Beruflich Dokumente
Kultur Dokumente
Examensarbete 30 hp
Mars 2016
Svante Karlsson
Abstract
Power Transformer Monitoring and Diagnosis using
Transformer Explorer
Svante Karlsson
1 Introduction 3
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Aim of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Outline of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Transformer Explorer 17
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 System Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.1 Instrument Transformers . . . . . . . . . . . . . . . . . . . . 18
3.2.2 Acquisition System . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.3 Phasor Calculation . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.4 Vector Group Transformation and Zero-Sequence Removal . 18
3.2.5 Extraction of Quantities . . . . . . . . . . . . . . . . . . . . 19
3.3 Transformer Explorer Software . . . . . . . . . . . . . . . . . . . . . 20
3.3.1 LabVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3.2 Program Structure . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.3 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.4 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.5 Display Options . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3.6 Logging of Data . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4 Field Installations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5 Fault Detection Ability . . . . . . . . . . . . . . . . . . . . . . . . . 29
1
3.6 Transformer Explorer Applications . . . . . . . . . . . . . . . . . . 30
3.7 Uncertainty and Accuracy . . . . . . . . . . . . . . . . . . . . . . . 30
3.7.1 Derivation of Uncertainty in a Linear Fit . . . . . . . . . . . 30
3.7.2 Uncertainty in the Estimated Parameters . . . . . . . . . . . 31
5 Analysis Improvements 52
5.1 Control of Data Distribution in Fit . . . . . . . . . . . . . . . . . . 52
5.2 An Alternative Fitting Algorithm . . . . . . . . . . . . . . . . . . . 59
6 Power Loss 64
6.1 Application Examples on Field Installation Data . . . . . . . . . . . 65
2
1 Introduction
Power transformers are one of the most important components in the electric power
system. By making it possible to interconnect different voltage levels, they permit
a economical way to transfer energy over long distances. Power transformers are
also among the most expensive components in the network and they are often not
manufactured until ordered [1]. A sudden failure which means that the transformer
needs to be taken out of service, is often associated with considerable costs for
the owner. To avoid such failures, it is important to detect problems in an early
stage so that a planned outage with necessary repairs and maintenance can be car-
ried out in time. A number of monitoring techniques has been developed during
the past few decades for this purpose. The majority of these methods are only sec-
ondary indicators such as temperature measurements and gas-in-oil analysis. When
a problem is indicated by the monitoring system, the transformer is disconnected
and more advanced diagnostic methods can be adopted to find the actual problem.
Very few of the conventional monitoring methods addresses the fundamental func-
tion of the transformer, i.e., to convert electric energy between two voltage levels
with power losses within acceptable limits and enough impedance to limit effects
of short-circuits. Two important parameters that ensures this function are turn
ratio and impedance; these are traditionally only measurable with the transformer
off-line.
1.1 Background
At ABB a new monitoring and diagnostic tool [1]: Transformer Explorer, is being
developed. The concept uses voltage and current signals obtained from the instru-
ment transformers in the substation to estimate some of the transformers most
fundamental parameters:
Turn ratio
Magnetizing current
Short-circuit impedance
Power loss
all which are strictly related to the transformers fundamental function, and thus
have a significant diagnostic value. By applying the measured quantities to a simple
transformer circuit model it is possible to estimate the three first of the aforemen-
tioned parameters by means of linear fitting. The power loss is simply given by the
difference in power going in and out of the transformer. The central part of the
concept is a software with the same name, which is used for both signal acquisition,
3
analysis and visualization. The current version of the program is written in the
graphical programming environment LabVIEW.
The concept is expected to find use in two different applications: either as a
permanent installation where the parameters are continuously monitored, or using
a temporary installation where the parameters are estimated based on data collected
during the measurement period.
Transformer Explorer has many advantages compared to conventional monitoring
and off-line diagnostics methods. Some of the most important are that no expensive
equipment is needed apart from the all-ready installed instrument transformers and
no outage is required for the installation or during the measurement, which is the
case when measuring both turn ratio and short-circuit impedance with conventional
methods. Further the power loss can be considered as a general problem indicator
and is much faster than secondary indicators, i.e., temperature [1].
4
2 The Power Transformer
To make this thesis understandable for a reader with limited knowledge in the field of
power transformers – this chapter is intended to give some necessary fundamentals.
5
2.2 Transformer Design
2.2.1 Ferromagnetic Core and Windings
The main active parts of a power transformer are the core and winding. The core is
usually build up by thin sheets of high permeability silicon steel which are stacked
together [4]. This is done to prevent eddy current losses.
It is the core that take care of the energy transfer between the two windings by
providing a low reluctance path for the magnetic flux. The high permeability en-
sures a low leakage flux and magnetizing current. The construction of the core could
be divided in two main categories: core type and shell type construction, where the
core type is the far most common for power transformers. All the transformers
studied in this work are of core type. A picture illustrating the salient difference
between the two types is shown in Fig. 2.2.
Figure 2.2: A sketch showing the two core configurations: core type and shell type [4].
The other fundamental part of the transformer is the windings. They are located
concentrically around the core. The conductors of the transformer winding are usu-
ally made of high conductivity copper. It has excellent mechanical properties and
the highest conductivity of all the commercial metals [5]. In order to optimize the
usage of space the conductors are commonly rectangular shaped. The conductors
are insulated with either varnish or cellulose base paper. The transformer windings
are typically constructed in four different ways: layer, helical, disk and foil windings
[4]. All these four types are similar in the aspect that they are cylindrical and have
a rectangular cross section. It is mainly the number of turns and the amount of
current that determines which type of winding that is preferred when constructing
a transformer. For large power transformers however, the ability to withstand short
circuit is considered of higher importance than the losses.
2.2.2 Insulation
Since various parts of the transformer operate at different voltage levels, sufficient
insulation is needed between the parts to avoid electrical breakdown. For example,
the core is usually grounded while parts of the high voltage winding operates at full
voltage.
The insulation in a power transformer could be divided in two parts: Solid insu-
lation and Oil insulation. The solid insulation typically consists of paper and press
board and is used for both electric and mechanical insulation between the windings
and core. Paper and press board are based on cellulose that consists of molecules in
6
Figure 2.3: The left picture shows a winding arrangement with a helical winding closest
to the core surrounded by a disk winding. The outer is a layer winding for voltage
regulation. The right picture shows example of barrier arrangement between windings
[4].
long chains. The cellulose materials undergo degradation over time by breaking up
into smaller molecule chains. The degradation drastically decrease the mechanical
properties of the insulation and it becomes brittle. There is a standardized method
for measuring the degradation process of cellulose: degree of polymerization, i.e.,
the DP number. A newly fabricated cellulose fiber has typically a DP number of
between 1200 and 1400, whereas the value can sink as low as 200 for an old trans-
former.When this happens, the mechanical strength has decreased to such an extent,
that the insulation’s ability to withstand the forces during an external short circuit
no longer may be sufficient. This could of course lead to metal-to-metal contact in
the winding with an internal short circuit as an result. The degradation process is
a growing process that is highly affected by the environment in the transformer [4].
