Advanced technique for partial discharge detection and analysis in power cables

A. Cavallini1, G. C. Montanari1, D. Fabiani1, L. Testa2*

1
DIE-LIMAT, Università di Bologna, Bologna (Italy)
2
TechImp, Bologna (Italy)
* E-mail: [email protected]

Abstract paper cables (e.g. high dielectric strength and

Polymeric cables are the most common choice resistivity, low dielectric losses and
for transmission and distribution in populated permittivity, low thermal resistance, chemical
areas, as well as for the connection of stability). The easy installation due to the
components (transformers, GIS, etc.) in new absence of oil in the system (resulting in lower
power stations, replacing paper-oil cables. costs) made the polymeric extruded cables to
Indeed, they do possess superior characteristics replace oil-paper cables in many plants during
with respect to oil-filled cables. As an example, the last years.
they do require lower maintenance, they are On the other hand, polymeric power cables
lighter (easier to deploy) and, in case of (especially those made of cross linked
breakdown, they are not flammable and do not polyethylene) are less resistant to partial
cause polluting leaks. The drawback of discharges (PD) and to water adsorption than
polymeric cables is their scarce ability to oil filled ones. PD, that can originate from local
withstand partial discharge activity over long defects (cavities, protrusions, contaminant),
periods of time. Therefore, partial discharge mostly in cable joints and terminations, as
measurements performed soon after cable consequence of bad installation, can lead to
laying (generally during the 1-hour AC breakdown of the insulation. The major and
withstand test) or on aged cables are getting a most effective tool to detect local damage,
key tool to ensure reliable performance of defects, and/or localized aging processes in
transmission lines. extruded cable systems is, as is well known, the
This paper introduces a novel technique for measurement and analysis of PD [1]–[2].
partial discharge detection, processing and Particularly, in the case of the HV polymeric
diagnosis. The technique is based on the use of power cables (150 kV to 400 kV), measurement
the information borne by the partial discharge of PD is performed just after the cable system
pulse waveforms. It enables noise rejection and installation (after laying test) in order to detect
dealing with contribution of different sources possible flaws of the insulation system that
separately and, therefore, allows accurate could bring to the breakdown of the system
diagnosis to be carried out. The Measurement during the commissioning HV test (usually
results enable, moreover, localization of partial performed applying a voltage higher than the
discharge sources to be performed. Some rated one for a short time interval, typically one
practical examples of on-field partial discharge hour), or could seriously damage the system
measurements and partial discharge source and lead to the breakdown in the earlier stage of
localization, performed applying the here the service. On the other hand, PD
presented technique, are reported and measurements are also more and more required
commented. in MV cable networks, when failure rate is
growing due e.g. to the increasing age of MV
Introduction: cable systems in most installations around
High voltage (HV) and medium voltage (MV) world.
polymeric extruded power cables have achieved This article describes the application of an
insulating properties comparable to those of oil- innovative PD detection and analysis tool to the
diagnosis of HV and MV polymeric cable
systems. The presented tool allows to separate
∫ (τ − t ) s(τ ) dτ
T 2 2
the signals generated from the PD inside the T= 0
0
object under test from signals coming from
other objects and from the background ∞ 2 ~
∫ f 2 s (τ ) S ( f ) df
2
electromagnetic noise, on the basis of the pulse W=
0
waveform characteristics. After the PD pulses
are separated from the other signals, the ~
where s, ~s , and S are the pulse, with time
identification of the PD source is performed in length T, the normalized pulse (Euclidean
an automatic way by an identification software, norm=1), and the Fourier transform of the
which is based on fuzzy logic algorithms [3]- normalized pulse, respectively. Quantities τ and
[6] and provides information about the type of f are integration constants in the time and
the defect of the insulating system that frequency domain. From a geometrical point of
generates the PD. By the separation technique, view, t0 is the gravity center of the pulse,
the fuzzy-logic tool is applied to homogeneous whereas T and W can be regarded as the
