See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/364331971

Thermal monitoring of power transformers

Thesis · September 2017

CITATIONS READS

3 584

1 author:

Mohammad Djamali
Fichtner GmbH & Co. KG
16 PUBLICATIONS 373 CITATIONS

SEE PROFILE

All content following this page was uploaded by Mohammad Djamali on 14 October 2022.

The user has requested enhancement of the downloaded file.

Thermal Monitoring of Power Transformers

Von der Fakultät Informatik, Elektrotechnik und Informationstechnik

der Universität Stuttgart zur Erlangung der Würde eines
Doktor-Ingenieurs (Dr.-Ing.) genehmigte Abhandlung

vorgelegt von
Mohammad Djamali
aus Teheran, Iran

Hauptberichter: Prof. Dr.-Ing. Stefan Tenbohlen

Mitberichterin: Prof. Dr. Zhongdong Wang
Tag der mündlichen Prüfung: 29. Januar 2018

Institut für Energieübertragung und Hochspannungstechnik

der Universität Stuttgart
2018
For the love of my life-
Elaheh
Acknowledgement I

Acknowledgement

I sincerely appreciate the people who have helped make this dissertation possible. First and
foremost I sincerely thank Prof. Stefan Tenbohlen for giving me the opportunity to work in
his outstanding team and for his valuable discussions and supports in the preparation and
completion of this research.

I express my sincere gratitude to Dr. Ulrich Schärli for his continuous supports and Prof.
Krzysztof Rudion for his elaborate discussions. Furthermore, I would like to express my
deepest gratitude and thanks to Ms. Nicole Schärli and Ms. Hermine Lwowski and all other
staff members of Institute of Power Transmission and High Voltage Technology (IEH) for
contributing to an inspiring and pleasant atmosphere.

I owe my thanks to my colleagues at IEH and my friends Farzaneh Vahidi and Shahrouz
Mardaneh for their friendly supports; and also Stephanie Hägele, Phillipp Hillenbrand,
Dr. Ahmad Abdelmajeed, and Dr. Mohsen Nemati for the enjoyable atmosphere that they
created in the institute. I would like to thank my colleagues Michael Beltle, Martin Siegel,
and Nicolas Schmidt for supporting me in collecting the necessary know-how. I am greatly
indebted to my friends and colleagues Dr. Mohammad Hamed Samimi, Saeed Khandan
Siar, and Mohammad Moradi-Dalvand for their valuable comments and their reviews of
this thesis. I sincerely appreciate Mr. Ebrahim Mashayekhi for the editorial review of this
thesis

Finally, I would like to express my profound gratitude to my family for their encourage-
ments and supports. Last, but certainly not least, I express my gratitude and lasting love
to my spouse, Elaheh, and acknowledge her outstanding moral and emotional supports in
completing this dissertation.

Stuttgart, 17.09.2017
Abstract III

Abstract

With the integration of the renewable energy-based power plants into power systems and
due to their volatile nature, the overloading of transformers is inevitable because of eco-
nomical reasons. Therefore, an accurate method should be used to determine the maxi-
mum loading capability of transformers under a specified ambient temperature. Loading
capability of transformers depends strongly on their thermal conditions which can be in-
vestigated by means of a dynamic thermal model. During the overloading of transformers,
it should be ensured that the transformer does not experience an accelerated thermal ag-
ing and that the whole cooling capacity of the transformer is available. Therefore, every
effort toward monitoring of the cooling system and determining the loading capability is
worthwhile.

In this dissertation, firstly, the state-of-the-art in the thermal modeling of transformers,

their applications, and factors affecting their accuracy are provided. Afterwards, a ther-
mal model for calculation of the top-oil temperature using an online monitoring system
is proposed which uses different heat transfer phenomena, namely the natural and forced
convection, the thermal radiation, the heat conduction, and the natural ventilation. Using
thermal-electric analogy, these phenomena are connected to each other with consideration
of their mutual impacts. Two variations of the model are proposed for indoor transformers.
Due to the simple structure of the model, it can be easily implemented and integrated into
online monitoring systems.

The developed top-oil temperature model is implemented into a monitoring system. Among
the tasks of the proposed monitoring system are online parametrization of the embedded
top-oil temperature model, after acquiring a specified number of measurements; the calcu-
lation of the top-oil and hot-spot temperatures, after parametrization; monitoring of cooling
system; and determining the loading capability. The proposed model and developed system
are validated using the data measured during normal operation of transformers with differ-
ent cooling systems and different ratings. The results of validation show that the proposed
model represents a good performance to calculate the characteristic temperatures of trans-
formers. Due to the reliability and the accuracy of the model, it can be used for monitoring
of the cooling system and determining the loading capability.

It is a matter of fact that during the overloading of transformers, the temperatures of the
oil and windings are near their maximum values. The cooling systems of transformers
play vital roles to transfer the heat generated in the active parts of the transformer to the
surrounding ambient. In this context, the monitoring of the cooling system is a crucial
task. Two algorithms are proposed for this purpose which use a top-oil temperature model
IV Abstract

as their cores. The proposed algorithms are validated using the data measured during the
normal and faulty operations of two transformers with ODAF and OFAF cooling systems.
The results show that both algorithms are able to detect failures in the cooling system as
soon as possible even if the failures are not severe.

Provided that an accurate thermal model for calculation of the hot-spot temperature as well
as a reliable online algorithm for monitoring of the cooling system are available, the max-
imum loading capability of transformers can be determined under specified environmental
conditions. In this dissertation, a method for determination of the loading capability of
transformers is provided which is a reverse procedure of the calculation of the hot-spot
temperature. This procedure is an optimization problem which should be solved using the
online monitoring system. This task takes, in some cases, one minute which should be
performed parallel with other tasks. The time for completion of this task varies with the
complexity of the top-oil temperature model and the microprocessor of the monitoring sys-
tem. However, the computational capability of monitoring systems is limited; hence, an
alternative way should be provided for determination of the loading capability. Afterwards,
the dependency of the loading capability on the ambient temperature is investigated. This
dependency is an equation in form of a quadratic or linear equation for accurate determi-
nation of the loading capability and contains every information needed for this purpose.
It is shown in this thesis that these equations can be used in an online monitoring system
instead of forcing the monitoring system to solve an optimization problem under dynamic
conditions. Both procedures have been applied on different types of transformers with
different cooling systems in order to show the performance of the quadratic and linear
equations.
Kurzfassung V

Kurzfassung

Wegen der Integration erneuerbarer Energien in die bestehenden Energienetze und deren
volatiles Erzeugungsprofil gewinnt eine möglichst hohe Ausnutzung eines Transformators
aus Kostengründen an Bedeutung. Eine akkurate Methode ist notwendig, um die max-
imale Belastbarkeit eines Transformators angesichts der aktuellen Umgebungsbedingun-
gen zu ermitteln. Die Belastbarkeit von Transformatoren hängt stark von deren thermis-
chem Zustand ab, welcher mittels eines dynamischen thermischen Modells beurteilt wer-
den kann. Während der Überlastung von Transformatoren muss gewährleistet werden,
dass sowohl der Transformator keine vorzeitige thermische Alterung erfährt als auch die
gesamte Kühlkapazität des Transformators zur Verfügung steht. Deswegen sind Beiträge
zur Überwachung der Kühleinrichtung und Ermittlung der maximalen Belastbarkeit von
Transformatoren von großer Bedeutung.

In dieser Arbeit wird zum Ersten der Stand der Technik vorgestellt. Dies enthält die thermis-
che Modellierung, ihre Anwendungen und Einflussfaktoren. Anschließend wird ein ther-
misches Modell zur Berechnung der oberen Öltemperatur mittels eines Monitoringssystems
vorgeschlagen. Das Modell zieht unterschiedliche Wärmeübertragungsphänomene im Be-
tracht, nämlich die natürliche und erzwungene Konvektion, die thermische Strahlung, die
Wärmeleitung, und die natürliche Belüftung. Die bereits vorgestellten Phänomene werden
mittels einer thermisch-elektrischen Analogie miteinander verbunden. Gleichzeitig wer-
den ihre gegenseitige Einflüsse berücksichtigt. Es werden zusätzlich zwei Variationen des
vorgestellten Modells für Transformatoren in Gebäuden vorgeschlagen. Aufgrund der nicht
aufwändigen Struktur der vorgeschlagenen Modelle können sie einfach in Monitoringssys-
temen integriert werden.

Im Rahmen dieser Arbeit wird das Modell zur Berechnung der oberen Öltemperatur
hardware- und softwaremäßig implementiert und ein Monitoringssystem entwickelt. Das
Monitoringssystem dient dazu, die obere Öltemperatur und die Heißpunkttemperatur
möglichst genau zu berechnen und auf deren Basis den Zustand der Kühlanlagen zu
überwachen und die maximale Belastbarkeit zu ermitteln. Anhand während des normalen
Betriebs von Transformatoren aufgenommenen Messdaten werden das entwickelte thermis-
che Modell und das Monitoringssystem validiert. Die Ergebnisse der Validierung zeigen,
dass das entwickelte Modell eine hohe Genauigkeit bei der Berechnung der Temperaturen
besitzt. Aufgrund der Zuverlässigkeit und Genauigkeit des Modells kann es zum Zwecke
des Monitorings verwendet werden.

Während der Überlastung eines Transformators kommen die Öl- und Wicklungstempera-
turen an ihre Grenzen. Um die im Aktivteil des Transformators erzeugte Wärme an die
VI Kurzfassung

Umgebung abzuführen, ist die Kühleinrichtung des Transformators von großer Bedeutung.
In diesem Zusammenhang bietet die Überwachung der Kühleinrichtung die Möglichkeit,
aufgetretene Fehler frühzeitig zu erkennen. In dieser Arbeit werden zwei Algorithmen zur
Überwachung der Kühleinrichtung vorgeschlagen. Die beiden Algorithmen verwenden das
thermische Modell als ihren Kern. Mittels der Messdaten werden die vorgeschlagenen Al-
gorithmen validiert. Die Ergebnisse zeigen, dass die beiden Algorithmen eine frühzeitige
Erkennung der Fehler bei der Kühleinrichtung ermöglichen.

Angenommen, dass sowohl ein genaues thermisches Modell zur Berechnung der Heißpunk-
ttemperatur als auch ein Algorithmus zur Überwachung der Kühleinrichtung zur Verfü-
gung stehen, lässt sich die maximale Belastbarkeit eines Transformators angesichts der
aktuellen Umgebungsbedingungen ermitteln. So stellt diese Arbeit eine Methode zur Er-
mittlung der maximalen Belastbarkeit vor, welche der umgekehrte Prozess der Berechnung
der Heißpunkttemperatur ist. Dieses Verfahren versteht sich als ein Optimierungsproblem,
dass in das Monitoringssystem implementiert werden muss. Es werden auch im Rahmen
dieser Arbeit der Zusammenhang zwischen der maximalen Belastbarkeit und der Umge-
bungstemperatur in Form einer linearen oder quadratischen Gleichung untersucht. Es wird
gezeigt, dass diese Gleichungen das zeitaufwändige Optimierungsverfahren gut treten kön-
nen.
Contents VII

Contents

Acknowledgement I

Abstract III

Kurzfassung V

List of Figures XI

List of Tables XV

Nomenclature XVII

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Outlines of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Literature Survey 7
2.1 Dynamic Thermal Modeling of Power Transformers . . . . . . . . . . . . . . 8
2.1.1 Loading Guides . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1.1 Final Temperature Rise . . . . . . . . . . . . . . . . . . . . . 10
2.1.1.2 Oil Time Constant . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1.3 Oil Time Constant Correction Factor . . . . . . . . . . . . . . 11
2.1.2 Other Conventional Models . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.2.1 Linear Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.2.2 Swift Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.2.3 Susa Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.3 Advanced Thermal Models . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Model Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Factors Affecting the Temperature Rise . . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Viscosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 Harmonic Contamination of Load Currents . . . . . . . . . . . . . . . 17
2.3.3 Environmental Factors . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.3.1 Solar Irradiation . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.3.2 Wind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.4 Tap Changer Position . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.5 Geomagnetically Induced Current (GIC) . . . . . . . . . . . . . . . . . 19
2.3.6 Moisture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Applications of the Thermal Models . . . . . . . . . . . . . . . . . . . . . . 20
2.4.1 Prediction of the Hot-Spot Temperature . . . . . . . . . . . . . . . . . 20
VIII Contents