It could be good to mention that the dielectric properties of the insulation stays
almost unaffected during the whole aging process, so without external events the
transformer may operate satisfactory even with a highly reduced DP number [4].
The transformer oil fills two functions in the transformer: cooling and insula-
tion between different parts. The transformer oil itself has a rather high dielectric
breakdown voltage (20-50 kV/mm vs 3 kV/mm for air), but gets excellent dielec-
tric properties when combined with paper. Between the windings and windings to
core the distances with only oil can be comparably large and the field stress can be
high enough to cause breakdown. This is solved by adding barrier systems of press
board sheets as additional insulation. The press board has good dielectric prop-
erties and also provides stability to the winding structure. Two different winding
arrangements with press board sheets can be seen in Fig. 2.3.
7
2.2.3 Tap Changer
Since the losses in the grid is load dependent and varies with time, compensation
needs to be done in order to sustain a stable voltage level. One way to achieve
this compensation is by letting the transformers operate at different turns ratio.
Therefore, most of the power transformers are equipped with tap changers.
Tap changers can be divided in two categories: Off-load and On-load tap chang-
ers [4]. The off-circuit type has a more simple design and is only operated when
the transformer is off-line. The On-load tap changer however is a sophisticated
mechanical device, which is able to change between adjacent winding blocks (1-2%)
without interrupting the current flow. An On-Line tap changer should be able to
perform a very large number of operations during it’s lifetime.
2.2.4 Bushings
The transformer windings needs to be connected to the power grid through the
earthed tank. This is achieved using different types of bushings. The bushings
function is to provide isolation between the conductor and the transformer tank
which is normally at ground potential. As the voltage could be very high, the
connection between bushing and tank is exposed to high electric stress. Since
the insulation medium outside the tank normally is air, the bushing also needs
to provide sufficient insulation distance between the incoming conductor and the
tank. There are two types of bushings: solid and capacitance graded, where the
latter one is used for higher voltage ratings.
8
Figure 2.4: Circuit representation of a current transformer with a connected burden [6].
sections. Just like ordinary full size transformers, instrument transformers are also
subjected to errors associated with magnetizing current and leakage reactance. Fur-
ther, the errors for current transformers are also load dependent, with a relatively
large error at lower loads. To specify the amount of error one can expect, instru-
ment transformers are divided into different accuracy classes specified by IEC and
IEEE [6].
Current Transformer
Fig. 2.4 shows a schematic diagram of a current transformer with its primary
winding connected to a power line. The primary winding usually only consists of
one turn of the main conductor wound around the current transformer core and is
always connected in series with the line. In comparison, the secondary side which
carries the low measurement current consists of many turns surrounding the core.
The current measurement equipment – in Fig. 2.4 represented with a burden – is
connected to the secondary side. The current transformer is constructed to give a
max value of 1 or 5 A at rated current at the secondary side according to certain
standards.
Voltage Transformer
There are mainly two types of voltage transformers: electromagnetic and capac-
itor type, where capacitor type is the most common in high voltage application
(>100kV) [6]. A voltage transformer is typically connected between line and ground.
An important difference to current transformers are that the errors are less load de-
pendent and the variation is small even if the voltage varies with wide limits.
9
Figure 2.5: A circuit diagram showing an ideal transformer together with a voltage
source and a load [7].
reactance and no active losses in the core. This leads to the relationship:
Vp VS
= (2.1)
NP NS
The permeability of the core is finite, i.e. the core has reluctance.
10
Figure 2.6: The practical transformer circuit.
11
Figure 2.7: Cross section of two windings used for Rabins’s Method [4].
reactance is the imaginary part of the short-circuit impedance, this will be depen-
dent on the magnetic leakage flux, and thus the parameter will be strictly related
to the geometrical properties of the primary and secondary windings.
Their are several methods developed for calculating the short-circuit impedance
from given geometrical properties of the winding. One example is Rabins’s Method.
It uses a simplified geometry including core, coils, and yokes of infinite extent
[5].Thus it is ignoring the surrounding tank walls and structures. The method
is still accurate enough to perform calculations of the forces and inductance’s of the
winding. The short-circuit reactance for the windings given by
(D1 + D2 ) M 2
Xw = 0.827 ∗ 10−9 f (t1 + t2 + 3t12 ) α (2.3)
2h Sl
where t is the radial width of shells and ducts, f is the power frequency, h is the
height, M is the total ampere turns of each shell, Sl is the rated power per wound
limb and D is the mean diameter of shells and ducts. The simplified geometry
used to derive the formula is shown in Fig. 2.7. The factor α is the so called Ro-
gowski factor which is a compensation factor for the leakage flux. It compensates
for the non vertical behavior of the field lines. From Eq. 2.3 it can be seen that
the impedance is highly dependent on the distance between the shells and thus the
geometrical properties of the winding.
12
Figure 2.8: Failure location distribution obtained in a survey from Cigré [8]. It is based
on 536 major failures in sub stations transformers. Note that the winding and tap changer
are major failure sources.
13
Figure 2.9: Buckling of the outer winding of a three phase transformer.
14
2.4.2 Core Related Faults
Another failure, however not so common as winding faults is problems related to the
core. Since the purpose of the core is to carry the magnetic flux, a problem could
be visible as an increase of the losses. Core problems often takes it’s visibility as an
increasing in the magnetizing current or additional heat development. The cause of
failure could have many reasons, e.g., damaged sheet lamination that increase eddy
current losses.
Temperature
Gas in oil
Partial discharge
Moisture
By knowing for example the number of tap changer operations performed; the wear
can be calculated and maintenance can be carried out when needed.
Different diagnostic methods could be adopted to more deeply investigate the prob-
lem cause when a warning signal has been received from some of the monitoring
sensors. Some widely used methods are described in the following sections:
Short-circuit Test
These test are carried out by measuring winding resistance and reactance with the
transformer decoupled from the grid. Could give indications of contact resistance
and winding deformation. In addition to this the magnetizing current can be mea-
sured in the same procedure.
15
Dissolved Gas Analysis
Analyzing samples of the transformer oil is a valuable source of information for the
transformers health condition and several analysis methods has been developed over
the years for this purpose. The main principle behind all the methods is that, when
transformer oil is subjected to electrical sparks or severe heat; chemical reactions
occur that generate substances in the oil [9]. One method that has gained much
attention is the Dissolved Gas Analysis (DGA). Samples of transformer oil is then
analyzed in order to find deviant contents of different gases. Gas production in
the oil could be a result of a variety of faults, e.g., disruptive discharges, partial
discharges and core heating, to mention a few. DGA should even be able to diagnose
which type of fault that is present in the particular case and the severity. DGA
could be considered as a standard test method that is used regularly on most power
transformers.
