Sie sind auf Seite 1von 4

Monitoring Harmonic Sources in Distribution System

by Neural Network Estimator


Yuriy Varetsky, Taras Nakonechny
Institute of Energy Engineering and Control Systems
Lviv Polytechnic National University
Lviv, Ukraine
j.varetsky@gmail.com
Abstract - A method is proposed for monitoring harmonic voltage
magnitudes in radial power distribution system with nonlinear
loads. This approach permits harmonic identification in the
system by a measuring instrument installed on the main feeder. It
is shown that an artificial neural network trained on the system
simulated data can be used to monitoring harmonic voltage
magnitudes.

instrumentation that is not generally permanently installed at


most busses. So, network operators do not identify effect of the
harmonic sources in the complicated system because of varying
the power system loads operation. In this sense, it would be
desirable to develop suitable procedures in an online
environment for monitoring and assessing harmonic levels by
limited harmonic meters.

Keywords - distribution system; power quality; nonlinear load;


time-varying harmonic; neural network; modeling.

Some methods [2-4] have been proposed to solve the


problem of adequate estimating the harmonic sources in the
condition of under-determined harmonic state by traditional
mathematics. On the other hand, some approaches based on
artificial intelligence, such as artificial neural networks (ANNs)
and fuzzy clustering, has been used to solve problems of
identification and location harmonic sources in distribution
systems [5,6].

I.

INTRODUCTION

Due to the rapid development of electronic and


semiconductor devices, harmonic problems have become one
of a major concern for present day engineers. Harmonics are
caused by nonlinear loads, especially power converters in
residential, commercial and industrial loads. The harmonic
current generated by any nonlinear load flows from the load
into the power system. This current causes a voltage drop
through the system. The harmonic voltage combines with the
commercial frequency voltage producing a distorted power
system voltage on the same feeder or on the adjacent feeders.
This voltage distortion may result in more harmonic currents
being produced as other linear loads experience the distorted
system voltage. It can also result in parallel or series resonance
at network substation busbars.
In response to this challenge, most electric supply
companies often assess the quality of electric power supplied to
customers through measurement programs, to determine
whether harmonic distortion is within limits of existing
standards. Field measurements clearly indicate that voltage and
current harmonics are time variant due to continual changes in
system configuration and load conditions [1]. Measurement of
harmonic distortion in the system requires specialized

Neural network computing is a method of using a large


number of simple parallel processors to classify or associate
input patterns with prescribed (so called, learned) output
patterns. This approach can be adapted to recognizing
behaviors in electrical networks where exact functional
relationships are not easily defined. In a paper [6] a structural
neural network with multiple parallel two-layer feedforward
nets was trained to identify the magnitude of harmonic sources
in power systems with nonlinear loads and relatively few
permanent harmonic measuring instruments. A method of
constraining the neural network output to conform to direct
harmonic measurements was proposed.
This paper presents the use of neural network to estimate
harmonic levels in distribution system with limited set of
harmonic measuring instruments. To be useful in identifying
harmonic sources, the neural network must be trained to
classify the magnitude of harmonic sources with patterns of
available power system state data, which have not simple

functional relationships to magnitudes and angles of the


harmonic sources. In the proposed approach the classifications
have been learned by data patterns obtained from simulation of
the system. Modeling of the system has been carried out by
Electromagnetic Transient Program (ETP), where real system
parameters and load characteristics were included. The neural
network was trained to estimate the harmonic voltage levels at
non-instrumented busses. The input data set of the neural
network consists of active and reactive powers at the buses and
harmonic distortion indices at permanent metering bus. The
outputs, where each from a separate neural network, are
numbers representing magnitudes of fifth, seventh, eleventh
and thirteenth voltage harmonics at non-instrumented busses.
The accuracy of the neural network was tested by non-trained
input-output data set obtained from power system operation
simulations.
II.