2.4.2 Estimation of Loss of Life . . . . . . . . . . . . . . . . . . . . . . . . 24

2.4.3 Determination of the Overload Capability . . . . . . . . . . . . . . . . 25
2.4.4 Online Monitoring of Cooling System . . . . . . . . . . . . . . . . . . 26
2.4.5 Other Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3 Improved Dynamic Thermal Model 29

3.1 Power Transformer Thermal Modeling . . . . . . . . . . . . . . . . . . . . . 30
3.1.1 Thermal Modeling Concept . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.2 Determination of the Thermal Resistance . . . . . . . . . . . . . . . . 30
3.1.2.1 Oil Convection . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.2.2 Oil Mixture in OFAF Transformers . . . . . . . . . . . . . . . 35
3.1.2.3 Equivalent Oil-Side Thermal Resistance . . . . . . . . . . . . 35
3.1.2.4 Air Convection . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.2.5 Heat Radiation . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.2.6 Equivalent Thermal Resistance of the Air-Side . . . . . . . . . 38
3.1.2.7 Equivalent Thermal Resistance . . . . . . . . . . . . . . . . . 38
3.2 Distribution Transformer Thermal Modeling . . . . . . . . . . . . . . . . . . 39
3.2.1 Thermal Modeling Concept based on Tank Temperature . . . . . . . . 39
3.2.2 Determination of the Thermal Resistance . . . . . . . . . . . . . . . . 40
3.2.2.1 Natural Convection in oil . . . . . . . . . . . . . . . . . . . . 42
3.2.2.2 Heat Conduction Through Tank Wall . . . . . . . . . . . . . . 42
3.2.2.3 Natural Convection in Air . . . . . . . . . . . . . . . . . . . . 42
3.2.2.4 Heat Radiation in Air . . . . . . . . . . . . . . . . . . . . . . 43
3.2.2.5 Natural Ventilation in Transformer Station . . . . . . . . . . . 43
3.2.2.6 Equivalent Thermal Resistance . . . . . . . . . . . . . . . . . 44
3.2.3 Calculation of the Top-Oil Temperature using Tank Temperature . . . . 44
3.2.4 Summary of the Proposed Thermal Model for Distribution Transformers 45

4 Hardware Implementation 47
4.1 Developed Monitoring System for ONAN Transformers . . . . . . . . . . . . 48
4.2 Field Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5 Validation of the Dynamic Thermal Model 55

5.1 Parametrization of the Thermal Model . . . . . . . . . . . . . . . . . . . . . 56
5.2 ODAF Transformers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2.1 Performance of the Model in High Loading Regimes . . . . . . . . . . 60
5.2.2 Performance of the Model at Low Ambient Temperatures . . . . . . . . 62
5.3 OFAF Transformers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.4 OFAN Transformers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.5 ONAF Transformers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.6 ONAN Transformers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.7 ONAN Indoor Distribution Transformers . . . . . . . . . . . . . . . . . . . . 71
5.8 Summary of the Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Contents IX

6 Monitoring of the Cooling System 77

6.1 Monitoring by the Error in the Calculation of the Top-Oil Temperature . . . . 78
6.1.1 An Algorithm for Analyzing the Error for Detection of Failures in the
Cooling System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.1.2 Validation of the Proposed Algorithm . . . . . . . . . . . . . . . . . . 82
6.1.2.1 Healthy Operation of the Cooling System . . . . . . . . . . . . 84
6.1.2.2 Detection of Fan and Pump Failures . . . . . . . . . . . . . . . 85
6.1.2.3 Two Case Studies . . . . . . . . . . . . . . . . . . . . . . . . 87
6.1.2.4 Incapability of the Loading Guide . . . . . . . . . . . . . . . . 89
6.2 Changes of the Estimated Parameters for Detection of Fan Failures . . . . . . 89
6.2.1 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.2.2 The Proposed Algorithm for Detection of Changes in the Estimated
Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.2.3 Validation of the Proposed Algorithm through Two Case Studies . . . . 94

7 Assessment of the Loading Capability 103

7.1 Loading Beyond Nameplate Rating . . . . . . . . . . . . . . . . . . . . . . . 104
7.2 Dependency of the Loading Capability on the Ambient Temperature . . . . . 107
7.2.1 Investigation of the Concavity of the Dependency . . . . . . . . . . . . 108
7.2.2 Loading Capability of Wind Farm Transformers . . . . . . . . . . . . . 112
7.2.3 Loading Capability of Indoor Distribution Transformers . . . . . . . . 113
7.3 Summary of the Loading Capability . . . . . . . . . . . . . . . . . . . . . . 116

8 Conclusions and Future Works 119

8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.2 Recommendations for Future Works . . . . . . . . . . . . . . . . . . . . . . 122

Bibliography 125

Appendix 139
A.1 Nonlinear Least-Squares Method for Parameter Estimation . . . . . . . . . . 139
A.2 Thermal Properties of Oil and Air . . . . . . . . . . . . . . . . . . . . . . . 141
A.3 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
List of Figures XI

List of Figures

2.1 Thermal-electric analogy of conventional models. . . . . . . . . . . . . . . 10

2.2 Thermal network of the 3-body model. . . . . . . . . . . . . . . . . . . . . 14
2.3 Thermal network of the 2-body model. . . . . . . . . . . . . . . . . . . . . 14
2.4 Change of the loss ratio (R) with the tap changer position (tap) . . . . . . . 19
2.5 Oil and winding temperature changes from the bottom to the top of the
winding according to the IEEE C57.91 . . . . . . . . . . . . . . . . . . . 22
2.6 Measurement Error of the hot-spot temperature due to incorrect or
inaccurate measurement of top-oil temperature rise and winding
temperature rise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.1 Schematic diagram of the oil flow path, and the oil temperature change
inside power transformers with different cooling systems . . . . . . . . . . 31
3.2 Thermal-electric analogy of the proposed top-oil temperature model. . . . . 32
3.3 Schematic of an indoor distribution transformer. . . . . . . . . . . . . . . . 41
3.4 Thermal-electric analogy of the proposed model for indoor distribution
transformers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.1 Concept of the developed system. . . . . . . . . . . . . . . . . . . . . . . 48

4.2 Photos of the prototype of the developed printed circuit board for logging
of the current and temperature in which the proposed thermal model and
the proposed algorithm for determination of the loading capability are
embedded. (a) Front side. (b) Back side. . . . . . . . . . . . . . . . . . . . 49
4.3 Prototype of the developed monitoring system for ONAN transformers. . . 49
4.4 Flowchart of the operation algorithm of the developed system. . . . . . . . 51
4.5 Measured values for parametrization of the thermal model during normal
operation of the indoor distribution transformer by the developed online
monitoring system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.6 Measured ambient and station temperatures, measured load, measured and
calculated tank temperatures using proposed model. . . . . . . . . . . . . . 53
4.7 Ambient and station temperatures, and loading capability of the
transformer calculated using thermal model and linear equation. . . . . . . 54

5.1 Schematic of the parametrization. . . . . . . . . . . . . . . . . . . . . . . . 57

5.2 Effect of parametrization period on the accuracy of the thermal model. . . . 58
5.3 The measured and calculated top-oil temperature during in-service
operation of Tr1 where the load factor is near the nameplate load and the
ambient temperature is high. The model has been parametrized by data
measured in June. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.4 Duration curve of the error in the calculation of the top-oil temperature
using different models for Tr1 during August when the load factor is near
the nameplate rating and the ambient temperature is high. . . . . . . . . . . 61
XII List of Figures

5.5 The measured and calculated top-oil temperatures for Tr1 when the
ambient temperature is very low. The model has been parametrized by
data measured in June. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.6 Duration curve of the error in the calculation of the top-oil temperature
using different models for Tr1 during December when the ambient
temperature is very low. The model has been parametrized by data
measured in June. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.7 The measured and calculated top-oil temperatures for Tr2. The model has
been parametrized by data measured in July. . . . . . . . . . . . . . . . . 64
5.8 Duration curve of the error in the calculation of the top-oil temperature
using different models for Tr2 during August. The model has been
parametrized by data measured in July. . . . . . . . . . . . . . . . . . . . . 65
5.9 The measured and calculated top-oil temperatures for Tr3. The model has
been parametrized by data measured in April. . . . . . . . . . . . . . . . . 67
5.10 Duration curve of the error in the calculation of the top-oil temperature
using different models for Tr3 during July. The model has been
parametrized by data measured in April. . . . . . . . . . . . . . . . . . . . 67
5.11 The measured and calculated top-oil temperatures for Tr4. The model has
been parametrized by data measured in December. . . . . . . . . . . . . . . 69
5.12 Duration curve of the error in the calculation of the top-oil temperature
using different models for Tr4 from January to May. The model has been
parametrized by data measured in December. . . . . . . . . . . . . . . . . 69
5.13 The measured and calculated top-oil temperatures for Tr5. The model has
been parametrized by data measured in December. . . . . . . . . . . . . . . 71
5.14 Duration curve of the error in the calculation of the top-oil temperature
using different models for Tr5 from end of January to end of March. The
model has been parametrized by data measured in December. . . . . . . . . 71
5.15 The measured and calculated top-oil temperatures for Tr6. The top-oil
temperature has been calculated using the calculation of the tank
temperature. The model has been parametrized by data measured in March. 73
5.16 Duration curve of the error in the calculation of the top-oil temperature
using different models for Tr6 in June, July, and August. The model has
been parametrized by data measured in March. . . . . . . . . . . . . . . . . 74

6.1 The measured and calculated top-oil temperatures for Tr1 before and after
occurrence of the failure. . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2 The measured and calculated top-oil temperatures for Tr2 before and after
occurrence of the failure. . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.3 Allowable band for recursive error (a) defined by (6.1), (b) defined by (6.4). 81
6.4 Flowchart of the proposed method for malfunction detection of the cooling
system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.5 The measured and calculated top-oil temperatures from January to May
for Tr4 when the operation of the cooling system is healthy. The model has
been parametrized in December. . . . . . . . . . . . . . . . . . . . . . . . 84
6.6 Standardized error Zi for Tr4 from January to May when the cooling
system is operated properly. . . . . . . . . . . . . . . . . . . . . . . . . . . 85
List of Figures XIII

6.7 The measured and calculated top-oil temperatures as well as the

standardized error Zi for Tr1 when one pump fails to operate. . . . . . . . . 86
6.8 The measured and calculated top-oil temperatures as well as the
standardized error Zi for Tr4 when one fan fails to operate. . . . . . . . . . 86
6.9 The measured and calculated top-oil temperatures as well as the
standardized error Zi for Tr4 when the effective cooling system is
degraded gradually. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.10 Standardized error Zi and alarm signal triggered using the proposed
algorithm with the proposed thermal model for Tr1 when two fans fail to
operate due to energization failure. . . . . . . . . . . . . . . . . . . . . . . 88
6.11 Standardized error Zi and alarm signal triggered using the proposed
algorithm with the proposed thermal model for Tr2 when three fans fail to
operate due to energization failure. . . . . . . . . . . . . . . . . . . . . . . 89
6.12 Standardized error Zi and alarm signal triggered using the proposed
algorithm and the IEC 60076-7 model for Tr1. . . . . . . . . . . . . . . . . 90
6.13 Changes in an estimated parameter of the air convection for Tr1. . . . . . . 91
6.14 Flowchart of the proposed algorithm for detection of changes in unknown
parameters of thermal models. . . . . . . . . . . . . . . . . . . . . . . . . 93
6.15 Qualitative representation of changes in the estimated parameters of the
proposed thermal model, when a failure occurs in the cooling system of Tr1. 95
6.16 The calculated LLR of the estimated parameters of the air-side for Tr1. . . . 97
6.17 The calculated LLR of the estimated parameters of the air-side for Tr2. . . . 97
6.18 Validation of the algorithm in the case of the healthy operation of fans for Tr1 99
6.19 Measured values during normal operation of the Tr9. . . . . . . . . . . . . 100
6.20 Qualitative representation of changes in the estimated parameters of the
proposed thermal model for Tr9 under severely volatile loading conditions. 101