16
3 Transformer Explorer
3.1 Introduction
As an answer to the markets increasing interest in on-line monitoring and diagnos-
tics of power transformers; a new on-line monitoring and diagnostic method is being
developed at ABB Corporate Research: Transformer Explorer. The method uses
fundamental frequency signals of currents and voltages obtained from ordinary volt-
age and current transformers available in a substation to estimate the transformers
performance. Some of the most important information, the so called transformer
fundamental parameters: turn ratio, short circuit impedance and power loss, are
extracted from the transformer equivalent circuit by means of the current and volt-
age signals. The method was first tested and verified using numerical computing
programs such as MATLAB with data collected from several transformers in ser-
vice. Later, a complete program was written in LabVIEW for both data acquisition,
analysis and visualization. It is this version of the program that will be used in this
thesis.
17
3.2.1 Instrument Transformers
The start of the signal chain is the current and voltage transformers. The analog
signals are taken from the secondary circuit of the current and voltage transformers.
As described in the theory part, instrument transformers are divided into different
accuracy classes, according to IEC and IEEE standards [4]. The instrument trans-
formers used in most substations is of protection class, which means that they
usually have an accuracy around 1 %.
18
one of them. Depending on the winding configuration of the analyzed transformer,
vector group transformation might need to be performed to obtain the right voltage
and current phasors for the transformer model.
Transformer Explorer does also have the option to remove zero-sequence com-
ponents (see [2] for definition of sequence components) from the phasors, before
they are used for fitting. Whether they should be removed or not depend on the
particular transformer configuration. It is however convenient in most cases when
estimating the impedance, since the transformers zero-sequence impedance is dif-
ferent from the positive and negative equivalents. The presence of zero-sequence
components could be for two reasons: there exist an actual phase imbalance, either
caused by the feeding grid itself or by a load imbalance on the supply side of the
transformer, or mismatch between the sensors, e.g, instrument transformers, caus-
ing an artificial imbalance. The zero-sequence component could also be excluded
from the current phasors when estimating the turn ratio for some transformer types.
This is due to the fact that a delta winding blocks the zero-sequence component on
one side.
x y a b
n I1 I2 1/n I0
Zw I2 /n V1 − nV2 Zw Z1 I0
19
The Impedance Fit
The voltage drop ∆V is the voltage difference between V1 and V12 , i.e., the input
voltage and the output voltage multiplied by the turns ratio. This is mainly de-
pendent on the load current, since I0 is relatively low. ∆V could be used to write
the next linear equation, which ia basically obtained by writing ∆V as a function
of the load current I2 . This gives the resulting equation, c.f., Fig. 2.6,
I2
∆V = V1 − nV2 = R1 + jX1 + R12 + jX12 ) + R1 + jX1 )I0
n
I2
= Zw + Z1 I0 (3.3)
n
The important quantity from this fit is the short-circuit impedance Zw which is given
in Ω. The short-circuit impedance is usually given on the transformers nameplate
as the percentage voltage drop Uk at rated load. This could easily be converted to
Ω using
Vnominal
Zw = Uk (3.4)
Snominal
This quantity has both an imaginary and a real part. The imaginary part reflects
the leakage flux and the real reflects winding and eventual contact resistance. Any
deviations from this fit is given in kV .
Power Loss
The last quantity that can be obtained is the total power loss. This is not obtained
using linear fits, instead it could simply be obtained by taking the difference be-
tween the power going in to the transformer and the power going out. Described
mathematically using the same notation as in Fig. 2.6, this is
Sloss = V1 I ∗ 1 − V2 I ∗ 2 (3.5)
One problem with this equation is that, since it formed as a difference, it is very
sensitive to errors in the measured quantities. This results in that the power loss
is difficult to measure with any absolute accuracy without applying some sort of
correction to the sensors (see more in chapter 6).
3.3.1 LabVIEW
LabVIEW is a graphical programming environment developed by National Instru-
ments. It is commonly used by scientists and engineers for fast implementation of
data acquisition and measurement systems. A program consists of two windows:
the front panel and the block diagram. The front panel works as the interface from
where the user can control the program. The programming code is written in the
block diagram, by connecting individual icon boxes together in a proper way. Each
box represent a block of underlying executable code that performs a certain task.
20
LabVIEW has several features that makes it an excellent programming language
when it comes to writing prototype programs for scientific applications. Its graph-
ical interface and large library, letting the user to quickly build a program and it’s
ability to divide the program into separate so called subVI where the code can be
executed separately, makes it excellent for debugging.
Figure 3.2: The program structure, the outermost frame is a never ending while loop
followed by an inner case structure. This is the main part of the program which in this
figure shows the default case. This case contains a event structure which responds to user
action.
21
3.3.3 Data Acquisition
The data acquisition is performed by the getNewData.vi, and this provides a number
of possibilities for the user. As described earlier there are three main options to feed
transformer explorer with data: from an ADC, from a waveform file, directly from
a protection relay or by loading previously analyzed files. Using the second option
it can acquire wave form files stored in Comtrade or ManyScope (ABB’s private
format). When acquiring data from a relay it both has the possibility to download
wave-form files, or in case of a newer protection relay, e.g., REx670V2, only phasors
need to be communicated.
It also have the possibility to read from pre analyzed table files in ASCII format.
Each measurement is then a line in this file, containing information of the phasor
given by two quantities: frequency and complex amplitude. The vi uses the a
frequency acceptance test in order to not let trough any phasors with too much
deviation leading to deceptive results when later analyzed. The amplitude could be
given in either complex form or with amplitude and phase angle in degrees. The
user could specify settings for the acquisition in the transformer explorer settings
window shown in Fig. 3.3.
Figure 3.3: The settings window with the input tab selected. Here the user can choose
which type of files that should be analyzed together with settings for reading intervals
and nominal frequency for the frequency acceptance test. Note that the input selector in
this example is set for reading table files.
22
Transformation from line voltages and currents to winding voltage and cur-
rents, as discussed in earlier section. This is performed by a sub.vi, which
transforms the quantities according to the right vector group. This is spec-
ified in the settings window tab Transformer. The vector groups that could
be chosen are YY, YD, DD, DY and the clock options are 1, 5, 6, 11, which
includes the most common configurations.
Since the turn ratio and impedance change with tap position, and the power
flow direction in the transformer could be different; the data has to be clas-
sified according to this properties when analyzed. This is performed by a
sub.vi, which can determine the tap belonging to each class in two ways: ei-
ther from currents or voltages. The latter one is preferably used since the
current method is used for determine the ratio
Transformer explorer will also find the deviations of each new data point from
the fitted fundamental quantities. This is considered as one of the main results
and is important when determining the uncertainty of the fit.
Actions that are performed only when enough data are collected includes of course
the fitting itself. The fitting could be divided into running fits and reference fit. The
running fit is done whenever enough new data are contained in the class and this is
done only with data newer than those included in the last fit. In order to simplify
the monitoring ability, a reference fit is also performed on all data included in the
class after a certain number of running fits have been done. A number of criteria
must be fulfilled in order for the program to make a fit, here the most important
are:
Minimum readings for fit: The class must contain at least this much data in
order to be able to make a fit.
Minimum load range in fit: The class must contain data where the load has
varied a certain specified minimum percentage value of the nominal power.
Minimum load in fit: The minimum load should not be lower than a certain
value, e.g., 10 percent. This is because effects of the magnetizing current will
become more significant at lower loads.