classes of pulses (i.e., coming from a single standard deviation of the pulse in the time and
source at a time), thus providing accurate frequency domains.
identification results. Results of a PD Once projected in the so-called T-W plane,
measurement performed off-line on a cable pulses coming from PD or noise, as well as
system, rated 220 kV, just after laying, is pulses belonging to different PD source types,
presented and used as example of the system might form different clusters corresponding to
application. PD-generating defect identification pulses characterized by similar waveforms.
and location are also discussed. Finally the Therefore, different clusters of pulses on the T-
results of online PD measurements, performed W plane, end up representing groups of pulses
on a MV cable system, are reported together having a common source. An example of this is
with the results of localization achieved by the shown in Fig. 1, where three types of pulses
Time Domain Reflectometry (TDR) technique, (detected in a high voltage cable system) can be
applied to a set of data obtained from the observed: (a) pulses with low value of
measurements by the separation technique. The equivalent frequency and high duration (e.g.
measurement reported in this paper were IGBT commutation pulses), (b) pulses with
performed by the TechImp PDbase and intermediate value of equivalent frequency
PDcheck systems. (possible PD pulses) and (c) pulses with high
value of equivalent frequency (external
PD Detection and Identification electromagnetic noise). Eventually, typical
The separation of PD pulses from different representations, such as PD height-phase
sources and of PD pulses from noise is based patterns, can be built up for the classes so that
on a clustering technique that relies on a time the acquired PD pattern can be divided in
frequency, T-W, transformation applied to each subpatterns showing pulse height and phase
recorded pulse. Two quantities (the equivalent coming from a single-source type and/or
time length, T, and the equivalent bandwidth, location (Fig. 2). It is noteworthy that pulses
W) are extracted from each detected pulse belonging to noise will differ, in general, from
according to the following expressions [7]–[8]: pulses generated by PD sources.
With regard to cables, it has been observed that
s (t ) = s(t ) / ∫ s(τ )dτ
T
~ pulses coming from different locations belong
0
to different clusters in the T-W map. In fact,
even in cases where they are generated by the
s (τ )dτ
T
t0 = ∫ τ ~ same type of defect, they face different transfer
0
functions between the source and the sensor [9].
 In particular, pulses generated by different
sources located along a cable route tend to lose
a frequency content (lower W values, larger T
c values) when they have to travel from longer
and longer distances to reach the measurement
b point [9].
The T-W map can constitute, therefore, a
fundamental tool for noise rejection, PD
activity separation, and even defect localization
in cable systems, as discussed later.
Once different phenomena have been
separated, a proper identification, to single out
the type of the defect generating PD, can be
achieved for each data subcluster. Automatic
a) identification is based on fuzzy logic algorithms
[3]–[5] applied to statistical markers extracted
from the distribution of the pulse quantities
(e.g., height, phase, and repetition rate) of each
separated class. In such a way, identification
becomes highly effective, because it is applied
to homogeneous data sets. Noise can be
b)
recognized through statistical algorithms (for
example based on correlation techniques) and
then removed [3].
Identification of PD source for each subclass
relies on a multi-level procedure [4]. The first
level provides a broad recognition stage based
c) on three fundamental categories, internal,
surface, and corona PD, which are defined as
Figure 1: Example of feature extraction of different follows. Internal PD: discharges occurring in
pulses. The differences, in terms of pulse shapes, air gaps surrounded by solid dielectric or solid
result in different positioning of the pulses in the dielectric and metallic electrodes involving
equivalent time-length/equivalent bandwidth plane. significant components of electric field
orthogonal to electrode surfaces. Surface PD:
discharges that develop on surfaces of solid
insulating materials, including solid-solid and
solid-liquid material interfaces, which involve
significant field components tangential to the
surface. Corona PD: discharges produced in
open air (gas) originating from a metallic
object.
These categories are general enough to achieve
identification with good likelihood and success
rate in most cases, but they are also an initial
