Documente Academic
Documente Profesional
Documente Cultură
Abstract — Early warning system based on the detection of Electrostatic force Grounded tank
UHF SF6 partial discharge (PD) signals is a necessary means for
the protection of Gas-Insulated Switchgear (GIS) in service as
well as the power system to which it is connected. In order to
ensure the safe and reliable operation of GIS, it is important to
High voltage conductor
adopt an effective diagnosing method, which is able to identify
signals of harmful defects promptly. Unlike approaches such as Metallic particle
Fast Fourier Transform (FFT) and Discrete Wavelet Transform PD sensor
(DWT) based techniques, a novel approach to extract
discriminative features directly from time-domain UHF signals is
introduced in this paper. With the proposed approach, UHF Partial discharge
signal waveforms of different sources are partitioned and Electromagnetic wave
quantified so that the harmful (SF6 PD) and not-so-harmful (air
corona) UHF signals are classified promptly. The investigation is Sensor output
based on the experiment data measured from a 300 KV GIS with
encouraging results, which proves the usefulness of this proposed Figure 1: UHF sensor mounted on the enclosure of GIS
approach.
It is understood that the reliability of the early warning
Index Terms—Gas-Insulated Substation, Partial Discharge,
Waveshape analysis, Pattern Recognition.
system is dependent on the accuracy of the diagnosing system
to identify various UHF signals. Therefore, it is crucial for a PD
diagnostic system to have reliable functions of differentiating
1. INTRODUCTION SF6 PD from air corona and other interferences. Comparing to
The use of Gas-insulated Switchgear (GIS) is being globally the UHF signals of air corona, signals originated from SF6 PD
increased due to its compact size, unaffected by atmospheric contain richer high frequency components as shown in Figure
conditions, low maintenance and high reliability [1]. SF6 is a 2. According to the nature of the UHF signal patterns, it is
nontoxic and nonflammable gas, which overcomes many of the understandable that dominant features of a UHF signal are
limitations of the conventional open-type HV switchgear. In embedded in its large resonance, which occurs about hundreds
addition, it has a dielectric strength two to three times that of air, of nanoseconds long after its initial surge. Meanwhile, it is also
and this ratio increases with increasing pressure [2]. However, well known that a number of Fast Fourier Transform (FFT) and
GIS has problems with the sharp deterioration of the gas Discrete Wavelet Transform (DWT) based techniques have
dielectric strength particularly when its voltage reaches to EHV been widely applied to the classification of various PD sources.
level [3]. The radical deterioration of the GIS insulation is When working with these approaches, FFT and Wavelet
mainly contributed by partial discharge (PD) occurred inside functions have to be used to process and transform the
the GIS chamber. PD occurring in SF6 gas is attributed by many time-domain UHF signals into frequency domain or relevant
causes such as mechanical assembling faults, protrusions on the Wavelet pattern, through which discriminative features are
enclosure and/or HV conductor as well as free metallic calculated. The classification is then achieved by finding
particles left inside the GIS chamber. Hence, PD that occurs in discriminative features based on the transformed patterns [6-8].
the GIS chamber due to the above defects may deteriorate the However, complicated FFT or DWT transformation could be
dielectric strength of the SF6 gas, may lead to the breakdown of avoided, if a feature with adequate discriminative ability can be
the GIS or even the power system to which it is connected. found and extracted directly from time-domain UHF signal
To prevent it happening, the detection of SF6 PD is an waveforms. In such a way, the diagnosing process will be
approach to ensure the safe and reliable operation of GIS as simplified and accelerated. Hence, a new approach to identify
remedy can be done before deterioration of GIS continues. As UHF signal sources directly through the time-domain
for the protection purpose, a PD monitoring system including waveforms is proposed in this paper, which aims to improve the
high sensitive sensors and efficient diagnosing algorithms is efficiency of the PD diagnosing system. The data used for this
attached to every GIS installation in conjunction with other investigation is measured from a 300 kV GIS. Encouraging
monitoring features such as self-checking, trend analysis of the results have been obtained for the successful identification of
gas density and circuit-breaker conditions etc. Figure 1 shows UHF signals that originated inside the GIS chamber.
the placement of an UHF sensor on the enclosure of the GIS.
2
0.000
-0.005
-0.010
0 64 128 192 256
Time Series
Sample Number
0.006
(b)
Am plitude (V)
Amplitude (V)
0.003
0.000
-0.003
-0.006
0 64 128 192 256
Time Series
Sample Number
0.008
(c)
Am plitude (V)
Amplitude (V)
0.004
0.000
-0.008
Apparently, the difference among the waveforms of UHF 0 64 128
Time Series
192 256
Sample Number
signals of air corona, SF6 PD and unknown interference is quite
obvious as illustrated in Figure 2. According to the initial visual Figure 3: Pre-selected signals: (a) Corona; (b) PD in SF6; (c)
Unknown Interference
3
Normalization is another important step after pre-selection. dividing the sampling period into 4 equal sub-periods, each
It is performed by linearly mapping the pre-selected signal contains 64 samples. Likewise, the magnitude axis is divided
samples into the value range of [-1, 1]. It is based on each into 4 equal portions, each covering a normalized range of 0.5.
pre-selection waveform to make sure that the distribution of the
pre-selected signal portion is dependent on the waveshape only
and is thus ready for the extraction of the discriminative feature.
The scatter plots shown in Figure 4 are the normalized
waveforms of those in Figure 3.
1.0
Normalized Amplitude
(a)
Norm alized Am plitude
0.5
0.0
-0.5
-1.0
0 64 128 192 256
Time SeriesNumber
Sample
1.0
Normalized Amplitude
(b)
Norm alized Am plitude
0.5
0.0
Figure 5: 16 windows partition
-0.5
-1.0
0 64 128 192 256
Time Series
Sample Number
1.0
Normalized Amplitude
(c)
Norm alized Am plitude
0.5
0.0
-0.5
-1.0
0 64 128 192 256
Time Series
Sample Number
restricted by the following conditions: the error of the output with respect to each weight in the
network and modifies each weight by a value that is
4 proportional to the error. After being trained, the MLP network
∑R
i =1
i, j = 1 , and Ri , j ≤ 1 is able to classify the pre-processed UHF waveforms of various
types. The schematic diagram of the MLP network is shown in
Figure 9.
SF6 PD contains higher frequency components comparing to
corona UHF signals. As for the entire partition, the number
non-zeros will be more for the SF6 PD signals than those for
other signals. On the other hand, the calculated non-zero values
for SF6 PD signals are smaller in general than those for other
signals due to the relative even distribution of the SF6 PD
signals. This phenomenon can be observed by comparing
sample distributions shown in Figures 7 and 8.