Exclude the most deviating data in fit: The program excludes and throws
away a specified percentage of the data that deviates most from the fit and
makes a new fit. This is done to reduce the impact of disturbed data on the
fit.
The criteria are administrated from the settings window, Criteria tab shown in Fig
3.4.
23
Figure 3.4: The Criteria tab, letting the user apply settings regarding the fitting pro-
cedure.
Status
The status display serves as an overview of the monitoring status. As default display
it is also the first thing the user sees when start running the program. The display
is divided in two sections: present status and fit status. The present status gives
information about the latest measurement with corresponding winding values, e.g.,
voltage and currents for each phase. Latest difference to measurement fit is also
presented for each of the quantities. A picture of the status display is shown in Fig.
3.5.
24
Figure 3.5: The status display.
Winding Values
This display is intended for analyzing the values that are used for the fits, but also
to obtain directly calculated quantities such as power and power loss. The display is
controlled by four pull-down menus: value, component, view and parameter. The
first menu is the most central for this display and the following options can be
chosen:
1. Voltages
2. Currents
3. Voltage ratio: ratio between the high and low voltage side.
4. Compensated voltage ratio: The same as above, but this ratio is compensated
for the internal voltage drop.
5. Current ratio: ratio between the corresponding currents in one winding phase.
6. Voltage drop: The internal voltage drop, i.e. the difference between the high
voltage and the low voltage multiplied by the turn ratio.
7. Current difference: difference between the high side current and the low side
current divided by the turn ratio.
8. Power: The power supplied to the high voltage side and the power delivered
by the low voltage side.
25
For all of these values the choices component and view could be varied. The compo-
nent menu allows the user to view the sequence component for each value, and the
view menu makes it possible to show absolute, real, imaginary or the phase angle.
The last pull-down menu is basically the x-axis on the graph and makes it possible
to plot the values against other parameters than the default: time, e.g., current or
real power.
Fundamental Quantities
This is probably the most valuable display option since it shows the fit results. From
the fit pull-down menu the three obtained fits can be chosen: ratio, impedance and
power, where only the first two will be treated in this work. The second menu
result, where the results from each type of fit can be chosen. Slope and intercept
are the most important here. The view menu is the same as for winding quantities
except for that the phase angle can’t be shown. Both the running and reference fit
is displayed which are shown in Fig 3.6. The reference fit is always displayed as a
fixed value.
Difference to Fit
This display shows the difference between the latest measurement and the reference
fit. It is a fast indicator or eventual problems and can thus be used for change
detection. The Fit and the parameter menus are the same as for the fundamental
quantities. In the pull-down menu it is possible to chose between the deviation
units: std. deviation, % of nominal and actual value.
26
Figure 3.7: The difference to fit display.
27
Figure 3.8: The logging tab in the criteria settings window. From here, different options
can be set for logging of the results. Note the Max memory which by default is set to
1000 data points, this and the Max Files setting serves to avoid memory overflow during
long monitoring sequences. Max memory limits how many measurements that are kept
in active memory and Max Files will activate file deletion if too many files are found.
28
Table 3.4: Name plate data for Field installation 3
Core Problems
Core problem could take its visibility in a change in the magnetizing current or
the active power loss. The magnetizing current is obtained as the intercept of the
turn ratio fit, and should thus be sensitive to core related problems. It is however
important to know that the magnetizing current also is dependent on the voltage
level and this must hence be checked before any actions is taken on an observed
change. A core problem could also be seen in the active power loss. Since the core
losses is not load dependent any change is observed as an decrease or increase in
the offset of the power curve. See section 2.4.2 for more information on core related
faults.
Winding Deformation
As discussed in section 2.3, the winding geometry is directly related to the short
circuit impedance. A winding deformation, e.g. caused by a short circuit, will cause
a permanent change in the short-circuit impedance. Such a problem could thus be
detected by the impedance fit.
29
Contact Problems
Contact problems, such as loose connections and coking tap changer contacts with
additional heat development as result, will cause a change in the winding resistance
and should thus be visible in the real part of the impedance.
Eddy Currents
Could be present both in the core, surrounding support structure and tank; in the
latter case they may be load dependent. Eddy currents could therefore be detected
by observing the magnetizing current and active power loss.
30
How is the uncertainty of the estimated quantities affected if there is noise present
in the measured values x and y? This is rather straight forward if one considers
the exact solution for only two pairs of measured values: (x1 , y1 ) and (x2 , y2 ). The
slope and intercept is given by:
y2 − y1
k= (3.7)
x2 − x1
and
x1 y2 − x2 y1
m= (3.8)
x1 − x 2
If one of the y’s is increased by an infinitesimal change, the change in k will be
dy
|∆k| = (3.9)
x1 − x2
σ(δ∆VHV )
∆Z = (3.11)
max(IHV ) − min(IHV )
31
4 Short Time Measurements
A possible application for transformer explorer that could give valuable information
for a transformer owner is short-time measurements using a temporary installation.
Two main questions that needs to be adressed are then:
How and what results should be presented to provide the most valuable in-
formation for the transformer owner?
What performance could be expected when the tool is used for short-time
analysis? Is the assumed uncertainty valid?
4.1.1 Method
The data used for the study was taken from the already existing data-base. To
cover eventual differences between different transformer types – data from three
transformers were used: Field installation 1, Field installation 2 and Field instal-
lation 3 as mentioned in section 3.4. The collected data available for the survey
consists of so called FundLog files, where each file corresponds to data for a period
of two days. For simplicity, one FundLog file was used for the short period and 4
for the longer period. That corresponds to data collected during a two day, and a
8 day measurement (roughly one week). The two time periods will here after be
refereed to as day and week.
32
The outcome of a short-time measurement with Transformer Explorer should of
course contain the three fundamental quantities: turn ratio, impedance and power
loss. This survey will deal with the first two, since the power loss needs special atten-
tion (see chapter 6). Other results that could be extracted and should be included
in the study is the magnetizing current. This because the magnetizing current, in
addition to the power loss, also could provide valuable information concerning the
transformers condition.
A result of any measurement should always be presented with a proper uncer-
tainty – a measurement value could be considered worth nothing unless the un-
certainty is properly specified [11]. The uncertainties in the fitted quantities were
derived analytically (see Section. 3.7) and thus, their validity for short-time mea-
surements needs to be checked.
The entire period used for each transformer could be seen in Table 4.1. The
sub-periods (days and weeks) were chosen randomly between these intervals. 10
sub-periods of each kind were used to get some statistical validity.
Table 4.1: Time periods from where the data was taken for each transformer.
One fundamental difference when the program is used for short-time measure-
ments instead of continuous monitoring, is that the program should be able to
make a fit using all data that has been collected. This functionality is not needed
when the program is used for monitoring since then a fit should only contain data
not older than the latest fit. Since the concept initially was developed for monitor-
ing purposes this functionality was missing. The problem was solved by adding an
extra case to the program structure: doFullFit. The user should be able to make
a full fit when all data has been read. Therefore a button was added to the front
panel which will be visible when the continuous acquisition has been turned off.