Figure 2: PD data processing. Example of two aid for risk assessment. In fact, it is generally
superimposed phenomena consisting of PD activity true that internal discharges are more harmful
and noise. Feature extraction and separation (T-W than surface discharges, and both are more
map), pulse classification (Class.), and identification harmful than corona discharges. According to
are carried out sequentially. the fuzzy nature of the inference system, a
pattern may belong to all three categories, Clearly, the pattern relative to the entire
although with different degrees of likelihood, acquisition is the result of the overlapping of
ranging from 0 to 100%. This means that a different activities. Specifically, Cluster A is
certain pattern can be traced back to a defect associated with PD occurring along the
whose nature might be considered intermediate connections between the HV bushing and the
among the three different categories. resonant test set (in fact, the equivalent
Two further categories, i.e. noise and invalid frequency content is quite low). Clusters B is
data, have been devised to identify and filter due to noise caused by electronic component
out possible noise, taking into account that switches in the resonant test set (correlated with
some acquired data set may be inconsistent or test voltage). Cluster C is relevant to high
not related to PD [3]. The second identification frequency background noise (uncorrelated with
level deals mainly with defect position within test voltage) because of distribution circuits
the insulation, that is, the degree of closeness to installed in proximity to the cable system under
HV or ground electrodes, and contains a routine test. Cluster D is due to PD in joint #5. It is
that is able to detect the presence of an worthwhile mentioning that phenomena A and
electrical tree, once the first level has D are overlapped thus, without any separation
recognized the defect as internal [4], [10]. The tool, PD would have gone undetected, since
third level is finally able to provide specific their maximum amplitude is lower than the
indications on the nature of the defect noise level.
generating PD for a given family of The phenomena corresponding to Clusters A,
apparatuses, such as rotating machines, B, and C were detected from all of the
transformers, cables [11]. For cables, the main measurement points. In particular, the activities
output categories are internal interface, external marked as A and B were subjected to signal
surface, and internal cavity. attenuation and distortion, because the pulses,
generated by these phenomena, were travelling
Field Measurements on HV cable and from the power source to the detection point (in
localization the case of detection from joint #5, the signals
The results of on-site tests performed off-line had travelled almost 3 km); Cluster C was
on a 220 kV cross-linked polyethylene (XLPE) detectable in every location, almost unchanged.
cable system are presented in the following. On the contrary, Cluster D was detected only at
The system under test was composed of 15 the link box connected with joint #5 and was
cable spans, 14 sectionalized joints, and two characterized by PD pulses with a large
outdoor terminations and was tested at 146 kV equivalent bandwidth. Therefore, the source of
(phase to ground voltage) at 37 Hz using a this PD contribution was attributed to a defect
variable frequency resonant test set (VFRTS) inside joint #5. The output of the fuzzy
and acquiring PD signals from a CT (Current identification tool applied to the subpattern
Transformer) installed in the link boxes corresponding to the cluster D is reported in
(bandwidth of the sensor and detector up to 40 Fig. 4. As can be seen, the fuzzy algorithm
MHz). As a whole, the cable system proved to provides a mixed response of “internal” and
be PD free with the exception of a single joint “surface” with a likelihood of 77 and 23%
(#5), where PD activity was detected below the respectively, but the tree alert (not shown in
noise level. The results relevant to this joint are Fig. 4), based on an algorithm aimed at
shown in Fig. 3. It provides the whole PD recognizing the evolution to an electrical
pattern detected at Joint #5, the relevant treeing from internal discharges, did not
classification (T-W) map and the subpatterns provide any warning. This identification is
relevant to the four separated clusters (A, B, C, compatible with a discharge at the interface
and D) together with examples of characteristic between a joint and the cable insulation. This
pulses. kind of discharge can occur where there is an
incorrect joint assembly.
 Entire Acquisition Subpattern Pulse ID
A
PD from HV
source