DISTRIBUTION SYSTEM DESCRIPTION

A one-line diagram of a described radial system is given in


Fig. 1. The system is supplied from 110 kV bus S0 of a high
power transformer substation. Nonlinear loads of the system
are six-pulse static power converters that are connected to
10 kV buses of S2 and S3 substations. It should be noted that
the buses are known to be variable harmonic generators, but
they are not permanently instrumented for harmonics, although
portable instruments can be connected to them to get training
data for the neural network. The same buses contain linear
loads as well.
Line, transformer and load characteristics are given in
Table I - III. Substations of the described system are provided
with conventional measurement instrumentation which is
permanently installed at substation buses including watt and
var meters. It is assumed that only line L2 of the network is
permanently instrumented for harmonic measurements. System
operator can observe load-time profile in the network by
telecommunicational visualization of the measurements. Linear
load of the substations changes according to daily load demand
curve of the distribution system. While nonlinear loads at the
10 kV buses of the S2 and S3 substations change with
stochastic demand of more rapid time varying industrial
processes. So, there is not clear corresponding between power
and harmonic variations with time.

SL2 , U5-1 ,THDU-1

S0

L1

L2

S1

S2

S3

L3

35 kV
T2

T1

T3

110 kV
10 kV

10 kV

PNL2 ,U5-2

10 kV
PNL3 ,U5-3

Figure 1. The one line-diagram of studied radial system.

TABLE I.

OVERHEAD LINE CHARACTERISTICS

Overhead
Line

Wire cross-section
(aluminum/steel), mm2

Length,
km

L1

120/19

10.1

L2

120/19

9.1

L3

120/19

10.0

TABLE II.
Transformer

Power
rating,
MVA

T1

TRANSFORMER CHARACTERISTICS
Voltage rating of
winding, kV
*

Me

16

115

38.5

T2

10

115

T3

10

115

Hi

Lo

Short circuit voltage, %


*

Hi-Me

Hi-Lo

Me-Lo

11

10.5

17.5

6.5

11

10.5

11

10.5

* Hi, Me, Lo high, medium and low voltage windings, accordingly

TABLE III.

LOAD CHARACTERISTICS

Substation

Full linear load, MVA

Limits of converter load


variations, MW

S1

3.5 + j1.5

S2

3.6 + j1.4

0 5.0

S3

1.6 + j0.8

0 5.0

III.

SYSTEM HARMONIC SIMULATIONS

In practice, training data for neural network could be


collected by portable recording instruments placed at harmonic
sources, which could later be synchronized with appropriate
input signals to find input-output training pairs. But it is
difficult to obtain appropriate field data during neural
networking. To generate a set of data series for neural network
training the distribution system simulations were performed.
Modeling of the system has been carried out by ETP [7] for
varying harmonic load conditions in three-phase equivalent
circuits. Transmission lines in the software are modeled by equivalents. Transformers are considered with their
magnetization system, transformation ratios, winding
connections and winding resistances and leakage inductances.
The linear loads are represented as parallel impedances of
resistance and inductance. The nonlinear loads are treated as
suitable converter topologies with the correspondent resistive
loads. The power system parameters are assumed to be three
phase balanced. The methodology of using software for
calculation of instantaneous values in time domain provides a
sufficiently accurate harmonic simulating the system
conditions. Voltage and current waveform diagrams obtained
by the simulations were presented for the analysis through its
Fourier series. So appropriate powers and harmonic indices
were selected to form training data series.
Example of a set of the simulation results selected for
training fifth harmonic neural net of the distribution system are
presented in Table IV. Values in the columns 6 and 7 were
used as the set of target data while values in columns 2, 3, 4, 5

and 8 were later used as the set of input data in neural network
training procedure.
TABLE IV.
*

SELECTED SET OF THE SIMULATIONS

No.

NL2
W

NL3
W

SL2 *
VA

U5-1*
%

U5-2*
%

U5-3*
%

THDU-1*
%

1
1
2
3
4
5
6
7
8
9
10
11
12

2
1
3
5
7
0
0
0
0
1
3
5
7

3
0
0
0
0
1
3
5
7
1
1
1
1

4
5,4+j1,6
7,2+j2,2
8,9+j2,9
10,5+j3,7
5,5+j1,6
7,3+j2,1
9,0+j2,8
10,7+j3,6
6,4+j1,8
8,2+j2,4
9,9+j3,1
11,5+j3,8