7.1 Loading capability as a function of the ambient temperature for Tr1

determined using the data set measured during in-service operation in June. 109
7.2 Loading capability of Tr1 in December determined using the proposed
model, the linear and quadratic equations. . . . . . . . . . . . . . . . . . . 110
7.3 Loading capability in April, July, and January as a function of ambient
temperature for Tr3 determined using the data set measured during
in-service operation in April. . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.4 Loading capability of Tr3 in July determined using the proposed model,
the linear and quadratic equations. . . . . . . . . . . . . . . . . . . . . . . 111
7.5 Loading capability of Tr5 from December to March as a function of the
ambient temperature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.6 Loading capability of Tr5 in June calculated using the proposed model, the
linear and quadratic equations. . . . . . . . . . . . . . . . . . . . . . . . . 114
7.7 Loading capability of Tr6 determined using the proposed model in March
which has been used to estimated the parameters of the linear dependency. . 115
7.8 Ambient and station temperatures, and the loading capability of Tr6
calculated using the thermal model and linear equation (7.6) in June, July,
and August. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
XIV List of Figures

7.9 Loading capability as a function of ambient temperature for Tr1, Tr2, Tr3,
Tr4, and Tr5 determined using the training data sets. . . . . . . . . . . . . . 117
List of Tables XV

List of Tables

3.1 Summary of the parameters, inputs, constants, and output of the proposed
model for outdoor power transformers. . . . . . . . . . . . . . . . . . . . . 39
3.2 Categorization of the parameters, inputs, constants, and outputs of the
proposed model for indoor distribution transformers. . . . . . . . . . . . . 45

4.1 Technical specifications of the proposed monitoring system for ONAN

transformers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2 Characteristics of the indoor ONAN transformer under test. . . . . . . . . 52

5.1 Characteristics of the transformers under study. . . . . . . . . . . . . . . . 56

5.2 The months during which the data sets have been measured. The data sets
are measured during normal operation of transformers and are used for
parametrization and validation of the models for the transformers under
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.3 Unknown parameters of the proposed thermal models estimated using
least-squares method for the transformers under study. . . . . . . . . . . . . 59
5.4 Indices of error in the calculation of the temperature during August for Tr1
using the IEC 60076-7, the linear model, and the proposed model. . . . . . 62
5.5 Indices of error in the calculation of the temperature during December for
Tr1 using IEC 60076-7, linear model and proposed model. . . . . . . . . . 63
5.6 Indices of error in the calculation of the temperature during August for Tr2
using the IEC 60076-7, the linear model and the proposed model. . . . . . . 65
5.7 Indices of error in the calculation of the temperature during July for Tr3
using the IEC 60076-7, the linear model and the proposed model. . . . . . . 68
5.8 Indices of error in the calculation of the temperature from January to May
for Tr4 using the IEC 60076-7, the linear model and the proposed model. . . 70
5.9 Indices of error in the calculation of the temperature from January to
March for Tr5 using the IEC 60076-7, the linear model and the proposed
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.10 Indices of error in the calculation of the temperature from June to August
for Tr6 using the IEC 60076-7, the linear model and the proposed model. . . 74
5.11 Indices of error in the calculation of the temperature using the IEC
60076-7, the linear model and the proposed model. . . . . . . . . . . . . . 75

6.1 Summary of the application of the proposed algorithm on both

transformers under study. . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

A.1 Temperature-dependent thermal properties of oil . . . . . . . . . . . . . . 141

A.2 Temperature-dependent thermal properties of air . . . . . . . . . . . . . . . 141
Nomenclature XVII

Nomenclature

Roman

· (i) Time step -

· (T ) Temperature-dependent value -

Aair Surface of heat transfer on the air-side m2

Acond Surface of the thermal conduction m2

Aoil Surface of heat transfer on the oil-side m2

Aopening Surface area of the openings m2

Arad Effective surface area of radiation m2

Cd Discharge coefficient -

Cair Empirical factor -

Coil Empirical factor -

Cth Thermal capacity W·s/K

cpair Specific heat of the air W·s/(kg·K)

cpoil Specific heat of the oil W·s/(kg·K)

D Difference in the associated variable -

dair Characteristic diameter of the air-side m

doil Characteristic diameter of the oil-side m

Dt Time difference between two sequential measurements s

E (·) Expected value of a parameter -

F Jacobian matrix -

f Jacobian vector -

g Gravitational acceleration m/s2

XVIII Nomenclature

gr Winding-oil temperature gradient at rated load K

Grair Grashof number for the air-side -

Groil Grashof number for the oil-side -

H Height difference of two openings m

hair,combiend Heat transfer coefficient for combined convection of the air W/(m2 ·K)

hair,f orced Heat transfer coefficient for forced convection of the air W/(m2 ·K)

hair,nat Heat transfer coefficient for natural convection of the air W/(m2 ·K)

hoil,combiend Heat transfer coefficient for combined convection of the oil W/(m2 ·K)

hoil,f orced Heat transfer coefficient for forced convection of the oil W/(m2 ·K)

hoil,nat Heat transfer coefficient for natural convection of the oil W/(m2 ·K)

HF Hot-spot factor -

K Load factor or the ratio of the specified load to the full load -

K11 Oil time constant correction factors -

K21 Constant depending on the cooling system -

K22 Constant depending on the cooling system -

L Lorenz number W·Ω/K2

l Thickness of the tank wall m

LC Loading capability -

LLR (·) Logarithm of likelihood ratio of a parameter -

m Empirical factor -

m′ Empirical factor -

mact Mass of active part T

moil Mass of Oil T

mtube Mass of tube T

n Empirical factor -

n′ Empirical factor -
Nomenclature XIX

N (·) Normal distribution -

NF ans Number of running fans -

NP umps Number of running pumps -

P The rate of heat generated by the losses W

p (·) Probability of the occurrence of a parameter -

P0 No-load loss W

PK Short-circuit loss W

pair Empirical factor -

Pf l Rated loss W

Pinlet Air pressure at the inlet opening Pa

poil Empirical factor -

P rair Prandtl number for the air-side -

P roil Prandtl number for the oil-side -

qair Empirical factor -

qoil Empirical factor -

R The ratio of the short circuit loss to the no-load loss -

r Residual in the estimation of parameters -

Rair,convection Thermal resistance corresponding to the natural convection of the air

K/W

RBC Thermal resistance between the top of the winding and the cooler inlet
K/W

Rcond Thermal resistance of the heat conduction through tank wall K/W

Roil,convection Thermal resistance corresponding to the natural convection of the oil

K/W

Rradiation Thermal resistance of the thermal radiation K/W

Rth,air Thermal resistance of the air-side K/W

Rth,oil Thermal resistance of the oil-side K/W

XX Nomenclature

Rth,R Rated thermal resistance K/W

Rth Thermal resistance K/W

Rvent Thermal resistance of the air ventilation K/W

Reair Reynolds number for the air-side -

Reoil Reynolds number for the oil-side -

sair Characteristic length of heat transfer on the air-side m

soil Characteristic length of heat transfer on the oil-side m

t Time s

Ttank Measured tank temperature °C

Tto Measured top-oil temperature °C

uair Proportionality constant -

uoil Proportionality constant -

Vm,air Average air velocity m/s

Vm,oil Average oil velocity m/s

w Recursive error K

X Input matrix of models -

x′ Oil mixture weight -

Y Output matrix of models -

y Winding exponent -

Z Standardized recursive error -

Greek

βair Volumetric thermal expansion coefficient of the air 1/K

βoil Volumetric thermal expansion coefficient of the oil 1/K

∆P Pressure difference between openings of different height Pa

∆θ The temperature rise K

∆θi The initial temperature rise K

Nomenclature XXI

∆θu The ultimate temperature rise caused by the loss K

∆θw Winding-oil temperature gradient at considered load K

∆θBC Temperature drop between the top of the winding and the cooler inlet K

∆θcond Temperature drop across the tank wall K

∆θf l temperature rise at the full-load condition K

∆θom the average oil temperature rise at the considered load K

∆θtank Tank temperature rise over ambient temperature at the considered load
K

∆θto Top- oil temperature rise at the considered load K

∆θw1 Increase of the winding gradient to a certain level at the considered load
K

∆θw2 Decrease of the winding gradient to a certain level at the considered load
K

ǫ Independent additive error K

γ+ Cumulative sum of the w in the case of exceeding the allowable band -

γ− Cumulative sum of the w in the case of falling below the allowable band
-

κair Thermal conductivity of the air W/(m·K)

κoil Thermal conductivity of the oil W/(m·K)

κsteel Thermal conductivity of the steel plates W/(m·K)

µ Mean value -

µpu Per unit value of dynamic viscosity of the oil -

νair Kinematic viscosity of the air m2 /s

νoil Kinematic viscosity of the oil m2 /s

φ Vector of unknown parameters -

ψstation buoyancy-driven air flow m3 /s

ρair air density kg/m3

XXII Nomenclature

ρoil Oil density kg/m3

Σ Electrical conductivity 1/(Ω · m)

σ Standard deviation -

τw Winding time constant s

τf l The oil time constant at the rated load s

τtank Tank time constant s

τto The oil time constant s

θa Ambient temperature °C

θB Temperature of the oil at the top height of the winding °C

θB Temperature of the oil at the top height of the winding in the parallel
path °C

θs Station temperature or outlet opening temperature °C

θw Temperature of the winding at the top height °C

θhs Hot-spot temperature °C

θo−a Logarithmic average temperature of the oil and the air (K) in the radia-
tors K

θtank Tank temperature at the considered condition °C

θto Top-oil temperature at the considered condition °C

εtank Emissivity of the tank -

εwall Emissivity of the station wall -

ς Stefan-Boltzmann constant W/(m2 ·K4 )

Indices

a Number of iteration

air Air-side

oil Oil-side

steel Steel-side

ˆ Estimated value of a parameter

Nomenclature XXIII

Acronyms

AF Air Forced

AN Air Natural

I2C Inter-Integrated Circuit

IEC International Electrotechnical Commission

IEEE Institute of Electrical and Electronics Engineers

LLR Logarithm of the Likelihood Ratio

LOL Loss of Life

NTP Network Time Protocol

OD Oil Directed

ODAF Oil Directed Air Forced

ODAN Oil Directed Air Natural

OF Oil Forced

OFAF Oil Forced Air Forced

OFAN Oil Forced Air Natural

ON Oil Natural

ONAF Oil Natural Air Forced

ONAN Oil Natural Air Natural

RTC Real Time Clock

SPI Serial Port Interface

 1

1 Introduction

"Nature should direct itself according to our subjective ground of apperception"

The Critique of Pure Reason

E. Kant

Power transformers are of the most important apparatuses and large capital items of power
systems. Their operations affect both economic and technical issues of power systems
and their sudden failures increase both economic and technical costs. Hence, the predic-
tion of the failure time of transformers before the occurrence of the failure helps owners
of power transformers to adjust plans for replacement of transformers. In this context,
condition-based monitoring and in-service maintenance have economical and technical jus-
tifications [1,2]. However, the prediction of the failure time requires additional investments
and planners should decide about tolerating it based on cost-benefit analysis since the cost
of a complete condition monitoring system is significant. Hence, because of the demand of
transformers owners to guarantee the continuous and reliable operation, issues of the online
condition monitoring of power transformers have received considerable attention especially
after the development of sensors and computer technologies [3].

1.1 Motivation

The reliability of electrical power systems depends on the performance and the availability
of their components such as power transformers. According to the CIGRÉ Brochure 642
[4], before 2010, only 11% of the recorded disturbances in high voltage and highest voltage
grids are related to transformers.

However, with the integration of renewable energy-based power plants into power systems
and due to their volatile nature as well as possible consequences of this integration, the issue
of diagnostic measurements and monitoring of transformers may become more crucial since
transformers are connected directly to these new components and may be affected seriously
by the new operating conditions.
2 1 Introduction

Moreover, with the increasing average age of the transformer population, there is an in-
creasing need to be aware of their internal conditions. For this purpose, online methods and
systems have been developed in recent years. Online monitoring can be used continuously
during operation of transformers. This offers in that way a possibility to record different rel-
evant stresses which can affect the lifetime. The automatic evaluation of these data allows
the early detection of an oncoming fault [4].

Over the period of 2000 to 2010, the failure analysis of 165 generator step-up transformers
has shown that 32.1% of failures were of thermal modes [4]. Similarly, 52 (6.5%) of 799
investigated substation transformers failed due to thermal issues [4]. In other words, 10.9%
of the failures of all transformers are thermal failures [4]. In comparison with another in-
ternational transformer reliability survey back in 1983 [5], the contribution of the thermal
failure modes has been increased for generator step-up transformers and decreased for sub-
station transformers. These statistics indicate that the thermal aspect of transformers is of
high prior aspects to be considered.