Another feature that was added was the ability to read a single file instead of an
entire file list, which was the only option in the initial version of the program. In
order to collect and analyze all the results from all the measurements in an efficient
way a MATLAB1 program was built. The program should be able to accomplish
the following:
1. Read the desired information from each RefFitLog file produced by Trans-
former Explorer.
2. Calculate results based on this information, e.g., uncertainties for each fit.
3. Write the results for each measurement to different tabular files which can
later be plotted and analyzed.
The presented procedure needs some further explanation. When Transformer Ex-
plorer writes to the RefFitLog the following parameters, except from date and time,
are included: tap position, power direction, total low load & high load, ratio slope &
intercept, ratio deviation, impedance slope & intercept and impedance deviation.
The quantities are further given by phase, each having a real and an imaginary
1
MATLAB version R2015b was used in this thesis work.
33
part. Since the output files is classified by tap position and power direction not
all of the information in the RefFigLog files are needed. As explained earlier the
results need to be presented with its corresponding uncertainty. Thus, the program
must calculate the uncertainty for each quantity and write in separate columns in
the tabular file. For the magnetizing current this corresponds to the standard de-
viation of the ratio fit and for the quantities: the standard deviation divided by
the range (see section 3.7). The resulting tabular files produced by the program
contains the following information:
Ratio uncertainty.
Impedance uncertainty.
Magnetizing current.
and again, the information was given for all of the three phases. Some of the
information was later considered as less valuable and not used in the study, i.e.,
number of data and difference to nominal for the quantities. Difference to nominal
for the magnetizing current was not included since no information about the nominal
mag current existed for any of the studied transformers.
4.1.2 Results
In this section the resulting plots and tables for each transformer will be presented.
To facilitate the comparison, the average of the actual variations between the mea-
surements and the average uncertainty obtained from each one of them (the error
bars) are shown in tables. They are referred to as several measurements and single
measurement, since the first column relate to the average uncertainty based on all
measurement and the second column relates to the average uncertainty estimated
from each of the single measurements. This is shown for all of the investigated
parameters. If the derived uncertainty is correct, these two should at least be of
the same order. Some of the parameters was also plotted against average load cur-
rent, this to investigate if any correlation with load exists. This is justified since
current transformers are slightly load dependent. Error bars with the uncertainty
was added to the plots to visualize if the expected uncertainty corresponds to the
variation between each measurement. When available, the nominal values were also
plotted with the quantities for comparison. It is important to mention that only
one tap position was studied: the one most commonly used. This is because not
enough load variation was observed for some of the day periods to perform a fit for
other tap positions. This is an important result itself since it implies that there is
34
no guaranty to obtain results from more than one tap position when a short time
measurement is carried out. However, it is very little chance to cover the majority
of the available tap positions even during a much longer period, as this depends on
the transformer usage. Generally, extreme taps are rarely used; most transformers
use only a few taps around the nominal.
Field Installation 1
The real part of the turn ratio looks rather stable for both the day and week
measurement with no significant deviations as can be seen in Fig. 4.1. Note here
that the individual phases differ so much from each other that the error bars is
barely visible in the graphs. Table 4.2 shows that the ratio variation corresponds
well to the single estimated uncertainties. Phase 0 shoes however a slightly higher
variation for both the periods. Fig 4.2 shows the magnetizing current as a function
of average load current. The variation looks also here to be in the same order
compared to the error bars. Table 4.4 and 4.5 reveals the same information and no
significant difference is observed. The imaginary part of the impedance is shown in
Fig. 4.3. Here the predicted uncertainty is higher than the actual variation and the
impedance varies slightly less for the week interval. An intriguing observation is the
wave shape pattern observed for both the day and weeks within roughly the same
load interval. Fig. 4.4 shows the real part of the impedance (the resistance) as a
function of average load current. Also for this the uncertainty seems well predicted
and no significant difference between the time periods can be observed. Note here
that the resistance is negative for Phase 2. This is of course not physically possible
and is a result of the low accuracy described in section 3.7. The variations between
each single measurement are however significantly lower and this is a good example
of the difference between systematic and random error.
Day Week
2.65 2.65
Phase0 Phase 0
Phase1 Phase 1
Phase2 Phase 2
Nominal Nominal
2.6 2.6
Ratio real part
2.55 2.55
2.5 2.5
10 12 14 16 10 15 20
Average load current [%] Average load current [%]
Figure 4.1: Real part of the turn ratio as a function of average load current, Field
installation 1.
35
Table 4.2: Turn Ratio Uncertainty2 , day-period, Installation 1
Day Week
Phase 0 Phase 0
0.6 Phase 1
0.6 Phase 1
Phase 2 Phase 2
0.5 0.5
Magnetizing current [A]
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
10 12 14 16 10 15 20 25
Average load current [%] Average load current [%]
Figure 4.2: The magnetizing as a function of average load current, Field installation 1.
2
The left column ”Several measurements” is here a measure of the actual uncertainty from all
the measurements obtained by taking the standard deviation of the turn ratio values. The right
column ”Single measurements” is the mean value of all the estimated uncertainties from each
single measurement ( the error bars).
36
Table 4.4: Magnetizing Current Uncertainty, day-period, Installation 1
Day Week
38.4 38.4
Phase 0 Phase 0
38.2 Phase 1 38.2 Phase 1
Phase 2 Phase 2
38 Nominal 38 Nominal
Impedance imaginary part [Ω]
37.6 37.6
37.4 37.4
37.2 37.2
37 37
36.8 36.8
36.6 36.6
36.4 36.4
10 12 14 16 10 15 20
Average load current [%] Average load current [%]
Figure 4.3: The imaginary part of the impedance as a function of average load current,
Field installation 1.
37
Day Week
0.8 0.8
Phase 0 Phase 0
Phase 1 Phase 1
0.6 Phase 2
0.6 Phase 2
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
10 12 14 16 10 15 20 25
Average load current [%] Average load current [%]
Figure 4.4: The real part of the impedance as a function of average load current, Field
installation 1.
38
Table 4.8: Real Impedance Uncertainty, day-period, Installation 1
Field installation 2
Also for this installation no significant change in the results between the two periods
are observed. Note here that the load profile is much less varying – the load variation
is similar for both periods. The turn ratio is shown in Fig 4.5. Table 4.8 and
4.9 shows that the uncertainty corresponds well to the variation for both periods.
Also for the magnetizing current shown in Fig. 4.6, the uncertainty seems in the
right order. There is one measurement around 14% for the shorter period, where
the currents are comparably higher but this day also has higher uncertainty. The
imaginary part of the impedance is shown in Fig. 4.7. Here it seems like the
variation is slightly smaller for the week measurements as also indicated by Table
4.12 and 4.13. The resistance can be seen in Fig. 4.8. One could possibly see a
decreasing trend for the resistance with higher average load current. The resistance
obtained below 5% deviates significantly from the other day measurements. This
is probably due to high error in the current transformers. This is probably also the
reason for the lower variation for the week periods seen in table 4.17.