A
B
Classification Map Voltage
correlated noise
B

C B
D C
A
High frequency
noise

C
D
PD in a Joint

D
Figure 3: Sub pattern representation of classification map single classes

Localization
The acquisition of pulse waveforms allows a
more effective approach to condition-based
maintenance to be performed, not only through
Figure 4: PD source identification relevant to the
identification of the PD source, but also by
subpattern D of Fig. 3 localization of insulation flaws. Generally, in
extended HV systems (e.g. HV cable systems),
The corrective action following the discovery the same PD phenomenon may be detected
of such PD was joint replacement. PD from different measurement locations. The T-W
measurements were then carried out again on map separation allows to obtain homogeneous
the newly installed joint, in this case with no (i.e. relative to the same PD phenomenon) data
indication of PD phenomena. The defective to be compared. Generally, the location
joint was analyzed in the laboratory, and showing, on average, larger magnitudes and W
forensic observations showed evidence of values, and lower T values is the one where the
incorrect assembly, with the edge of the HV defect can be found. Moreover, propagation
cable electrode just outside the shielded zone. models can be used to fit the measurement
This finding agrees with the indication of results and to obtain more accurate indications.
interface discharges provided by the In case of short cable systems, TDR can be
identification system. used to assess the exact location of a PD source.
The system described in this paper is able to A
record the pulse waveforms with various time
length (1 to 50 µs), so both the PD pulse and its
reflections, generated in correspondence of the
propagation impedance discontinuities, can be
recorded in the same oscillograph. The analysis
of the time lag between the PD pulse and its
reflections provides the distance between the
PD source and the measurement point.
Eventually, in long cable lines, the arrival time B
technique can be used to localize PD sources. It
consists of the simultaneous detection of PD
pulses from different measurement locations;
the analysis of the time lag among the recorded
pulses allows to estimate the defect position.
The T-W mapping technique can be used, also
in this case, to separate noise from PD and PD
pulses due to different sources, allowing for a
more comprehensive analysis of cable system C
state.

Example PD detection in a MV cable system

and TDR localization
Here, an example of PD detection, separation
on the T-W plane identification and localization
is reported, applied to the case of on-field PD
measurements performed on a medium voltage D
power cable system. The measurements were
performed online. A part of each cable has
polymeric insulation (EPR), the other part is
paper-oil insulated, with a transition joint Figure 5: Measurement results performed on an
online medium voltage cable. A) Entire pattern
between the two parts. Partial discharge
acquisition; B) Classification map; C) Sub-pattern
measurements, performed by the described corresponding to the red cluster on the classification
method, showed that an internal partial map; D) Output of the automatic identification.
discharge activity was present in the insulation
system. The measurement results are reported An example of pulse, extracted from those
in Fig. 5, together with the PD pattern of the belonging to the red cluster of Fig. 5.B, is
detected phenomenon, obtained from the entire reported in Fig. 6.A. An averaging process,
acquisition by classification on the T-W map, performed on a homogeneous cluster of pulses,
and with the output of the automatic allows to reduce the noise level and to make
identification system. clearly visible the original PD pulse and its
reflections at the cable terminations, as shown
in Fig. 6 (wavelet transformation can be also
applied, when pulses are distorted
significantly).
 PD pulse First reflection Second reflection
A

t1 t2 t3

Figure 6: Results of the PD localization by the TDR technique, performed on a homogeneous cluster of
waveform. A) Average pulse waveform; B) Localization of the PD source.

The analysis of the time lag among the PD frequency analysis prove to be effective for the
pulse and the reflections, as shown in Fig. 6.B, separation of different sources of PD, but also
indicate that the PD source is located at about for PD location. The latter can be strengthen by
200 m from the terminal 1 (detection point), the Time Domain approach, which, however,
corresponding to the exact location of transition can be strengthened significantly using the
joint. It is noteworthy that almost one month time-frequency separation of PD pulses. In such
after these measurements, the cable failed a way, in fact, only pulses coming for a certain
exactly in the joint indicated as affected by PD, source, which is evaluated as internal to the
after having shown a growing amplitude trend. cable system, can be processed.