Thermal monitoring of transformers provides the ability for continuous assessment of the
internal thermal behavior of transformers under different operating conditions. The thermal
behavior of transformers is reflected in characteristic temperatures of transformers, namely
the top-oil and the hot-spot temperatures [6]. With comparison to the measurement of the
top-oil temperature, the measurement of the hot-spot temperature is not cost-effective. The
installation of temperature sensors is time-consuming and can not retrofit the transformers
already installed. In addition, with the measurement of the hot-spot temperature, the inter-
dependencies between the affecting factors on thermal behavior of a transformer is not re-
vealed. Therefore, it is the best practice to calculate the hot-spot temperature. However, the
calculation of the hot-spot temperature during in-service operation of transformers requires
an accurate and simple thermal model for calculation of the top-oil temperature. For effi-
cient thermal monitoring of power transformers, a thermal model should be integrated into
the monitoring system for calculation of the top-oil temperature fulfilling several criteria as
follows [7]:

• The model should use just a few design-dependent variables. In other words, the
model should be almost independent of transformers.

• The model should be sufficiently accurate.

• The sensors which deliver the input values to the model must be able to retrofit to
transformers already installed in substations. This prevents, for example, the use of
fiber optics for direct measurement of the hot-spot temperature.
1.2 Objectives of Thesis 3

Nevertheless, thermal models fulfilling the above mentioned criteria which are widely used
for thermal monitoring of power transformers during their normal operations are based on
modeling of complex heat transfer phenomena from top of the tank to the surrounding am-
bient with equations that may limit their accuracy in some cases (specially in case of forced
cooling systems and extremely fluctuating operating conditions) since they consider the
thermal resistance of transformers as a constant value. In other cases, they are complicated
to be used in online monitoring systems and it is necessary to use some characteristics of
transformers which are normally not provided by transformer manufacturers. Therefore,
there is a need to improve the accuracy of these models with a more accurate determination
of the thermal resistance.

Moreover, in order to make the best use of the existing assets, the calculation of the hot-spot
temperature is a promising way to determine the maximum loading of transformers with
consideration of the current operating condition which can be more economically advan-
tageous than the extension of networks by installing new transformers. The determination
of the loading capability prevents the overheating of transformers and reduction of their
lifetime; as a result, it increases their availability.

In addition, the loading capability of transformers is affected by the healthy or faulty op-
eration of their cooling systems. Before the online determination of the loading capability,
the possible malfunction of the cooling system should be detected using an online algo-
rithm.

1.2 Objectives of Thesis

The main objectives and contributions of this thesis can be mentioned as follows:

• To provide a comprehensive, critical, and fundamental review of the state-of-the-art in

thermal modeling and thermal monitoring of transformers. Since the researchers have
invested more than hundred years on the subject of thermal modeling of transformers,
it is worthwhile to briefly investigate and to summarize their achievements from a
neutral point of view. Moreover, there are some vague points in the IEC [8] and
IEEE [9] loading guides and their applications, the clarification of which may help to
better understand their boundaries.

• To propose an accurate and simple thermal model for calculation of the top-oil tem-
perature fulfilling the following criteria:
4 1 Introduction

(a) The model should be used in an online thermal monitoring for calculating the
top-oil and hot-spot temperatures, for determining the loading capability, and
for monitoring of the cooling system.

(b) The proposed thermal model should be an empirical-based and adoptive model
being able to retrofit transformers already installed and independent of the trans-
former cooling mode and design details.

(c) The model should be applicable on transformers which have no temperature

pocket for measurement of the temperature. In these cases, the tank temperature
can be measured and calculated; and the top-oil temperature can be determined
by the calculated tank temperature. This method can be also utilized on trans-
formers having temperature pocket and allows the installation of sensors during
in-service operation of transformers.

(d) In order to use a physical thermal model in an online monitoring system, the per-
formance and accuracy of the model should be firstly validated. The calculated
top-oil temperature using the proposed model should be in sufficient agreement
with the measured top-oil temperature.

• To propose online algorithms for monitoring of the cooling system in air-forced oil-
immersed power transformers. If an accurate thermal model is used to calculate the
top-oil temperature, the deviation between the measured and calculated temperatures
is normally negligible. In the case of failure, the calculated temperature deviates from
the measured temperature significantly. The online algorithms benefit from this de-
viation and use allowable bands for the deviation in order to detect the malfunction
of the cooling system. The proposed allowable bands are applicable for all trans-
formers with different cooling systems. Based on this band, the alarm signal can be
triggered indicating that a problem has occurred. The proposed algorithm should be
implemented and integrated into monitoring systems.

• To propose methods for online determination of the loading capability and investiga-
tion of its dependency on the ambient and station (for indoor transformers) tempera-
tures. The methods should be easily integrated and implemented into the monitoring
system.

• To construct a prototype of a model-based monitoring system for online determination

of the loading capability using the proposed methods as well as the prediction of the
hot-spot temperature using the proposed top-oil temperature model. Moreover, the
proposed monitoring system is able to parametrize the embedded top-oil temperature
model. In other word, the system is a self-parametrized one.
1.3 Outlines of Thesis 5

1.3 Outlines of Thesis

The thesis is organized as follows:

• In chapter 2, the state-of-the-art is presented where different thermal models pro-

posed in the literature, their merits and demerits, and their applications are discussed.
In addition, the factors affecting the performance of thermal models are presented.
The considered thermal models are ones proposed in the IEC 60076-7 [8] and IEEE
C57.91 [9] loading guides, the linear model [10], the Susa model [11], the Swift
model [12, 13], and advanced thermal models [14–16] which are not yet widely used
but their aims are to overcome deficiencies of the conventional thermal models. The
factors affecting the accuracy of thermal models considered in this chapter include
the viscosity of the medium, harmonic distorted currents, environmental factors, tap
changer position, geomagnetically induced currents, and moisture in insulation. Fi-
nally, the applications of physical thermal models presented in this chapter include
the prediction of the hot-spot temperature, the estimation of loss-of-life, the determi-
nation of the loading capability, and monitoring of the cooling system.

• In chapter 3, the proposed thermal model in this thesis for oil-immersed air-cooled
power and distribution transformers is presented. Moreover, a method for calculation
of the top-oil temperature based on the calculated tank temperature is proposed in this
chapter followed by the modeling of the natural ventilation of the station building for
indoor transformers.

• In chapter 4, a low-cost hardware implementation of the proposed model in chapter

3 is introduced which has been developed at the University of Stuttgart, Institute of
Power Transmission and High Voltage Engineering. Moreover, the results of the in-
stallation of the monitoring system on a 1250 kVA indoor distribution transformer are
presented.

• In chapter 5, the proposed models is further validated using data measured during
in-service operations of nine transformers with different cooling systems. Moreover,
different operating conditions of transformers are taken into consideration to evaluate
the performance of the proposed model. The accuracy of the proposed dynamic ther-
mal model is compared with the IEC 60076-7 and linear models using error duration
curve, the mean absolute error, and the maximum absolute error.

• In chapter 6, two online algorithms are proposed for detection of failures in the cool-
ing system of oil-immersed air-cooled power transformers. The proposed thermal
6 1 Introduction

models in chapter 3 constructs the core of both algorithms. The algorithms are after-
wards validated using data measured during in-service operations through two sce-
narios. The first scenario concerns a non-severe failure in fan operation of an ODAF
power transformer whereas the second scenario is with regard to a severe failure in
fan operation of an OFAF transformer. Both algorithms are able to detect the failure
sufficiently fast in both scenarios.

• In chapter 7, a method for online determination of the loading capability of trans-

formers with different cooling systems is proposed. The dependency of the loading
capability on the ambient temperature is afterwards investigated. Finally, the pro-
posed method is applied on seven transformers.

• In chapter 8, the main achievements of this thesis are summarized and recommenda-
tions for future works are provided.
 7

2 Literature Survey

"Since history is the embodiment of spirit in the form of events, that is, of direct
natural reality, the stages of development are present as direct natural princi-
ples. Because they are natural, they conform to the nature of a multiplicity, and
exists one outside the other. Hence, to each nation is to be ascribed a single
principle, comprised under its geographical and anthropological existence."

Elements of the Philosophy of Right

G.W.F. Hegel

L
OADING capability of transformers is limited by its winding temperature which should
not exceed the maximum allowed temperature [8,9]. As an example, according to the
IEC loading guide, the aging rate of the insulation is doubled for each 6 degrees of tem-
perature increase above the maximum allowable insulation temperature [8]. Nevertheless,
in this context, the measurement of the winding and the hot-spot temperature would not be
a competent measured value in an online monitoring system since the installation of sen-
sors inside the transformer is time-consuming and the transformer should be out of service
for a long period of time. However, the hot-spot temperature can be calculated using two
different ways as follows:

• Physical and semi-physical modeling approaches describing the complex heat transfer
modes in transformers with first or second order differential equations. The inputs of
such models can be measured during operation of transformers. These models can be
used in online monitoring systems.

• Computational fluid dynamic modeling approach which is very accurate but time-
consuming and requires design details of the transformer for modeling [17]. Using
these models, the hydraulic behavior of a transformer can also be investigated [18,
19]. However, such models can not be used in online monitoring systems which
are assumed to be independent of monitored transformers. The thermal behavior of
transformers under dynamic loading can not be accurately investigated and requires
high-performance computers [20].
8 2 Literature Survey

On the contrary, the measurement of the top-oil temperature is cost effective, can be done
on the transformers already installed, and provides adequate information about the thermal
condition of transformers.

CIGRÉ Brochure 659 [6] provides a comprehensive review on the thermal modeling of
power transformers, where the IEEE and IEC loading guides [9, 21, 22] have been intro-
duced as the thermal models for the dynamic loading of power transformers; however, it is
also worthwhile to provide a review on the origins of the loading guides, their advantages
and disadvantages, other widely used conventional and advanced thermal models, and their
applications.

This section provides a review on the physical and semi-physical thermal models and their
applications. Although the inputs of thermal models based on machine learning meth-
ods (e.g. Artificial Neural Network [23], Neurofuzzy networks [24, 25], and evolving
Fuzzy systems [26]) are the same as the physical and semi-physical models, they are ex-
cluded from this review since their suitability for usage in online monitoring systems have
not been yet proved. Moreover, the loading capability of transformers can not be deter-
mined using artificial intelligence methods since, in most cases, there is no available mea-
sured data during overloading of transformers for the training of machine learning meth-
ods.

2.1 Dynamic Thermal Modeling of Power Transformers

To describe the dynamic thermal behavior of power transformers, some models have been
proposed in the literature which are based on describing a complex heat transfer phenomena
with first or second order differential equations. Despite the high interest of researchers in
this type of modeling, due to their simple structures, it is always necessary to improve the
accuracy of these models in order to represent the temperature changes rather perfectly.
These thermal models can be divided into the following groups:

• The IEEE C57.91 [9], IEC 60076-7 [8], and IEC 60354 [21] loading guides which
provide simple equations for calculation of the top-oil temperature based on the real-
time measurements during normal operation of transformers. However, their accuracy
is limited since they consider the thermal resistance of transformers independent of
the loading condition and do not account for changing of the number of operative fans
and pumps directly [27].
2.1 Dynamic Thermal Modeling of Power Transformers 9

• Other conventional models based on the physical phenomena namely Susa [28] and
Swift [12] models, which do not account for the forced convection; and therefore,
their accuracy is limited in these cases.

• Advanced thermal models based on physical phenomena which have been proposed
to overcome the deficiencies of the loading guides and the conventional models [15,
27, 29, 30].