Day Week
13.26 Phase 0 13.26 Phase 0
Phase 1 Phase 1
Phase 2 Phase 2
Nominal Nominal
13.24 13.24
Ratio real part
13.22 13.22
13.2 13.2
13.18 13.18
13.16 13.16
5 10 15 8 10 12 14 16
Average load current [%] Average load current [%]
Figure 4.5: Turn ratio as a function of average load current, Field installation 2
39
Table 4.10: Turn Ratio Uncertainty, day-period, Installation 2
Day Week
Phase 0 Phase 0
3 3
Phase 1 Phase 1
Phase 2 Phase 2
2.8 2.8
Magnetizing current [A]
2.6 2.6
2.4 2.4
2.2 2.2
2 2
1.8 1.8
0 5 10 15 20 8 10 12 14 16 18
Average load current [%] Average load current [%]
Figure 4.6: Magnetizing current as a function of average load current, Field installation
2
40
Table 4.12: Magnetizing Current Uncertainty, day-period, Installation 2
Day Week
57 Phase 0 57 Phase 0
Phase 1 Phase 1
Phase 2 Phase 2
Nominal Nominal
Impedance imaginary part [Ω]
55 55
54 54
53 53
52 52
5 10 15 8 10 12 14 16
Average load current [%] Average load current [%]
Figure 4.7: Imaginary part of impedance as a function of average load current, Instal-
lation 2.
41
Table 4.14: Imaginary Impedance Uncertainty, day-period, Installation 2
Day Week
5 5
Phase 0 Phase 0
Phase 1 Phase 1
4 Phase 2 4 Phase 2
Impedance real part [Ω]
2 2
1 1
0 0
-1 -1
-2 -2
5 10 15 8 10 12 14 16 18
Average load current [%] Average load current [%]
Figure 4.8: Real part of impedance as a function of average load current, Field instal-
lation 2.
42
Table 4.16: Real Impedance Uncertainty, day-period, Installation 2
Field installation 3
The last investigated transformer was field installation 3. This is the smallest of the
transformers in this survey, in terms of rated power. The turn ratio as a function
of average load current is shown in Fig. 4.9. For both measurement periods, one
can see a convergence with higher loads. Further the uncertainty is decreasing for
higher currents. Fig 4.10 showing plots for the magnetizing current one can instead
observe a divergence pattern for the week measurements. A similar converging as
in Fig. 4.9 can also be seen for the imaginary part of impedance shown in Fig. 4.11.
Finally the resistance is shown in Fig. 4.12. Also here the converging for higher
current values can be seen but towards lower resistance. All the observed patterns
for this transformer is most likely load dependent errors introduced by the current
transformers. It was later discovered that the primary side current transformer for
this unit has very high current rating compared to the average currents normally
flowing, i.e., the currents through the current transformers are around 5% for the
lower loads seen in this survey. The 1% maximum error are only guaranteed when
the current transformer are operated with a current above 10% of rated.
43
Day Week
Phase0 Phase 0
7.62 Phase1
7.62 Phase 1
Phase2 Phase 2
Nominal Nominal
7.6 7.6
Ratio real part
7.56 7.56
7.54 7.54
7.52 7.52
7.5 7.5
10 15 20 25 30 10 15 20 25 30
Average load current [%] Average load current [%]
Figure 4.9: Turn ratio as a function of average load current, Field installation 3.
44
Day Week
2.5 2.5
Phase 0 Phase 0
Phase 1 Phase 1
Phase 2 Phase 2
2 2
Magnetizing current [A]
1 1
0.5 0.5
10 15 20 25 30 10 15 20 25 30
Average load current [%] Average load current [%]
Figure 4.10: Magnetizing current as a function of average load current, Field Installation
3.
45
Day Week
85 85
Phase 0 Phase 0
Phase 1 Phase 1
Phase 2 Phase 2
84.5 Nominal 84.5 Nominal
Impedance imaginary part [Ω]
83.5 83.5
83 83
82.5 82.5
10 15 20 25 30 10 15 20 25 30
Average load current [%] Average load current [%]
Figure 4.11: Imaginary part of impedance as a function of average load current, Field
installation 3.
46
Day Week
3.8 3.8
Phase 0 Phase 0
3.6 Phase 1 3.6 Phase 1
Phase 2 Phase 2
3.4 3.4
Impedance real part [Ω]
3 3
2.8 2.8
2.6 2.6
2.4 2.4
2.2 2.2
2 2
10 15 20 25 30 10 15 20 25 30
Average load current [%] Average load current [%]
Figure 4.12: Real part of impedance as a function of average load current, Field instal-
lation 3.
47
4.2 Data rate
Transformer Explorer can acquire data for analysis with several different data rates.
The minimum rate that can be used is 1 minute rate since this is needed to obtain
stable voltage and current phasors. A similar study as in section 4.1 was conducted
in order to find out if any difference between the used data rate exists. Here the
same set used for the 2-day study at field installation 1 was used, but with two
different data rates: 10 min and 30 min. A difference is especially observed for the
turn ratio where the variation seem slightly lower when the higher rate is used. No
difference is observed between the 10 min and the 1 min rate for the same periods
obtained in section 4.1.2. No significant difference can be seen for the impedance.
48
Table 4.28: Imaginary Impedance Uncertainty, 10 min rate, Installation 1
49
Serve as a basis for future measurements.
All of these points except the last could appear to be more or less obvious for the
reader. The latter one however is really essential. As shown from the results in
section 4.1.2 the accuracy in a measurement can be really low. It is in that case
very important to have the results from earlier reports for comparison. The data
used to simulate the measurement was a two day period from installation 2. The
MATLAB program described in section 4.1.1 was used to calculate the uncertainty
and power loss.
The collaboration resulted in a final version of the report with the following
contents
Executive Summary
Background.
Transformer data.
Method.
Results.
Conclusions.
In the background a brief description of what has been measured and why is
presented. In the next section available data on the transformer is presented. The
Method describes how the measurement is carried out and what equipment that
was used. In results the results are presented in a brief and concise way and finally
conclusions and when the next measurement should take place. In the following
sections each content will be described in more detail.
Executive Summary
Probably the most essential part of the report. The report is briefly summarized
and recommendations are highlighted. It should be possible for the customer to
make decisions from only reading this part.
Background
The background to the measurement is presented; what are measured and what
causes have initiated the decision to perform a measurement. Since the used method
is entirely new, the reader are also shortly introduced to the concept.
Transformer Data
Here is basically the transformer name plate data presented in a table. Since turn
ratio and impedance is not directly visible here, these are presented in a separate
table with actual values for the tap positions during the measurement. If any
additional information is available, e.g, Magnetizing current values from factory
tests, these are also presented here.
50
Method
Here the measurement system is described, the measurement period and if there is
any problems regarding the procedure that arose during the time. This section is
divided into the following subsections: System Description and Instrument Trans-
formers. The first describes the different part of the system, e.g, specifications on
DAC and current clamps. In the latter any information about the instrument trans-
formers is presented in tables. Also the absolute error that could be introduced is
specified.
