lhouse
http://www.trop.uha.fr

**
Laboratoire LEPSI (Laboratoire dElectronique et de Physique des Systèmes Instrumentaux)
Université Louis Pasteur, Strasbourg
http://www-lepsi.in2p3.fr

***
Laboratory for Integrated Micro Mechanical Systems (LIMMS)/CNRS-IIS
University of Tokyo, Tokyo
http://www.fujita3.iis.u-tokyo.ac.jp/~limms/




ABSTRACT
Harmonic estimation is the foundation of every active
noise canceling method in low-voltage power systems.
Reference currents are generated and re-injected in
phase opposition through an active power line
conditioner. Active Power Filters (APFs) are today the
most widely used systems to compensate harmonics in
industrial power plants. We propose to improve the
performances of conventional APFs by using artificial
neural networks (ANNs) for harmonics estimation. This
new method combines both the advantages of
conventional APF to compute instantaneous real and
imaginary powers and the learning capabilities of ANNs
to adaptively choose the parameters of the power
system. In fact, the separation of the powers is
implemented with an Adaline neural network which
uses a priori known frequencies as inputs. Furthermore,
multilayer feedforward networks are used to
approximate the instantaneous powers and to compute
the reference currents. Simulation results show the
reliability of the method and better performances than
conventional APFs.
I. INTRODUCTION
Due to the increasing presence of loads absorbing non
sinusoidal currents such as regulators, motors, etc.,
harmonic distortions have become a significant issue for
power consumers. These distorsions may have
damaging effects on the equipment.
Active Power Filters (APFs) are proposed to
compensate harmonics in existing power systems [1].
APFs are able to correct the power factor without any
additional equipment.


Rigorous identification of harmonics is crucial in terms
of current compensation performances and several
methods have been developed in recent years. For
example, Kalman filters have been applied with the
need for a dynamic state model [2] . The Fast Fourier
Transform (FFT), which needs a lot of computation
resources [3], has also been applied. More recently,
artificial neural networks (ANNs) have been introduced
as a complement or an alternative to conventional
methods [4]. ANNs have been associated with the Park
vectors representation [5] and can be used with a direct
measure of the current [6].
Our method uses ANNs to compute the instantaneous
real and imaginary powers as described in [1] allowing
thus a precise selection of the harmonics. The proposed
method replaces the conventional Concordia
transformations (direct and inverse) and the
computation of the instantaneous real and imaginary
powers, with multilayer feedforward networks. The
harmonics identification itself is implemented with an
Adaline neural network.
We introduce ANNs in harmonic estimation methods by
considering two aims. Firstly, the adaptability of ANNs
must lead to similar or better performances than
conventional methods for varying loads. Secondly, the
structure must be suited for hardware implementation.
The computation of the instantaneous real and
imaginary powers is presented in Section II. Section III
introduces the ANNs for voltage and current wave
forms estimation. The application of the proposed
method is demonstrated in Section IV. Finally, Section
V concludes the paper. II. COMPUTING THE INSTANTANEOUS
REAL AND IMAGINARY POWERS
The principle of harmonic compensation in power
systems is shown in Fig. 1. The presence of the
nonlinear load introduces harmonics distortions in the
source current I
S
and transforms it in a load current I
L
.
The APF has to identify the harmonics distortions to
restore the initial form of the current.
Active power compensation schemes have two main
parts: the first one generates the reference signals and
the second one carries out the control signals. The
identification strategy is decomposed in several blocks
as detailed by Fig. 2.
At first, we introduce a phase-locked-loop (PLL) to
allow the computation of the instantaneous real and
imaginary powers, whatever the environment, condition
or load [1]. Indeed, the instantaneous real and imaginary
powers computation is only possible if the input of the
identifier is a phase-equilibrated system where voltage
waveforms are sinusoids.
The next step consists in transforming the phase
voltages and load currents into an orthogonal
coordinates system according to the Direct Concordia
Transformation (DCT). This transformation, from a
three-phase system to a two-phase system, simplifies the
mathematical expressions and reduces the
computational costs:

0
1
2
3
1
1
1
2
2
2
2
1
1
. 1
3
2
2
3
3
0
2
2
V
V
V
V
V
V


=


, (1)

0
1
2
3
1
1
1
2
2
2
2
1
1
. 1
3
2
2
3
3
0
2
2
S
S
S
I
I
I
I
I
I


=

. (2)
Instantaneous active and reactive powers, respectively
p
and
q
, are then calculated as:

.
V
V
I
p
V
V
I
q
=



. (3)
Filter
AC Mains
50 Hz

Nonlinear
load
PLL
DCT
(I)
DCT
(V)
Computation of

instantan
e
ous po
wers
ICT
(I)
V
s1,2,3
I
s1,2,3
I

I

V

V

p
q
Filter
+
-
-
+
p
q
I
I
PWM
controller
Inverter and
passive filter
I
ref1,2,3
u
I
inj1,2,3
Identification
p
q
+
-
V
1,2,3
Figure 2.
The AFPs general structure
Computation of

current
harmonic distor
tions

Active
power
filter


Nonlinear
load
AC
Mains
50 Hz
I
S
I
L
Figure 1.
The APFs principle in a power system Instantaneous active and reactive powers can be
decomposed into DC components
p
and
q
related to
the fundamental frequency and into AC components
p

and
q
which represents the terms produced by the
harmonic distortions. Thus,
p
p p
= +
and
q q q
= +
.
Filters are used to separate the terms produced by the
harmonic distortion from the DC components related to
the fundamental frequency. The terms
p
and
q
are
thus rejected. Inverting (3) using the instantaneous
powers
p
p p
= +
and
q q q
= +
leads to:

0
1
1
0
1
s
s
s
s
s
s
s
s
s
s
s
s
I
V
V
V
V
p
I
V
V
V
V
q
V
V
p
V
V
q



=
+






+

,(4)
where
2
2
s
s
V
V
=
+
.
The harmonic distortion for the currents in the
orthogonal coordinates can be identified as

1
s
s
s
s
V
V
I
p
V
V
I
q

=





, (5)
and the following currents can be determined with the
Inverse Concordia Transform (ICT):

1
2
3
1
0
2
1
3
3
2
2
1
3
2
2
ref
p
ref
p
ref
I
I
I
I
I
=

. (6)

These currents serve as reference currents and are
injected equal but opposite into the power plant through
the Pulse Width Modulation (PWM) controller, thereby
canceling the original distortions.
III. NEURAL NETWORKS FOR
VOLTAGE AND CURRENT
WAVEFORMS ESTIMATION
This work proposes a new approach based on the
application of ANNs to the estimation of the magnitude
of the symmetrical components of individual harmonics.
The structure of APFs is kept but each block is replaced
by ANNs as shown by Fig. 3. Its advantage is that the
ANNs estimate the reference currents from direct
measurements of the currents and voltages.
Principle of compensation currents
The phase voltages into orthogonal coordinates
and the instantaneous active and reactive powers are
now approximated with the use of two multilayer
feedforward networks (block 1) which replace Eq. (1) to
(3). An Adaline (block 2) identifies and separates the
AC components
p
and
q
from the DC components
p

and
q
. This Adaline takes a priori frequencies as inputs
and the filter delivers two outputs, estimating the
powers due to the harmonic distortions. The signals
p

and
q
, combined with the voltages V and V are then
converted into reference currents. This transformation,
previously given by Eq. (6), is implemented with a
multilayer feedforward network (block 3). The benefit
of this method over other conventional identifiers used
in APFs is that it is based on instantaneous power
computation. Power signals allow to estimate only one
continuous component rather than several alternative
components which is more difficult.


s1
V

s2
V

s3
V

I
s1
I
s2
I
s3
p
q
p
q
p
q


I
ref1
I
ref2
I
ref3
+
-
Block 1
Block 2
Block 3
Figure 3.
Neural network identifier
V
V
-


1
sin 6
wt
cos
nwt
+
-

1
sin 6
wt
cos
nwt
+
-

+ The implementation of the transformations with
feedforward ANN before and after the filtering are
presented in [7]. In this paper, we focus on the filtering
and harmonic estimation stages.
Adaline for harmonic estimation
The Adaline is a simple dynamical learning system by
means of a linear combination of time-dependent
signals. Introduced by Widrow [8] with the LMS (Least
Mean Square) learning rule, the Adaline is now widely
used in signal processing theory for signal estimation
and prediction. Recently, it has emerged as a new
harmonic estimation technique [6, 9]. We will use this
principle and introduce
a priori frequency knowledge to
identify the harmonic distortions. Based on an Adaline,
the compensation scheme can then adapt itself to
changes to any load current waveform.
As nonlinear loads are present in a power plant, load
current waveforms are nonsinusoidal. E