2.1.1 Loading Guides

The basic principle of the dynamic calculation of the temperature rise of transformers fol-
lows an exponential law [31]. After the generation of losses in transformers, the tempera-
ture increases rapidly since, at first, the temperature of the transformer is not sufficient to
dissipate any loss and, consequently, all the loss is stored. As the temperature increases,
the dissipation to the surrounding medium increases, until finally no further heat is stored
and all the generated heat is dissipated. This law can be expressed with a first order dif-
ferential equation corresponding to the thermal-electric analogy depicted in Fig. 2.1 as
follows [32]:
d∆θ 1
P = Cth · + · ∆θ, (2.1)
dt Rth
where P is the losses including the no-load and load losses (W), Cth and Rth are the
thermal capacity (W·s/K) and the thermal resistance (K/W) respectively, and ∆θ is the
temperature rise (K). The solution of (2.1) in heating and cooling of the oil is as fol-
lows [33]:  
− τt
∆θ = (∆θu − ∆θi ) · 1 − e to + ∆θi , (2.2)

where ∆θu is the ultimate temperature rise (K) caused by the loss P , ∆θi is the initial
temperature rise (K), t is time (s), and τto is the oil time constant (s). Furthermore, for
monitoring purposes and the cases in which the data set are discrete, an approximation
of the exponential equation (2.2) can be used [8]. In a general form, the approximated
difference equation can be written as follows [8]:

Dt
Dθto (i) = (∆θu − θto + θa ) , (2.3)
K11 · τto
Dθto (i) = θto (i) − θto (i − 1) , (2.4)

where operator D implies a difference in the associated variable that corresponds to the time
difference Dt between to sequential time steps (s), K11 is the oil time constant correction
factors, θto is the top-oil temperature at the considered condition (°C), and θa is the ambient
temperature (°C). As it can be seen from (2.3), in order to calculate the temperature of the
10 2 Literature Survey

q to
R th

P C th

qa

Figure 2.1: Thermal-electric analogy of conventional models.

oil, there are three parameters which should be determined accurately, namely ∆θu , τ , and
K11 .

2.1.1.1 Final Temperature Rise

The heat generated in the active part of transformers is dissipated to the surrounding am-
bient through conduction, convection, and radiation. Montsinger in [34] provided some
empirical equations for these heat transfer modes inside an oil-natural cooled transformer
and concluded that the final temperature rise is proportional to the xth power of the gen-
erated losses (P ) where the value of x depends on the cooling system of the transformer
[35, 36]:
∆θu = C · P x , (2.5)

where C is a constant. Equation (2.5) assumes the Poisseuille’s law of flow through the oil
ducts where no turbulence is appreciable [37]. The parameter x means that there are two
different coolants in the transformer with different thermal characteristics which respond
differently to a specific loss. In other words, if the heat is dissipated to the surrounding
ambient across the tank wall, the air outside of the tank moves faster rather than the oil
inside of the tank [12, 13]; and therefore, the temperature rise in oil and air sides will be
different [38].

For an specified load, the temperature rise may be calculated as follows [39]:
x
1 + R · K2

∆θu = ∆θf l · , (2.6)
1+R

where ∆θf l is the temperature rise at the full-load condition, K is the ratio of the
specified load to the full load, and R is the ratio of the short circuit loss to the no-
load loss. This concept has been used by the IEEE [9, 40] and IEC loading guides
[8, 21, 22].
2.1 Dynamic Thermal Modeling of Power Transformers 11

2.1.1.2 Oil Time Constant

According to the IEEE loading guides, the oil time constant at rated load is calculated as
follows [32, 36]:
Cth · ∆θf l
τf l = , (2.7)
Pf l
where Pf l is the full load loss. Equation (2.7) is valid if x in (2.6) is equal to 1. If x
is less than 1, the oil time constant for the considered load is calculated as follows [9,
40]:

   
∆θu ∆θi
∆θf l
− ∆θf l
τto = τf l ·   x1   x1 . (2.8)
∆θu ∆θi
∆θf l
− ∆θf l

However, IEC 60076-2 [22] and IEC 60076-7 [8] use other equations for determination
of the time constant for using in (2.3). The former describes the oil time constant as fol-
lows:
Cth · ∆θu
τto = . (2.9)
P

However, IEC 60076-7 [8] recommends to use the average oil time constant in (2.3) instead
of the final oil temperature rise in (2.9) as follows:

Cth · ∆θom
τto = , (2.10)
P

where ∆θom is the average oil temperature rise at the considered load.

2.1.1.3 Oil Time Constant Correction Factor

Investigations have shown that in several ONAN and ONAF transformers, the top-oil (in
tank) time constant resulted from the rising part of the temperature curve has been less than
the simultaneously deduced average oil (in the tank) time constant [41]. Therefore, IEC
60076-7 [8] suggests to multiply a factor K11 , whose value is less than 1, by the oil time
constant. This factor is considered equal to 1 in IEEE loading guides.

2.1.2 Other Conventional Models

Although the equations of the loading guides are simple, their accuracy is limited in the
calculation of the top-oil and hot-spot temperatures in some cases; therefore, it is always
12 2 Literature Survey

necessary to improve the accuracy of the thermal models while keeping them simple in
order to be used easily for practical applications. Therefore, other simple models have
been proposed whose basic principles are the same as the loading guides. Despite their
simplicity which is a big merit and the main reason for using them in the industry and
academic researches, their accuracy in the calculation of the temperature is limited because
they do not account for the heat radiation and the heat conduction; and moreover, they
consider the thermal resistance corresponding to the natural convection as a constant value;
yet it depends on the temperature of the medium.

2.1.2.1 Linear Model

The linear model is a model developed based on the IEEE loading guide clause 7 in which
the oil exponent in (2.6) is considered equal to 1 and instead of the top-oil temperature rise
the top-oil temperature and ambient temperature are used separately [10]. One of the sim-
plifications made in this model is that the oil time constant does not change with variations
in operating conditions of a transformer; however, the main advantage of the model is its
simple form. In other words, the model can be easily used in an online monitoring system
and can be easily parametrized [10]. The governing equation for calculation of the top-oil
temperature is as follows:

θto (i + 1) = k1 · θto (i) + k2 · θa (i) + k3 · K(i)2 + k4 , (2.11)

where k1 through k4 are constants which can be determined based on the measurements
during normal operation of transformers, and i is the time step.

Moreover, Tylavsky et al in [42] investigated the errors in the prediction of the top-oil
temperature using linear model and tried to improve the accuracy of the linear model. They
claimed that the bulk of the error in the calculation of the top-oil temperature is caused by
the input data set. Moreover, some solutions have been provided to reduce the impacts of
the input data on the error [43].

2.1.2.2 Swift Model

Swift et al in [12, 13] started with the principle of the natural heat convection and provided
a differential equation with consideration of the ambient temperature for calculation of
the hot-spot and top-oil temperatures. Comparing with loading guides, they considered
the parameter x as the exponent of the temperature rise in (2.5). The proposed top-oil
temperature model is as follow:
2.1 Dynamic Thermal Modeling of Power Transformers 13

1 + R · K2
 
dθto
τf l + (θto − θa )1/x = ∆θf l 1/x . , (2.12)
dt 1+R
τf l = Rth,R · Cth , (2.13)

where Rth,R is the rated thermal resistance.

It should be noted that the right term of the (2.12) is no longer the ultimate temperature rise
but a value needed for the equality (2.12) [44].

2.1.2.3 Susa Model

Susa et al in [11] provided the dynamic top-oil temperature and hot-spot temperature mod-
els based on the principle of the natural heat convection and considered the dynamic vis-
cosity of the oil in their proposed models. Moreover, they proposed a bottom winding
temperature rise over bottom-oil temperature as well as bottom oil temperature rise over
ambient temperature based on the law of energy conservation [45, 46]. The governing dif-
ferential equation for calculation of the top-oil temperature according to the Susa model is
as follows:

dθto (θto − θa )1+x 1 + R · K2

 
τf l + = ∆θf l . , (2.14)
dt (µpu · ∆θf l )x 1+R
where µpu is the per unit value of dynamic viscosity of the oil.

2.1.3 Advanced Thermal Models

Advanced thermal models are models whose parameters, as opposed to the conventional
models, varies with the temperature and load of transformers. The advanced thermal models
are principally based on the n-body modeling concept with consideration of one or more
heat transfer modes. It is worthwhile to note that the conventional models use the same
concepts; nevertheless, some simplifications result in the limitation of their performances.
The concept of n-body modeling is based on the thermal-electric analogy in which the heat
path inside transformers is represented by an electric circuit containing lumped parameters,
namely resistors representing the thermal resistance, capacitors representing the thermal
capacity, voltage sources representing the temperature, and current sources representing
the losses in transformers [47]. According to this concept each body, containing an RC
circuit and a current source, is related to an active part of a transformer and represents a
temperature difference under both dynamic and steady state loads.
14 2 Literature Survey

G6
G5
G4
1 2 3 4 5
G1 G2 G3 G7 G8
C1 C2 C3 C4 C5

Figure 2.2: Thermal network of the 3-body model. 1. High voltage winding, 2. Low volt-
age winding, 3. Core, 4. Oil, 5. Tank; G and C correspond to the thermal conductance
and the thermal capacitance respectively [14].

θCu θOil
1 Λ1 2 Λ2

C1 C2
P1 P2
ETEP

Figure 2.3: Thermal network of the 2-body model. P , Λ, and C correspond to the generated
heat, the thermal conductance, and thermal capacitance in each body respectively [15,
48].

The most common n-body model is the 3-body model [14, 15] each of which is related to a
specific heat source. Lindsay in [14] considered the high voltage and low voltage winding
losses as well as core losses as the heat sources, which is depicted in Fig. 2.2, whereas
Tang et al in [30] considered the winding losses, the core losses, and the stray losses. If
the aim of the modeling is the calculation of the hot-spot temperature, the model can be
simplified and the 2-body model can be used [15, 48] as is depicted in Fig. 2.3. If the aim
of the model is to calculate the top-oil temperature, the model can be further reduced to
the one-body model with simple exponential relationship between the thermal resistance
and the temperature drop [29,49] or with consideration of the complete heat transfer modes
(i.e. natural and forced convection, heat radiation, and heat conduction) [16] in which all
electrical losses are perceived as an input of equally distributed heat and assumed to be the
sum of the load and no-load losses given by the transformer design.

Moreover, for indoor transformers, the heat transfer path from inside to the outside of the
station should be taken into consideration, which may contain the natural or forced venti-
lation, the heat conduction through the station wall and door, and the cross radiation be-
2.2 Model Training 15

tween the transformer tank and the adjacent wall each of which affects the performance
of the transformer’s radiator [50, 51]. In the same way, for underground distribution trans-
formers, modeling of the vault has a significant impact on the accuracy of the thermal
model [52–54].

2.2 Model Training

Each thermal model has some parameters which are either design-dependent or empirical
parameters. The accuracy of thermal models depends upon these parameters. In the case
of the conventional models, the parameters are the temperature rise over ambient tempera-
ture at rated load and the oil exponent which can be determined either using heat run test
or using measurements during normal operation of transformers [8]. However, for most
transformers especially old ones, heat run test results are not available; moreover, provided
that they are available, the result of modeling will be conservative [55]. Therefore, these
parameters should be determined using measurements during normal operation of trans-
formers [55].

In the same way, advanced thermal models use several empirical and design-dependent pa-
rameters (e.g. surface of heat convection, length of oil ducts, and etc.). However, for most
of transformers, no design details of the transformer is available for usage in a thermal
model. Therefore, the advanced thermal models should be also parametrized using opera-
tional data. The measured quantities depend upon the application and the complexity of the
thermal model.

The model training is a fitting problem aiming to determine the parameters of the ther-
mal model (φ) using measured values during normal operation of transformers denoted by
X in such a way that the sum of squares of the deviation of the calculated temperature
(f (X, φ)) from the measured temperature (Y ) is minimized. This can be formulated as
follows:
minimize kY − f (X, φ)k22 . (2.15)

However, the method used for the determination of the parameters depends on the thermal
model. If the thermal model is linear, the linear least-squares method can be used [10, 44].
Nevertheless, for non-linear models, the most common method for parameterizing of the
models is the non-linear least-squares method [13, 15, 55, 56].
16 2 Literature Survey

2.3 Factors Affecting the Temperature Rise

There are several factors affecting the accuracy of the thermal models. Not all of them
can be easily used in the dynamic thermal models since they need some unconventional
additional information about some parts of transformers which are not available normally.
Accounting for these factors may increase the complexity of the models. Therefore, based
on the application of the thermal model, one or more of the affecting factors can be elimi-
nated from the considerations.

2.3.1 Viscosity

The viscosity and the temperature of the cooling medium of a transformer have mutual
impacts. Although the variation of the oil viscosity affects the top-oil temperature rise and
the oil time constant, it is assumed that the oil viscosity and the winding electrical resistance
have opposite impacts on the variations [57, 58]. Due to the temperature coefficient of
the winding electrical resistance, the DC part of the load loss of transformers (RI 2 ) will
increase with the temperature whereas the eddy loss will decrease with the temperature [28].
Moreover, as the load losses change, the average oil and winding temperature rises and the
winding and oil time constants change respectively [8]. The load increase may lead to a
faster oil flow since the viscosity of the oil changes with the temperature [59]. Whether the
faster oil flow due to natural convection necessarily leads to better transferring of the heat
to the ambient is a complex phenomenon and should be investigated accurately. Since as
the temperature changes, not only the viscosity but also the winding electrical resistance as
well as other thermal characteristics (e.g. density, thermal conductivity, etc.) of the medium
will change.