Any details about Transformer Explorer is not presented in the method section.
This since a brief description is not enough to understand the concept. This will
instead be available in an appendix.
Results
The actual results from the measurement is here presented as condensed as possible,
with one table for each measured quantity. The results are presented for two tap
position, since enough data was obtained for both tap 12 and 13 in this case. There
is no guarantee however, that this is possible for other 2-day measurements as
concluded in section 4.1.2. In each table the measured value is presented for each
phase, together with: deviation from last measurement, deviation from nominal (if
available) and uncertainty. The most important parameter here is deviation from
last measurement since the accuracy is low. The last quantity presented is power
loss. This is not provided in a table, instead a plot with power loss as a function
of current will be shown, both uncorrected and corrected. This is supplemented
with a table with the correction factors used. The method for correcting loss is
still under development and no uncertainty is thus shown (see chapter 7). Finally a
table of the uncertainty and accuracy is shown. This is done as a comparison with
the IEC standard values for off-line measurement. This is the only reference values
available up to date.
Conclusions
Conclusions based on the measurement results are presented together with recom-
mendations. This could for example be that more advanced diagnostic methods
need to be adopted for further investigation or just a new suggested time period to
next routine measurement.
51
5 Analysis Improvements
When transformer explorer is used for monitoring purposes, the criteria controlling
decision to make a fit becomes much more critical. The program should no longer
fit on all present data, instead the program must be able to judge autonomously
when the time is right to make a fit. This is done using a number of criteria that
must be fulfilled. The most important from where program makes its judgment are:
number of data in a fit, minimum load range in fit, minimum load in fit.
An additional factor is to exclude the 30% most deviating data. Whenever a fit is
made based on these criteria, some data points is always removed according to the
specified removal criteria. This is an efficient procedure for removing outliers that
are not normally distributed, e.g., disturbances that would otherwise affect the fit.
A problem however is that the procedure has the tendency to also remove evenly
distributed data.
Another problem that has been discovered which is a special case of the first one,
is that sometimes when the transformer changes tap position, the class from the
previous tap is almost but not entirely fulfilling all of the criteria settings. It could
for example be that one or two data are missing to fulfill the minimum number of
data. When the transformer suddenly changes back to the old tap position, the load
could be quite different and thus the last data appear far away from the others. If
the last point causes the criteria to be fulfilled the resulting fit has been seen to be
deviating significantly in some cases.
52
Figure 5.1: The labVIEW code block controlling the distribution of data. The desired
value in percent is set by the user and is multiplied by the difference between the min
and max value of the data corresponding to the low side current. The low side current is
given by the ratioVar. The difference between each value in the vector in ascending order
is then calculated in the for loop and saved in a new vector. The maximum value is then
compared to the minimum distribution criteria, and if it is larger the result will be false
and no fit will be performed on the actual data set.
compared to a deviating point more close to the others. An outlier can in this case
have a large impact on the entire fit in a negative way. More evenly distributed
data will reduce the impact of the outlier. Fig 5.3 illustrates this situation with
the same data set but with distribution control introduced. The proposed method
was tested on a known troublesome data set from installation 3 shown in Fig. 5.4,
which shows the continuous ratio fits for tap 12. Here the transformer was shut
down for a few weeks and then reconnected at a different load situation operating
at tap 11. Somewhere around 23 December it returned to tap 12 and the fitting
criteria were fulfilled a few days later, resulting in significantly deviating fit. A
closer analysis of the data in this fit is shown in Fig. 5.5. The continuous fits
for the same period but with a 15% distribution introduced is shown in Fig 5.4
where the effects of the bad fit suppressed. Note that the fits in both cases seem
to stabilize at another level after shut down more close to each other. A possible
explanation is the higher load level when the transformer is reconnected showed in
Fig. 5.7. Thus, the effect is probably the same as in Fig. 4.3, i.e., load dependent
errors in the current transformers. Another example where the use of distribution
control also reduced the impact of outliers are shown in Fig. 5.8 and Fig. 5.9.
53
0.18
0.16
0.15
0.14
0.13
0.12
0.05 0.052 0.054 0.056 0.058 0.06 0.062 0.064 0.066 0.068
Secondary current [kA]
Figure 5.2: A general example of an outlier causing trouble. Note the single deviating
point at a significantly lower load current than the others.
0.18
0.17
Primary current [kA]
0.16
0.15
0.14
0.13
0.12
0.05 0.052 0.054 0.056 0.058 0.06 0.062 0.064 0.066 0.068
Secondary current [kA]
Figure 5.3: The data used in Fig. 5.2 but more evenly distributed, suppressing the bad
effect on the fit caused by the deviating point at lower load.
54
Figure 5.4: A situation where an unwanted fit was done after the transformer was shut
down for a few weeks. Note that the ratio seem to stabilize at and deviate less from each
other after the shut down. This is most likely due to the nonlinear error characteristic of
the instrument transformers.
Figure 5.5: The data from the deviating turn ratio fit in Fig. 5.4.
55
Figure 5.6: Same period as in Fig. 5.4 but with distribution control of 20% introduced.
Note how the deviating fits in all three phases where suppressed but not entirely eliminated
since the troublesome data will still have an impact.
Figure 5.7: The apparent power for the same period as in Fig. 5.4. Here the higher
load after the reconnection explaining the convergence is confirmed.
Figure 5.8: A analyzed period with an observed deviating turn ratio fit in phase 1
(white).
56
Figure 5.9: The same data as in Fig. 5.2, with a distribution control of 20%. Note how
the deviating fit in Fig. 5.8 is avoided.
The distribution method was also tested on another data set from installation 3
where a different kind of problem were discovered. Fig. 5.10 shows the continuous
ratio fits for the particular period. Here something significant happens to the es-
timated turn ratio for all the three phases after 2014-02-07. Fig. 5.11 shows the
same period but with 15% distribution control. The overall result seem more stable
and less fluctuating, however, the proposed method has no significant effect on the
period with deviating turn ratio. A look at the reactive power delivered to the
transformer during the same period shows that significantly higher reactive power
shown in Fig. 5.12 coincide with the deviating fits of Fig. 5.11. A dependence of the
current ratio on reactive power is however not expected from the simple transformer
model used. This observation could therefore warrant deeper investigation.
Figure 5.10: The turn ratio observed at installation during a month. Something happens
to the fits for a few days which causes the ratio to deviate a lot from the reference.
57
Figure 5.11: The same period with 15% distribution control introduced. The ratio
seems more stable but the distribution has barley no effect on the significant deviations.
Figure 5.12: The reactive power going in to the transformer. The period of higher
reactive power coincide with the period with deviating turn ratio fits.
