Digital Signal Processing Reference
In-Depth Information
Fig. 7.7 Basic scheme of CT
saturation detection/
compensation with ANN
ANN
Compensator
A/D
Antialiasing
filter
Further signal
processing
CT
• correction performed is based on the information extracted from the secondary
current during the saturation interval, which reproduces the primary current
waveshape [
3
,
4
],
• correction is based on information extracted from the secondary current during
the non-saturation interval, which aims at determination of the fundamental and
DC component of the primary current [
5
,
11
].
The first step of all the correction procedures is to determine the beginning and
end of each saturation span. It is not a simple task, however a few approaches may
be applied to solve it:
• analysis of wavelet details of the secondary current [
7
],
• comparison of the fundamental component and its scaled 2nd derivative [
2
],
• calculation of the difference between outputs of mean and median filters [
2
],
• application of neural networks for the purpose [
8
,
10
,
12
,
14
].
Among the algorithms for CT correction the ones based on algorithmic
approach as well as novel Artificial Intelligence schemes are proposed.
The method of reconstruction of the secondary CT current on the ground of the
current data during the non-saturation period presented in [
5
] requires that the CT
magnetizing characteristic and its load are a priori known, which is not always the
case. On the contrary, the algorithms developed by the authors of this topic and
given in [
11
] are free of this limitation.
Promising results are also reported for application of Artificial Neural Net-
works. The basic idea of CT saturation compensation with neural networks is
depicted in Fig.
7.7
. The CT secondary current, after low-pass filtering and con-
verting to digital form, is passed to an ANN, at the output of which the recon-
structed primary CT current should appear. The ANN output is further digitally
processed to get certain relay criterion values, the ones needed for decision making
in the particular relay under consideration.
Independent of the ANN version (non-recursive or recursive) the neural net-
work is trained to resemble the non-linear inverse ''transfer function'' of the CT.
As final result the corrected secondary current samples are expected to be pro-
duced at the output of ANN that should be as close as possible to the non-
measurable samples of the unsaturated primary CT current.
The performance of ANN-based correction methods depend to a large degree
on the cases used to train the networks. More details on that one can be found in
Search WWH ::
Custom Search