However, investigations have shown that the loading guides are not accurate enough if
the ambient temperature is below 0 °C since the thermal models proposed in the load-
ing guides do not account for the oil viscosity and the winding resistance [60]. More-
over, the hot-spot location may change as the load changes which could be attributed to
changes in flow patterns around the disc coils as the viscosity changes [61]. Furthermore,
the oil viscosity is especially important under cold-load pick up conditions where the oil
viscosity can lead to a temporary excessive temperature rise in the winding when the load
is suddenly applied to a thoroughly cooled transformer at a very low ambient tempera-
ture [58, 62].

Moreover, different equations as a function of temperature have been proposed for viscos-
ity of different oils in the literature. Grubb et al in [63] provided some experimentally-
2.3 Factors Affecting the Temperature Rise 17

derived equations for the thermal properties of six different transformer oil types among
them Naphthenic, Silicone oil, Askarel, and Paraffinic. The behavior of the oil viscos-
ity of the Paraffinic and Naphtanic oils at very low temperature between -50 °C and 0 °C
have been studied in [64] and the relevant regression-based relationships between the vis-
cosity and the temperature for these types of oils have been proposed. Furthermore, for
ester oil FR3, a relationship between the viscosity and the temperature has been proposed
in [65].

The comparison of the mineral oil with the ester oil shows that under the same conditions,
the top-oil temperature, the hot-spot temperature, and the average winding temperature rise
of power transformers filled with mineral oil is lower than that of the ester oil due to the
higher viscosity of the ester oils [66]. On the contrary, for high temperature application
such as wind farm transformers, the top-oil temperature can reach 135 °C which is near the
flash point of mineral oil; however, natural and synthetic ester oils have significantly higher
flash points [67].

2.3.2 Harmonic Contamination of Load Currents

The existence of non-linear and unbalanced loads and the usage of power electronic con-
verters lead to the contamination of the load current with harmonics. The harmonic current
can consequently increase the oil temperature rise as well as the winding temperature gra-
dient of transformers by affecting the load-losses (by increasing the load factor, the eddy
and the stray losses) [68]. Under harmonic loads, the loading capability of transformers
is reduced. Moreover, due to the presence of the stray losses in the structural parts or the
tank, local overheating may also occur which is not considered in the loading guides equa-
tions [58]. IEEE C57.110 [69] provides a method for calculation of the actual losses under
harmonic currents in which at least the eddy current loss should be known. Moreover, an-
other method has been proposed in [70] for this purpose where the rated stray losses in
components other than winding is necessary to be known. However, for most of transform-
ers especially old transformers, these losses are unknown since the manufacturers do not
provide the decomposition of load losses.

2.3.3 Environmental Factors

The maximum peak load that a transformer can carry depends on the service conditions
at the time of applying the load and the maximum operating temperatures allowed by the
18 2 Literature Survey

user [71]. The dissipation of the heat generated inside a transformer is affected by ex-
ternal conditions such as wind velocity and its direction, and solar irradiation which are
continuously changing. Therefore, environmental conditions affect the temperature rise of
transformers and explicitly so their loading capabilities.

2.3.3.1 Solar Irradiation

The rated temperature rise under continuous nameplate load is based on tests in a virtually
still air environment with no solar or other radiation input to the transformer [58]. However,
for outdoor transformers especially those directly exposed to the solar radiations, the sun
intensity is of important external parameters which may increase the temperature rise and
consequently decrease the loading capability of the transformer [71]. The high sun intensity
may impose additional losses to the transformer to a degree that may reach 25% of the total
losses of the transformer [72]. Nevertheless, the loading guides do not account for the
incident solar energy or varying weather conditions and their effects on the temperature of
transformers.

The solar irradiation power can be either calculated based on the declination angle of the
sun and the latitude of a certain place [72–74] or based on the measurement of the solar
irradiation. The effect of the solar irradiation on the temperature rise can be considered in
thermal models as:

• an additional loss added to the total loss of the transformer [72, 73].

• a heat flux affecting the thermal resistance corresponding to the thermal radiation
[16,74,75]. In this way, the effect of tank color on the absorption of the incident solar
radiation may be taken into account. The tank color affects the emissivity constant of
the transformer tank [76].

• a time varying factor affecting the ambient temperature [77].

2.3.3.2 Wind

Wind can be considered as an extra forced convection which enhances the heat transfer
coefficient on the air-side of the transformer and thus changing the performance of the
radiator depending on the speed and direction of the wind [75, 78]. No significant research
is dedicated to the effect of the wind on the cooling of transformers specially wind farm
transformers which are exposed directly to strong winds.
2.3 Factors Affecting the Temperature Rise 19

2.3.4 Tap Changer Position

The rated oil temperature rise, the losses including stray losses and DC losses, and the
winding gradient will change if the tap position of a transformer changes; hence, these
values should be measured or calculated for a certain tap [8, 58]. Accounting for the tap
changer position according to the loading guides has the merit that the procedure is simple
enough to integrate into the thermal model of online monitoring systems. IEC 60076-7
loading guide [8] introduces a linear relationship between the loss ratio and the tap position
as is shown in Fig. 2.4. However, IEEE loading guide [9] uses the total losses at rated load
instead of the loss ratio.

In order to take into account the tap changer position in a thermal model, Radakovic et
al in [79] calculate the stray losses and DC losses and the hot-spot factor for different tap
position of a transformer accurately using finite element method for seven load steps and
used a 3-body model proposed in [15] to calculate the oil temperature in each time step, and
parametrized this model with a thermal-hydraulic network model proposed in [80]. More-
over, Djamali et al in [81] used an advanced thermal model with the IEC 60076-7 method
for consideration of the tap changer position. In addition to the conventional inputs used
¤
in a thermal model, they used the tap changer position which was measured during normal
operation of the transformers by an online monitoring system.

2.3.5 Geomagnetically Induced Current (GIC)

GIC leads to flowing a quasi-stationary current between two substations with solidly
grounded neutral of power transformers [82]. This may result in displacement of the op-
erating point of transformer into the saturation area of the magnetic circuit. But power

Rmin

Rmax
m2
m1
Rr+1
Rr

tapmin tapr tapr+1 tapmax X

Figure 2.4: Change of the loss ratio (R) with tap changer position (tap) [8].
20 2 Literature Survey

transformers are designed to be operated in the linear region of their magnetizing charac-
teristic. If a transformer is operated in the saturation area, a half-cycle saturation may occur
leading to a greater share of flux leaking out beyond the core, inducing additional eddy
currents in various parts of the core and winding assembly, potentially leading to additional
core and winding losses as well as harmonic contents of the load current [83]. These effects
may result in excessive local overheating in some parts of a transformer [84]. In addition,
due to the saturation of the core, a part of the main flux strays to the tank which causes
higher losses and temperature in the tank [84]. It is worthwhile to note that the high voltage
direct current transmission lines which are located beside an AC transmission line on the
same tower may induce a DC current in the AC circuit. This DC current flows through the
power transformer resulting in the same effects as GIC [85].

2.3.6 Moisture

The lifetime estimation of transformers depends not only on the operating temperature but
also on the moisture content of the solid insulation [86]. Moreover, the thermal properties
of the oil and cellulose are affected by the temperature and moisture which may lead to the
increase of the temperature gradient of the winding and consequently decrease the loading
capability of transformers [87]. Hence, taking the moisture content into account as an input
of the thermal model with an additional thermal resistance representing the heat conduction
in the solid insulation and a thermal capacity, and using online moisture-in-oil measurement
leads to a more accurate prediction of the hot-spot temperature and the top-oil temperature
[88].

2.4 Applications of the Thermal Models

2.4.1 Prediction of the Hot-Spot Temperature

One of the main applications of top-oil temperature models is the prediction of the hot-
spot temperature. The highest temperature of the winding insulation depends mainly on
the loading conditions and should be under a certain value specified based on the insulation
thermal class [35]. Moreover, the relative aging rate of power transformers depends on its
hot-spot temperature [8, 58]. Therefore, the determination of the hot-spot temperature is a
vital task for owners of power transformers in order to make the best use of their assets. The
hot-spot temperature can be either measured by means of sensors installed in transformers
or it can be calculated based on the top-oil temperature and the load current. However, the
2.4 Applications of the Thermal Models 21

hot-spot temperature is not a competent measured value for modeling and monitoring of
transformers for the following reasons:

• It is an expensive and not reliable technology.

• Installation of temperature sensors can only be done at manufacturing stage.

Retrofitting on previously installed transformers is not possible.

• The inter-dependency of the factors affecting the hot-spot temperature of a trans-

former is not possible to be investigated by means of measurement.

• The location of the hot-spot temperature can not be revealed using measurement since
it changes with operating conditions of transformers.

Fig. 2.5 shows the simplified temperature change inside a transformer [71] based on which
the loading guides have been built up. The very simplified governing equation for calcula-
tion of the hot-spot temperature θhs is as follows [9, 21]:

θhs = θto + HF · gr · K y , (2.16)

where HF is the hot-spot factor, gr is the winding-oil temperature gradient at rated load,
and y is the winding exponent.

Moreover, IEEE Annex G [9] provides a different way with the same temperature rises
for calculation of the hot-spot temperature which is able to handle the time lag between
the top-oil temperature rise and the oil temperature rise in the winding cooling ducts [89].
However, the hot-spot temperature changes dynamically with load variation and rises much
faster at the load increase than what an exponential function based on the time constant
of the top-oil temperature would predict [41]. Furthermore, the dynamic behavior of the
hot-spot temperature has an overshoot whose value depends on the load of the transformer.
Subjecting a cold transformer to a step load change from low load to high load or vice versa
will lead to different overshoot values [48]. For online monitoring purposes or in cases in
which the input data are discrete, the dynamic behavior of the hot-spot temperature over
time can be then written as follows [8, 41]:

θhs (i) = θto (i) + ∆θw (i) , (2.17)

∆θw (i) = ∆θw1 (i) − ∆θw2 (i) , (2.18)
Dt
D∆θw1 = [K21 · HF · gr · K y − ∆θw1 ] , (2.19)
K22 · τw
K22 · Dt
D∆θw2 = [(K21 − 1) · HF · gr · K y − ∆θw2 ] , (2.20)
τto
D∆θw1,2 (i) = ∆θw1,2 (i) − ∆θw1,2 (i − 1) , (2.21)
22 2 Literature Survey

Top-oil level
Top-oil temperature Top-oil temperature
(OF cooling) (ON and OD cooling)
Bottom oil
bypassing coil Winding
Oil from
top duct DqH hottest spot
temperature
Top coil
Vertical Height

Average oil Average winding

temperature temperature

Cooling oil
temperature
Winding
temperature

Bottom coil

Dqw

qa
æ 1 + RK 2 ö 2(DqAo,W - DqBO) HF × Dqw
Dq BO çç ÷÷
è 1+ R ø
Temperature (°C)

Figure 2.5: Oil and winding temperature changes from the bottom to the top of the winding
according to the IEEE C57.91 [71].

where ∆θw is the winding gradient at the considered load, ∆θw1 and ∆θw2 are the winding
gradients corresponding to the increase and decrease of the load to a certain level respec-
tively (in the case of the load increase ∆θw2 is equal to zero and in the case of the load
decrease ∆θw1 is equal to zero), K21 and K22 are constants depending on the cooling sys-
tem of the transformer, and τw is the winding time constant.