5
0,09
0,25
0,41
0,56
0,09
0,26
0,43
0,58
0,17
0,34
0,49
0,62

6
1,34
3,80
6,30
8,47
0,11
0,34
0,55
0,74
1,41
3,77
6,11
8,21

7
0,12
0,34
0,56
0,75
1,31
3,85
6, 23
8,61
1,50
1,74
2,01
1,93

8
0,56
0,99
1,46
1,78
0,90
1,61
2,23
2,66
1,04
0,95
1,91
1,54

21
22
23
24

1
3
5
7

7
7
7
7

1,95
4,52
6,75
8,80

8,69
8,94
9,15
9,28

2,3
2,11
2,44
3,53

25
26
27
28

2
5
4
6

5
2
6
6

11,6+j3,8
0,65
13,3+j4,5
0,82
15,0+j5,2
0,99
16,6+j6,0
1,13
Testing set
10,8+j3,4
0,59
10,8+j3,4
0,58
13,3+j4,5
0,84
15,1+j5,3
0,97

3,07
6,14
5,57
7,15

6,55
3,21
7,99
7,68

1,56
1,71
2,10
2,22

In the case in question, the information taken to be input


patterns represents powers NL2, NL3, SL2 (as PL2 and QL2) and
harmonic indices U5-1, THDU-1 supplied from the power
network permanently available measurements. The neural
network was trained to estimate the harmonic sources by
voltage harmonics U5-2, U5-3 at buses 10 kV of the power
system substations 2 and 3 accordingly. To increase
information storage capability and aid convergence, the neural
network is divided into multiple parallel networks, all with the
same input, but each has two quantities as output, as shown in
Fig. 2.
Q

L2

PL2
PNL3
PNL2
U 5-1
THD U-1

U 5-2

5 th harmonic
U 5-3

* Labels agree with Fig.1


IV.

NEURAL NETWORK ARCHITECTURE

U 7-2

A neural network is able to map an unfamiliar input pattern


to the closest (best represented) output pattern. In other words,
a neural network can identify the output that best represents an
unknown input pattern in cases where an exact input-output
functional relationship is not easily defined. The input and
output patterns may be binary or analog signals.
A typical two-layer feedforward neural network was used
in this paper [8]. The neural network in general is divided into
three layers: input, hidden and output layers, including nodes.
The nodes are processing units that receive input from their
lower side and deliver output on the upper side. A set of input
signals, comprising an input pattern, is applied to the input. The
pattern is transmitted to the input of the hidden layer through
the weighted network connections. The weighted pattern is
received by the hidden layer units, where the signals are
combined in an activation function. The activation function is
commonly chosen to be close to the signum in order to
resemble the two-state output of biological neurons. The
outputs of the hidden layer units are then transmitted to the
inputs of the output layer units through another weighted
network. The neural network is thus a pattern classifier,
receiving patterns directly from data at the input and delivering
output patterns that will give some quantitative information
about the system. The problem is to train the network by
adjusting the weights so that the proper prescribed
classifications will be made.

7 th harmonic
U 7-3

.
.
.
U 13-2

13 th harmonic
U 13-3

input

hidden

output

Figure 2. The neural network architecture.

layers

The feedforward neural network with back-propagation


error was used in this work. The total number of output nodes
is the same as the number of the system busbars multiply by the
number of harmonics described. Only one hidden layer is used
in the network. The node number for the hidden layer in the
prime harmonic net was taken as (input node number + output
node number)/2 or putting it numerically is (6 + 2)/2 = 4. The
neural network was designed and trained by MathLab software
resources. A tangent hyperbolic activation function was used
for the neurons in the hidden layer and for the neurons at the
output layer the activation function was pure linear function.
Network weights was adjusted to minimize total squared error
S between the set of training outputs Xpi for all patterns p and
the set of the actual outputs Vpi :
S=

( X pi V pi
p

U = U actual U trained .

)2 .

VI.

NEURAL NETWORK TRAINING RESULTS

The applicability of the proposed method to monitoring


harmonic sources in the distribution system is verified by the
neural network training results. Table V illustrates the results
for the fifth harmonic neural network path.
TABLE V.
No.
1
2
3
4
5
6
7
8
9
10
11
12