58
5.2 An Alternative Fitting Algorithm
The method which Transformer Explorer uses to exclude outliers is based on a
technique where the least square fit is performed twice. First to make an initial
estimate, and then a second time, but this time the data points that deviate ac-
cording to a specified criteria are removed before the final fit is made. This method
efficiently removes outliers and skewed data points in most cases. One problem
however, arises when the fit must be made to very few data points where several of
them are heavily disturbed (so called outliers). This is due to the fact that the least
square method is relatively sensitive to outliers (since it minimizes the least square
sum of all data in the fit). Thus, when the first fit is done containing disturbed
data this will heavily affect the slope of the resulting fit. Then, when the outliers
are removed before the final fit some of the ”good” data will also be removed. This
is unwanted since you always want to fit to as many data as possible when there
is noise present. A more robust technique that could be used is the Theil-Sen Es-
timator [12]. This method is much less sensitive to outliers than the least square
method.
The method estimates the slope β by taking the median slope among all lines
trough pair of sample points X and Y. It is given by
( )
(Yi − Yj )
β̃n = M ed : Xi 6= Xj , 1 ≤ i < j ≥ n (5.1)
(Xi − Xj )
59
100
Data points
Least Square Fit
80
Theil-Sen
60
y
40
20
0
0 10 20 30 40 50 60 70 80 90 100
x
Figure 5.13: The two estimators tested on the corrupted data set. Note how the Theil-
Sen estimator is unaffected by the outliers while the least square methods slope estimate
is clearly affected.
Finally the most deviating data were removed from the two fits (10% in this case).
The final remaining data and the fit with the least square method is shown in Fig.
5.14. The figure shows precisely what was predicted. The final fit has some of the
good data removed and two outliers remaining which will have a negative effect on
the fit estimate.
100
Data points
Least Square Fit
80
60
y
40
20
0
0 10 20 30 40 50 60 70 80 90 100
x
Figure 5.14: Least square fit performed twice with outliers removed. Most of the outliers
are removed with this method, however, several of the good data were also removed which
is undesirable.
In Fig. 5.15 the same thing is shown but with the Theil-Sen method to obtain
the first fit. 10% outliers is removed and the final fit is done with the standard
method. Here one can see that all of the true outliers were efficiently removed but
unlike the original method more of the good data was kept. This is desirable since
the fit should be performed with as many good data points as possible.
Finally the proposed method was implemented in the sub.vi controlling the fitting,
as a part of the transformer explorer LabVIEW program. The block diagram is
60
100
Data points
Theil-Sen + Least Square Fit
80
60
y
40
20
0
0 10 20 30 40 50 60 70 80 90 100
x
Figure 5.15: The proposed method where the Theil-Sen is used on the original data
set, 10% most deviating data points are removed and finally the remaining data is fitted
with linear regression. Note how all of the outliers were removed, while much more of the
good data was kept in the fit.
shown in Fig. 5.16. The proposed method should be able to find application in two
Figure 5.16: The code block for the Theil-sen estimator. The main frame is a case
structure containing both the standard least square algorithm and the new method. The
main of the code is the double for loop which calculates all the slopes for each x and y
pair and saves in a new vector. The median is then calculated in the main case. The
lower small case structure in the figure calculates the intercept according to Eq. 5.2.
61
versions. Either as described above with outlier removal, or by simply performing
the a fit once with the new algorithm. The latter means that no concern has to be
taken on how many outliers that should be removed. The method was tested on
some of the field installation data. One improved case were a significant difference
can be seen was observed with data from installation 2 when no removal of outliers
was done. The data set was known from earlier to have significant outliers deviating
up to 70 standard deviations. The data plotted in MATLAB is shown in Fig 5.17.
A plot showing the difference to fit for the particular set is shown in Fig. 5.18.
Fig 5.19 shows the ratio fit for tap 12 with the standard method. Fig 5.20 shows
the same but with the proposed method. The fits for the latter case appears more
stable. This fact confirms that the implemented Theil-Sen algorithm is working and
is less sensitive to the outliers.
0.55
0.5
Primary current [kA]
0.45
0.4
0.35
0.026 0.028 0.03 0.032 0.034 0.036 0.038 0.04
Secondary current [kA]
Figure 5.17: The data with observed outliers plotted in MATLAB. This is the total
amount of data used for the fits in Fig. 5.18.
Figure 5.18: Difference to reference fit for tap 12 with the troublesome data set from
installation 2. Most of the data are deviating less than 5 standard deviations from the
reference fit, while several outliers deviate up to 70 standard deviations.
62
Figure 5.19: Turn ratio fits for tap 12, performed with the standard method.
Figure 5.20: Turn ratio fits for tap 12 with the proposed method. The fits appear more
stable compared to Fig. 5.19 where the standard method was used.
63
6 Power Loss
As mentioned earlier the power loss of a transformer is simply given by the difference
between the input power supplied by the primary side and the output power deliv-
ered by the secondary side according to Eq. 3.5. In reality however, the measured
voltage and current signals are always subjected to a certain amount of error. The
error consists of a stochastic part associated with noise and a systematic (static)
part associated with the accuracy of the measurement equipment, i.e., Instrument
transformers and additional sensors. If one assumes that each sensor gives a value
with some error ε, then the voltage and current signals provided by the sensors can
be written as
Ṽ = (1 + εV )V (6.1)
I˜ = (1 + εI )I (6.2)
where ε is a complex number as the sensors can introduce an error in both the
amplitude and the phase angle of the phasors. Because of the errors introduced by
the sensor the power loss become something quite different than the actual, since
the measured loss becomes
where the subscripts H and L refer to the high(primary) and low(secondary) voltage
side of the transformer respectively.
The power loss of a typical power transformer is less than one percent and the
accuracy class of the voltage and current transformers is normally in the same or a
higher order of magnitude. As a consequence Eq. 6.3 is very ill-conditioned as it
is formed as a difference between quantities that are approximately equal in size.
This makes the equation very sensitive to errors and thus the active power loss can
appear to decrease with load or even adopt negative values. The aim of this chapter
is to present a method for minimizing the impact of the systematic errors on the
measured power loss.
The technique to obtain the power loss correction is currently subject for a patent
application and can thus not presently be disclosed here. Instead, a few examples
indicating the effectiveness of the proposed correction method are presented in the
following section.
64
6.1 Application Examples on Field Installation
Data
The method was implemented in MATLAB and tested with field data from Field
installation 2. As can be seen in Fig. 6.1 and Fig. 6.2, for both the individual and
total power loss, the sensors were so poorly matched that the loss became negative.
Further, Fig 6.1 shows that the proposed method has the ability to turn the loss
to positive values. The curves also got a more characteristic shape (parabolic) and
the power loss is also in the right order of magnitude (If 1 % loss is assumed, this
gives approximately 10 kW losses per phase at 0.14 p.u).
Figure 6.1: Uncorrected and corrected power loss for each phase at field installation 2.
Figure 6.2: Uncorrected and corrected total power loss at field installation 2.
65
7 Conclusions & Discussion
66
distribution tends to get more data in to a fit. It also seems logical to have some
distribution of the data from the statistical point of view.
67
Bibliography
[2] J.Duncan Glover, Mulukutla S. Sarma, and Thomas J. Overbye. POWER SYS-
TEM ANALYSIS AND DESIGN. Cengage Learning, Stamford, 2012.
[5] Martin J. Heathcote. The J & P Transformer Book. Newnes, Oxford, 1998.
68