] W.Dai, T.Huang, N.Lin “Single-phase Shunt Hybrid Active Power
Power Filter Compensation’, IEEE TRANSACTIONS ON POWER Filter Based on ANN” Fourth International Conference on Fuzzy
DELIVERY, VOL. 23, NO. 4, OCTOBER 2008 Systems and Knowledge Discovery (FSKD 2007), 0-7695-2874-0/07 ©
[6] M. Rafiei, Verification of global optimality of the OFC active power 2007,IEEE.
filters by means of genetic algorithms, Proc. WSEAS Int. Conf. on [23] W.Dai , Y.Wang “Active Power Filter of Three-phase Based on Neural
Systems, Vouliagmeni, Athens, Greece, Jul. 10-12, 2006, pp. 559-564 Network” Fourth International Conference on Natural Computation,
[7] H. Akagi, S. Srianthumrong and Y. Tamai, "Comparisons in circuit 978-0-7695-3304-9/08 © 2008 IEEE.
configuration and filtering performance between hybrid and pure shunt [24] N.Pecharanin, M.Sone, H.Mitsui “An Application Of Neural Network
active fitters", IEEE 38th Conf on Industry Applications, Vol. 2, pp. For Harmonic Detection In Active Filter” 0-7803-1901-X/94 1994
1195-1202, October 2003. IEEE,pp 3756-3760.
[8] S.Rahmani, K. Al-Haddad, . F.Fnaiech. "A Three-Phase Shunt Active
Power Filter for Damping of Harmonic Propagation in Power
Distribution Systems " IEEE ISIE 2006, July 9-12, 2006, Montreal,
 
Quebec, Canada, pp. 1760-1764.

1064
 World Academy of Science, Engineering and Technology 33 2009

R.Dehini : received the stage license degree in electrical & engineering from
the national high school of technical teachings (ENSET) ALGERIA.
Currently, He is working toward the Doctorate degree. His interests are
in electrical power quality.

A. Bassou :received the state engineer degree in Electronic Engineering in

1997 from the University of Abou Bekr Belkaïd de (Tlemencen-Algeria)
and the M.S. degree in 2000 from University of Djillali Liabes (Sidi Bel
Abbes–Algeria). In 2006 he received the doctorate degree from
University of Djillali Liabes (Sidi Bel Abbes–Algeria) (USTO), Algeria
and currently holding the post of Assistant Professor. His current
research and teaching interests are in the areas of Artificial Neural
Networks, telecommunication and signal. Presently he is supervising
doctoral students working in the field of telecommunication and
Artificial Neural Networks.

B. Ferdi :received the stage engineer degree in electrical & engineering from
the University of Boumerdes (INELEC), in 1991 and the MS degree in
2008 from Bechar University ALGERIA. Currently, He is working
toward the Doctorate degree. His areas of interest are hybrid active
power filters, applications of power electronics, and power quality
improvement.

1065