21
22
23
24

1,34
3,80
6,30
8,47
0,11
0,34
0,55
0,74
1,41
3,77
6,11
8,21

1,95
4,52
6,75
8,80

Trained set

*
5-3

0,12
0,34
0,56
0,75
1,31
3,85
6, 23
8,61
1,50
1,74
2,01
1,93

8,69
8,94
9,15
9,28

Testing set
25
26
27
28

3,07
6,14
5,57
7,15

*
5-2

1,35
3,70
6,17
8,49
0,29
0,39
0,57
0,76
1,36
3,72
6,18
8,39

1,87
4,34
6,82
9,02

*
5-3

Error
%

0,20
0,44
0,65
0,81
1,26
3,78
6,32
8,67
1,39
1,64
1,83
2,07

8,83
9,07
9,31
9,42

U5-2 %
-0,01
0,1
0,13
-0,02
-0,18
-0,05
-0,02
-0,02
0,05
0,05
-0,07
-0,18

U5-3 %
-0,08
-0,1
-0,09
-0,06
0,05
0,07
-0,09
-0,06
0,11
0,1
0,18
-0,14

0,08
0,18
-0,07
-0,22

-0,14
-0,13
-0,16
-0,14

0,15
-0,13
-0,03
-0,04

-0,05
0,1
-0,06
-0,17

REFERENCES
[1]

[2]
[3]

[4]

[5]

[6]

[7]

Estimated set
6,55
3,21
7,99
7,68

2,92
6,27
5,60
7,19

6,60
3,11
8,05
7,85

CONCLUSIONS

A method of monitoring harmonic sources in radial power


distribution system by limited harmonic meters has been
proposed. The proposed method is based on using artificial
neural network in the monitoring procedure. Distinctive feature
of the method is using relative few commonly available
measurements within system substations and collecting training
set through the power system behavior simulating in the neural
network designing. It is evident that the method could be
implemented to more complex radial networks with adequate
choice of neural network architecture.
The results of testing the neural network show that
monitoring harmonic sources in the described system can be
efficiently implemented by single harmonic meter.

RESULTS OF TRAINING THE NEURAL NETWORK

Actual set
*
5-2

(2)

(1)

Adjusting the network weights was carried out by scaled


conjugate gradient back-propagation function. The function has
provided a minimum of iterations of the training procedure.
V.

Actual set in the rows 1-24 in the Table V is taken from the
rows 1-24 in the Table IV (columns 6 and 7). After the neural
network was trained, the power system was tested by
monitoring known harmonic sources, i.e. obtained from
power system behavior simulating but did not included to the
training set. The rows 25-28 in the Table V show results of
estimating harmonic sources for the power system behavior
data which were out of the selected training set. Accuracy of
the harmonic sources neural estimates (the neural network
outputs) for the non-trained input data was found with regard
to appropriate results of the power system behavior
simulations. The error in the Table V is the difference between
the actual and trained (or estimated) outputs:

* Labels agree with Fig.1

[8]

Time varying harmonics: part I characterizing measured data,


Probabistic Aspects Task Force of the Harmonic Working Group
Subcommittee of the T&D Committee, IEEE Trans. Power Deliv., vol.
13, pp. 938944, July 1998.
G. T. Heydt, Identification of harmonic sorces by a state estimation
technique, IEEE Trans. Power Deliv., vol. 4, pp. 569575, no.1, 1989.
M. Najjar and G. T. Heydt, A hybrid nonlinear least squares estimation
of harmonic signal levels in power systems, IEEE Trans. Power Deliv.,
vol. 6, pp. 282288, no.1, 1991.
J. E. Farach, M. V. Grady, A. Arapostathis An optimal procedure
placing sensors and estimating the locations of harmonic sources in
power systems, IEEE Trans. Power Deliv., vol. 8, pp. 13031310, no.3,
1993.
R. K. Hong, and Y. C. Chen, Application of algorithms and artificialintelligence approach fo locating multiple harmonics in distributin
systems, IEEE Trans. Industry Appl., vol. 29, pp. 202208, no.1, 1993.
Y. Y. Hartana, and G. G. Richards, Constrained neural network-based
identification of harmonic sorces, IEE Proc. - Gener. Transm. Distr.,
vol. 146, pp. 325329, May 1999.
A. Ravlyk, T. Gretchyn, Digital complex for modelling of transient
processes in electric circuits, Proc. of III Int. Symp. Metody
matematyczne w elektroenergetyce, Zakopane, 1993, pp. 17-20.
H. S. Mark and R. F. Burch Applicability of Neural Networks to
Industrial and Commercial Power Systems: A Tutorial Overview, IEEE
Trans. Industry Appl., vol.33, pp. 1355-1361, no.5 1997.

Das könnte Ihnen auch gefallen