Conclusions REFERENCES
The PD assessment procedure here discussed [1] Partial Discharge Detection in Installed HV
can be very effective for on-site cable
Extruded Cable Systems, CIGRE Tech.
diagnosis. Actually, the method of separation
Brochure 182, CIGRE WG 21.16, Apr. 2001.
based on the decomposition of the pulse
[2] J. Densley, “Aging mechanisms and
frequency and time characteristics allows
diagnostics for power cables An overview,”
enhanced noise rejection and PD separation to
IEEE Elect. Insul. Mag., vol. 17, no. 1, pp. 14–
be achieved; artificial intelligence methods
22, Jan. 2001.
applied to homogeneous PD data sets prove to
[3] A. Cavallini, A. Contin, G.C. Montanari,
be very effective for PD source identification.
and F. Puletti, “Advanced PD inference in on-
Thus, the combination of “separation and
field measurements. Part I. Noise rejection,”
identification” leads to improved reliability of
IEEE Trans. Dielect. Elect. Insul., vol. 10, no.
PD evaluation and diagnosis. Therefore the 2, pp. 216–224, Apr. 2003.
presented technique, provides information that [4] A. Cavallini, M. Conti, A. Contin, and G.C.
can be exploited to enhance maintenance Montanari, “Advanced PD inference in on-field
procedure plans. This is fundamental in measurements. Part.2: Identification of defects
reducing operational costs in a utility cable in solid insulation systems,” IEEE Trans.
network. Moreover, not only did the time-
Dielect. Elect. Insul., vol.10, no. 3, pp. 528–
538, Jun. 2003. signals,” IEEE Trans. Dielect. Elect. Insul., vol.
[5] A. Cavallini, M. Conti, G.C. Montanari, C. 9, no. 3, pp. 335–348, Jun. 2002.
Arlotti, and A. Contin, “PD inference for the [9] S. Boggs, A. Pathak, and P. Walker, “Partial
early detection of electrical tree in insulation discharge XXII: High frequency attenuation in
systems,” IEEE Trans. Dielect. Elect. Insul., shielded solid dielectric power cable and
vol. 11, no. 4, pp. 724–735, Aug. 2004. implication thereof for PD location,” IEEE
[6] G.C. Montanari, A. Cavallini, F. Puletti, “A Elect. Insul. Mag., vol. 12, no. 1, pp. 9–16,
new approach to partial discharge testing of HV Jan./Feb. 1996.
cable systems”, IEEE Electrical Insulation [10] A. Cavallini, G.C. Montanari, and F.
Magazine, vol. 22, no. 1, pp:14 – 23, Jan.-Feb. Puletti, “A fuzzy logic algorithm to detect
2006. electrical trees in polymeric insulation
[7] A. Cavallini, G.C. Montanari, A. Contin, systems,” accepted for publication in IEEE
and F. Puletti, “A new approach to the Trans. Dielect. Elect. Insul., vol. 12, Apr. 2005.
diagnosis of solid insulation systems based on [11] A. Cavallini, M. Conti, A. Contin, G.C.
PD signal inference,” IEEE Elect. Insul. Mag., Montanari, and F. Puletti, “A new algorithm for
vol. 19, no. 2, pp. 23–30, Apr. 2003. the identification of defects generating partial
[8] A. Contin, A. Cavallini, G.C. Montanari, G. discharges in rotating machines,” in Proc. IEEE
Pasini, and F. Puletti, “Digital detection and ISEI, Indianapolis, IN, Sep. 2004, pp. 204–207.
fuzzy classification of partial discharge