Nevertheless, there are some simplifications in these models for the prediction of the hot-
spot temperature leading to the limitation of the accuracy as follows:

• It is assumed that the oil temperature measured in the temperature pocket of ON and
OD transformers is equal to the maximum oil temperature. Yet, measurements using
fiber optics in the winding outlet show higher temperature (see Fig. 2.6). This as-
sumption leads to overestimation of the calculation of the ∆θw during heat-run test
since it is normally calculated as the difference between the mean winding temper-
ature rise and the average oil temperature rise (average top and bottom temperature
rise) [90] which is referred to the winding gradient in the loading guides. The wind-
2.4 Applications of the Thermal Models 23

Figure 2.6: Measurement Error of the hot-spot temperature due to incorrect or inaccurate
measurement of top-oil temperature rise and winding temperature rise [91]. (i) Wrong
top-oil temperature rise at the radiator inlet and the assumed winding temperature. (ii)
Wrong hot-spot temperature calculated based on i. (iii) Correct top-oil temperature rise
and the corresponding winding temperature. (iv) Correct winding temperature profile
measured by optical sensors. (v) Correct hot-spot temperature.

ing gradient is calculated using DC resistance of the winding during heat-run test.
Moreover, for OF transformers, due to the oil mixture, the measured oil tempera-
ture is significantly lower than the oil temperature next to the winding at the hot-spot
location [75].

• It is assumed that the winding temperature increases linearly from bottom of the wind-
ing to the top of the winding (see Fig. 2.6) which may lead to inaccurate calculation
of the winding temperature [91].

• It is assumed that the hot-spot location is at the top of the winding. However, it is a
matter of fact that there are some turns that have higher temperatures than the temper-
ature at the top of the winding, which is mainly due to the magnetic flux leakage and
fluctuations in the local oil flow inside the winding (see Fig. 2.6).

• According to the loading guides, it is assumed that the hot-spot factor is a constant
value; however, its value depends on several parameters, namely loading condition,
type of cooling duct in the winding, type of cooling system, water and acidity con-
tents, oil velocity in the radial cooling ducts, ratio of DC to additional loss, mixing of
the hot oil in the cooler, average value of the losses over all conductors of the winding,
and etc. [6, 91, 92].
24 2 Literature Survey

It is worthwhile to note that the top-oil temperature models introduced in section 2.1 are
accompanied by an hot-spot temperature model which is more or less based on the same
concept used for the top-oil temperature modeling, where the winding temperature rise
over the top-oil temperature is used instead of the top-oil temperature rise, the winding
time constant is used instead of the oil time constant, and the total loss contains only the
copper losses.

2.4.2 Estimation of Loss of Life

Since the temperature distribution is not uniform in transformers, the part which is operating
at the highest temperature will be exposed to the greatest deterioration. Therefore, in aging
studies, it is usual to consider the aging effects produced by hot-spot temperature [58].
The degradation time of homogeneous cellulose systems can be derived as follows [93,
94]:
1 1 −Ea
− = Ae R·θhs (i) · t (i) , (2.22)
DP (i) DP (i − 1)
where DP is the degree of polymerization, A is a constant, Ea is the activation en-
ergy required to start the reaction (J · mol−1 ), R is the ideal gas constant equals to
8.314 (J · mol−1 · K −1 ), θhs is the hot-spot temperature (K), and t is the degradation
time.

According to the IEEE C57.91 loading guide clause 5 and annex 5, the aging acceleration
factor is calculated as (2.23) based on (2.22) where the ERa is considered equals to 15000
based on the experience and the maximum allowable temperature of the winding equals to
110 °C [58]: h i
15000
− θ 15000
FAA = e 383 hs +273 . (2.23)

The equivalent aging factor in a specified period of time for a given temperature cycle is as
follows:
N
P
FAA,i · ∆ti
i=1
FEQA = N
, (2.24)
P
∆ti
i=1

where FAA,i is the aging acceleration factor for the temperature that exists during time
interval ∆ti , and N is the total number of time intervals.

Finally, the loss-of-life is calculated as follows:

FEQA · t
LOL = , (2.25)
Normal insulation life
2.4 Applications of the Thermal Models 25

where the normal insulation life is the nameplate life under nominal loading condition at
normal ambient temperature.

However, IEC loading guide uses the relative aging rate for non-thermally upgraded paper
as follows [8]:
(θhs −98)
V =2 6 , (2.26)

which means that the aging rate of the insulation is doubled for each 6 degrees of tem-
perature increase above the maximum allowable insulation temperature. For thermally-
upgraded paper, IEC loading guide uses (2.24) to calculate the relative aging rate. The
loss-of-life over a certain period of time is equal to:

N
X
LOL ≈ Vi × ti , (2.27)
i=1

where Vi is the relative aging rate in time step i and ti is the ith time inter-
val.

It is worthwhile to note that no significant difference exists between the IEC and IEEE
models [95]. Although the methods proposed in the loading guides are widely used in the
literature and industry, aging of insulation is a function of temperature, moisture content,
and oxygen content [95–97]; therefore, the DP in (2.22) should be used to estimate the life
expectancy of a transformer accurately. The moisture and oxygen contents affect the value
of A in (2.22) [98] meaning explicitly that the loss-of-life equations according to loading
guides are valid only under constant moisture and oxygen contents.

2.4.3 Determination of the Overload Capability

There are three sorts of transformer overloading regimes: the normal cyclic loading which
is proposed for planned overload, the long-term emergency overload resulting from the pro-
longed outage of some system elements that will not be reconnected before the transformer
reaches a new and higher steady-state temperature, and the short time emergency overload
resulting from the occurrence of one or more unlikely events which seriously disturb nor-
mal system loading [8]. In order to prevent the improper operating conditions, an accurate
supervision is necessary in all sorts of overloading.

According to IEC loading guides [8], a higher-than-rated load may be applied on a trans-
former, but regarding the relative thermal aging rate, this loading should be equivalent to
the rated load at factory ambient temperature. This is achieved by taking the advantage
of low ambient temperature during the load cycle. In other words, long-term overloading
26 2 Literature Survey

can be achieved as long as the hot-spot temperature of the transformer does not exceed
the allowable hot-spot temperature corresponding to unity relative aging rate. Therefore,
prediction of the hot-spot temperature is a key word for determining the loading capabil-
ity of a transformer. In order to have a unity relative aging rate during long-term over-
load periods, an accurate thermal model should be used for the calculation of the hot-
spot temperature. Thereby, the ambient temperature-dependent loading of a transformer
can be determined in such a way that the hot-spot temperature remains below a certain
value [86, 99, 100].

Moreover, limitations concerning other parts of a transformer, e.g. bushings and on-load
tap changer during overloading periods, should also be kept in mind. IEEE loading guide
Annex B [58] and IEC 60076-7 cl. 5 [8] provide a deep insight into the effects of the loading
beyond the nameplate rating of a transformer and its drawbacks and what should be taken
into consideration when it is decided to overload a transformer.

2.4.4 Online Monitoring of Cooling System

The performance of the cooling system has a direct impact on the temperature of the wind-
ing and consequently on the lifetime of power transformers. The early detection of fan
failures based on a proper real-time online algorithm leads to increasing of the loading ca-
pability of transformers which is not well reported in the literature.

The loss of one or two fans means a severe failure for some designs of ONAF transformers;
however, for OD and OF cooling types, one or two inoperative fans with the pump still in
operation is not a severe failure but the failure of one or two pumps causes an excessive
winding temperature gradient [58]. These failures in addition to other failures related to the
cooling system such as speed reduction of fans due to contamination may not be detected
without online monitoring system of the cooling system. Different methods have been in-
vestigated in the literature for failure detection of the cooling system:

• The thermal resistance between oil at the cooler inlet and the surrounding ambient
describes the efficiency of the cooling system which can be used for monitoring of
the cooling system [101]. Nevertheless, this failure detection method may be slow.

• Big changes in the estimated parameters of a top-oil temperature model can be used
for fast failure detection of the cooling system [102, 103].

• The error in the calculation of the top-oil temperature can also be a fast method for
malfunction detection of the cooling system [27].
2.4 Applications of the Thermal Models 27

2.4.5 Other Applications

The integration of Photovoltaic (PV) and solar panels may increase the loading of distribu-
tion transformers [104]. On the contrary, the high penetration of Plug-in Electric Vehicles
(PEV) results in an increased load, reverse power flow and injection of harmonics into
the power system which can dramatically reduce the life expectancy of distribution trans-
formers [70, 105]. But the combination of the PV and PEV may increase the loading of
a distribution transformer because the PV shifts the peak load applied on the distribution
transformer to another time of the day, and therefore, reduces the temperature of the trans-
former [106]. These effects can be investigated by means of a smart charging algorithm
which accounts for the temperature and loss-of-life of transformers [107] leading to the re-
duction of the thermal aging of transformers. In this context, a multi-objective optimization
can be performed which makes a trade off between the loss-of-life of a transformer and
service quality of PEV charging [108].
 29

3 Improved Dynamic Thermal Model

"We make to ourselves pictures of facts. The picture is a model of reality."

Tractatus Logico-Philosophicus
L. Wittgenstein

D
IFFERENT dynamic thermal models have been introduced in the previous section with
the aim of providing a simplified solution for complex three-dimensional physical
phenomena of heat transfer inside transformers in the form of critical transformer temper-
atures (i.e. top-oil and hot-spot temperatures) that can be applied for real-time monitoring,
diagnostics, and transformer protection applications [6]. However, in some cases, their ac-
curacy is limited since they use a constant thermal resistance and in other cases, they are
complicated to be used in online monitoring systems and it is necessary to use some charac-
teristics of transformers which are normally not provided by the transformer manufacturers.
Therefore, although the researchers have invested more than hundred years on the topic of
dynamic thermal modeling of transformers, this theme is still open and more researches
may be assigned in this direction.

In the rest of this section, a thermal model for calculation of the top-oil temperature in
transformers with different cooling systems is proposed which utilizes different heat trans-
fer phenomena, namely natural and forced convection in oil and air, radiation, and conduc-
tion in the steel plate of the transformer tank. Moreover, a variation of the proposed model
is presented which is applicable to indoor distribution transformers for calculation of the
top-oil temperature using calculated temperature of the transformer tank. In this way, the
sensors can be installed during in-service operation of distribution transformers; moreover,
the model is also applicable to transformers having no temperature pocket for calculating
the top-oil temperature.
30 3 Improved Dynamic Thermal Model

3.1 Power Transformer Thermal Modeling

3.1.1 Thermal Modeling Concept

The differential equation for calculation of the top-oil temperature in oil-immersed trans-
formers as described in section 2.1 is as follows:

dθto
τto · + θto − θa = ∆θto , (3.1)
dt

where θto is the calculated top-oil temperature (°C), θa is the ambient temperature (°C), ∆θto
is the top-oil temperature rise over ambient temperature (K) at the considered load, and τto
is the oil time constant (s) at the considered load. The difference equation corresponding to
(3.1) for the ith time step is as follows:

Dt (i)
θto (i + 1) = θto (i) + (∆θto (i) − θto (i) + θa (i)) , (3.2)
τto (i)
∆θto (i) = Qtot (i) · Rth (i) , (3.3)
Qtot (i) = P0 + PK · K 2 (i) , (3.4)
τto (i) = Rth (i) · Cth , (3.5)

where Dt is the time between two sequential measurements, P0 is the no-load loss (W), PK
is the short-circuit loss (W) which depends on the temperature [11] and the tap changer posi-
tion [8], K is the load factor of the transformer (%), Rth is the thermal resistance (K/W), and
Cth is the thermal capacity (W·s/K) and is calculated as follows [58]:

0.1323 0.0882 0.4

Cth = × mact + × mtube + × moil . (3.6)
3600 3600 3600

As it was mentioned before, the thermal resistance is considered as a constant value in the
conventional models leading to the limitation of their accuracy; nevertheless, it is affected
by several factors such as the number of running fans and pumps, the load, and the am-
bient temperature. The accurate determination of the thermal resistance leads to the more
accurate calculation of the top-oil temperature.

3.1.2 Determination of the Thermal Resistance

The proposed model in this section is based on modeling of different heat transfer modes
inside transformers. Fig. 3.1 illustrates the oil flow path inside OFAF, ODAF, ONAF, and
ONAN transformers. The oil enters the winding at point A in Fig. 3.1a and is heated within
3.1 Power Transformer Thermal Modeling 31

the winding, streams upwards due to natural and/or forced convection of the oil and leaves
the winding at point B. The OFAF cooling method is characterized by the mixing of the
oil heated up in the winding and the oil streaming along the tank wall through the shunt
path. Therefore, the temperature of the oil entering the cooler (at point C in Fig. 3.1b) is
lower than the maximum oil temperature. Hence, in order to model the top-oil temperature
precisely, the oil mixture should also be considered. It is noteworthy that due to the direct
pumping of the oil into the winding, the warm and cold oil are not mixed significantly in
the OD cooling type. So point B and point C in Fig. 3.1c are assumed to be approximately
at the same temperatures. Afterwards, the heat is transferred to steel plates of radiators.
The steel plates are warmed up and the absorbed heat is dissipated into the air by radiation,
natural and/or forced convection [75]. Moreover, as it can be seen from line AB in Fig.
3.1b,c,d,e, the temperature change from the bottom of the winding to its top for OF and ON
cooling types are concave, compared with OD cooling type where the temperature change
is approximately linear from the bottom to the top. The reason is that in ON and OF cooling
types, the oil streams upward in the winding due to buoyancy-driven forces, whereas in OD
cooling type, the oil is pumped directly to the winding. On the contrary, in transformers
with AF cooling type, the oil in the radiator cools down faster than the transformers with
AN cooling type and therefore the line CD in Fig. 3.1b,c,d is concave but in Fig. 3.1e is
linear.

The thermal-electric analogy of the heat transfer modes inside transformers proposed in
this section is depicted in Fig. 3.2 in which all of the thermal resistances are temperature-
dependent. The reference temperatures for calculation of the thermal resistances are af-
fected by the number of running pumps and fans as well as the load factor and the ambient
temperature. As it can be seen from Fig. 3.2, the heat generated in the active parts is

B B‘ Height/m
C Height/m Height/m Height/m
C C C C
Fan
B‘ B B B B

D D D D
A A A A
Pump
D
A Temperature/°C Temperature/°C Temperature/°C Temperature/°C
(a) (b) (c) (d) (e)

Figure 3.1: (a) Schematic of the oil flow path, (b) oil temperature change inside OFAF
power transformers, (c) oil temperature change inside ODAF power transformers, (d) oil
temperature change inside ONAF power transformers, (e) oil temperature change inside
ONAN transformers [75].
32 3 Improved Dynamic Thermal Model

Dq w Dq BC Dqto

qw qB qC R
air,convection
Roil , convection R BC
R radiation
Qtot
C th qa

Roil , convection Convection of the oil R radiation Radiation

R BC Oil mixture Rair , convection Convection of the air
Qtot Total losses C th Thermal capacity
qa Ambient temperature

Figure 3.2: Thermal-electric analogy of the proposed top-oil temperature model.

represented as a current source [109]. The heat is transferred from the winding at the top
height with the temperature of θw to the oil with the temperature of θB by either natural or
forced, or combined convection denoted by Roil,convection . If there are pumps in operation
in OD transformers, the natural convection is negligible although the forced convection is
always accompanied by the natural convection since the oil flow due to the operation of
pumps is significantly higher than the oil flow due to the natural convection. On the con-
trary, if there is no running pump in OD transformers, the heat is transferred to the oil by
natural convection [78]. Moreover, for OF transformers, the operation of pumps leads to
faster oil movement in the radiators but the oil streams upward within the winding due to
buoyancy-driven forces (natural convection). In order to take into consideration the tem-
perature drop between the top of the winding and the cooler inlet (line BC in Fig. 3.1b),
due to the oil mixture in an OFAF transformer, a thermal resistance is considered in the
model which is denoted by RBC . Neglecting the heat transfer by conduction in the steel
plate of the radiator, the heat is dissipated from steel plate to the surrounding ambient by
radiation denoted by Rradiation and convection denoted by Rair,convection . It should be noted
again that for the same reason stated before on the oil-side, either the forced convection
or the natural convection, or the combined convection should be considered on the air-side
too.

The heat generated in the windings is dissipated to the surrounding ambient using two
coolants, namely oil in the tank and air outside of the tank. Hence, the governing equations
of the oil-side and the air-side in the modeling procedure are separated with consideration
of their mutual impacts and inter-dependencies. The oil-side equations consist of two parts,
3.1 Power Transformer Thermal Modeling 33

namely the oil convection and the oil temperature drop due to the oil mixture at the top of
the tank. The air-side equations contain two thermal resistances corresponding to the heat
radiation and the air convection.

It is worthwhile to note that all of the thermal resistances proposed in this model are depen-
dent on the thermal properties of the oil and the air whose values change with temperature.
These dependencies are listed in Tables A.1 and A.2.

3.1.2.1 Oil Convection

The oil convection is one of the three following phenomena:

- Natural Convection: In ON and OF transformers or OD transformers which are oper-

ating in ON mode, the heat is transferred with natural convection of the oil.

- Forced Convection: In OD transformers, the dominant part of the convection of the

oil is the forced convection representing the contribution of the pumps in transferring
the heat to the radiators.

- Combined Natural and Forced Convection: In OD transformers in which the oil ve-
locity in oil channels is not high due to the malfunction of the pumps or small number
of operative pumps, the forced convection is not the dominant part of the convection;
therefore, the convection of the oil is a combined natural and forced convection.

Natural Convection of the Oil: The thermal resistance (K/W) corresponding to the natural
convection of the oil can be written as:

1
Roil,convection = , (3.7)
Aoil · hoil,nat

where Aoil is the surface of heat transfer (m2 ) of the oil which is a design-dependent pa-
rameter, and hoil,nat is the heat transfer coefficient (W/(m2 ·K)) for natural convection of the
oil defined as follows [110]:

κoil (T )
hoil.nat = · (qoil · (Groil (T ) · P roil (T ))poil ), (3.8)
soil

where κoil is the thermal conductivity (W/(m·K)) of the oil, T is the reference temperature
for calculation of the thermal properties of the oil, soil is the characteristic length of heat
transfer on the oil-side (m) considered as a design-dependent parameter, qoil and poil are
empirical factors, Groil and P roil are the Grashof number and the Prandtl number on the
34 3 Improved Dynamic Thermal Model

oil-side respectively. The Grashof and Prandtl numbers are also temperature-dependent
numbers described as follows [110]:

g.βoil · ∆θw
Groil (T ) = s3oil · , (3.9)
(νoil (T ))2
νoil (T ) · ρoil (T ) · cpoil (T )
P roil (T ) = , (3.10)
κoil (T )
∆θw = gr · k y , (3.11)

where νoil is the kinematic viscosity (m2 /s), ρoil is the fluid density (kg/m3), cpoil is the
specific heat (W·s/(kg·K)), g=9.8 is the acceleration due to gravity (m/s2 ), βoil is the vol-
umetric thermal expansion coefficient (1/K), ∆θw is the temperature gradient between the
winding and the oil at the top of the winding (K), gr is the average winding to average oil
temperature gradient [K] at the rated load, and y is the winding exponent. If the results
of the heat-run test of the transformer are available, gr and y are known constants of the
model, otherwise the former can be considered as a design-dependent parameter and the
latter as an empirical factor.

Forced Convection of the Oil: The thermal resistance of the forced convection of the oil
due to the operation of pumps is as follows [110]:

1
Roil,convection = , (3.12)
Aoil · hoil,f orced

where hoil,f orced is the heat transfer coefficient for forced convection defined as follows
[110]:
κoil (T )
hoil,f orced = · Coil · (Rem n
oil (T ) · P roil (T )), (3.13)
soil
where Coil , m, and n are empirical factors, and Reoil is the Reynolds number defined as
follows [110]:

Vm,oil · doil
Reoil = , (3.14)
νoil (T )
Vm,oil = uoil · NP umps , (3.15)

where Vm,oil is the average oil velocity assumed to be proportional to the number of running
pumps (NP umps ) with a proportionality constant uoil considered as a design-dependent fac-
tor, and doil is the characteristic diameter of the oil-side considered as a design-dependent
factor.
3.1 Power Transformer Thermal Modeling 35

Combined Natural and Forced Convections of the Oil: The thermal resistance of the com-
bined natural and forced convection can be written as follows [78]:

1
Roil,convection = , (3.16)
Aoil · hoil,combiend

0.25
hoil,combiend = h4oil,nat + h4oil,f orced . (3.17)

3.1.2.2 Oil Mixture in OFAF Transformers

The oil temperature drop between the top of the winding and the cooler inlet (line BC in
Fig. 3.1) which models the oil mixture at the top of the tank in OFAF transformers can
be considered as a thermal resistance. If the temperature rise along the parallel path (line
DB’ in Fig. 3.1b) is neglected, the thermal resistance due to the oil mixture (RBC ) can be
calculated as follows:
(1 − x′ ) · (θB − θB′ )
RBC = , (3.18)
(P0 + PK · K 2 )
where x′ is the oil mixture weight considered as an empirical factor whose value is less
than 1 and indicates how the temperature in the cooler inlet is affected by the top-oil tem-
perature and the bottom-oil temperature., θB is the oil temperature at the top of the winding
corresponding to the point B in Fig. 3.1, and θB′ is the temperature of the point B’ in Fig.
3.1 which is approximately equal to the bottom-oil temperature.

3.1.2.3 Equivalent Oil-Side Thermal Resistance

The equivalent thermal resistance on the oil-side corresponding to the thermal-electric anal-
ogy depicted in Fig. 3.2 is as follows:

Rth−oil = Roil,convection + RBC . (3.19)

3.1.2.4 Air Convection

The same as the oil convection, the air convection is one of three phenomena as fol-
lows:

- Natural Convection: In AN transformers or AF transformers which are operating in

AN mode, a part of the generated heat is transferred to the surrounding ambient with
the natural convection of the air.
36 3 Improved Dynamic Thermal Model

- Forced Convection: In AF power transformers, the dominant part of the convection of

the air is the forced convection representing the contribution of the fans in transferring
the heat to the surrounding ambient.

- Combined Natural and Forced Convection: In AF power transformers in which the

air velocity is not high due to the malfunction of fans or small number of operative
fans, the forced convection is not the dominant part of the convection; therefore, the
convection of the air is a combined natural and forced convection.

Natural Convection of the Air: The thermal resistance corresponding to the natural con-
vection of the air can be written as:

1
Rair,convection = , (3.20)
Aair · hair,nat

where Aair is the surface of heat transfer (m2 ) of the air considered as a design-dependent
factor, and hair is the heat transfer coefficient for the natural convection of the air
[110]:

κair (T )
hair,nat = · (qair · (Grair (T ) · P rair (T ))pair ) , (3.21)
sair
where κair is the thermal conductivity (W/(m·K)) of the air, T is the reference temperature
for calculation of the thermal properties of the air, sair is the characteristic length of the
heat transfer on the air-side (m), qair and pair are empirical factors, Grair and P rair are the
Grashof number and the Prandtl number on the air-side respectively, described as follows
[110]:

g · βair (T ) · (θo−a − θa )
Grair (T ) = s3air · , (3.22)
(νair (T ))2
νair (T ) · ρair (T ) · cpair (T )
P rair (T ) = , (3.23)
κair (T )

where νair is the kinematic viscosity (m2 /s) of the air, ρair is the fluid density (kg/m3) of
the air, cpair is the specific heat (W·s/(kg·K)) of the air, and βair is the volumetric thermal
expansion coefficient of the air (1/K), and θo−a is the logarithmic average temperature of
the oil and the air (K) in the radiators.

Forced Convection of the air: The thermal resistance of the forced convection on the air-
side due to the operation of fans is as follows [110]:

1
Rair,convection = , (3.24)
Aair · hair,f orced
3.1 Power Transformer Thermal Modeling 37

where hair,f orced is the heat transfer coefficient for forced convection defined as follows
[110]:
κair (T ) ′ n′
hair,f orced = · Cair · (Rem
air (T ) · P rair (T )), (3.25)
sair
where Cair , m′ , and n′ are empirical factors, and Reair is the Reynolds number defined as
follows [110]:

Vm,air · dair
Reair = , (3.26)
νair (T )
Vm,air = uair · NF ans , (3.27)

where Vm,air is the average air velocity assumed to be proportional to the number of running
fans (NF ans ) with a proportionality constant uair considered as a design-dependent factor,
and dair is the characteristic diameter of the air-side considered as a design-dependent fac-
tor.

Combined Natural and Forced Convections of the Air: The thermal resistance of the com-
bined natural and forced convections can be written as follows [78]:

1
Rair,convection = , (3.28)
Aair · hair,combiend

0.25
hair,combiend = h4air,nat + h4air,f orced . (3.29)

3.1.2.5 Heat Radiation

According to Stefan-Boltzmann law, the heat transferred from the steel plate of trans-
former’s tank to its surroundings by radiation from its surface Arad is as follows
[78]:
4
− θa4 , (3.30)


Q = ς · ε · Arad · θo−a

where ς is the Stefan-Boltzmann constant (W/(m2 ·K4 )), ε is the emissivity factor, and
Arad is the effective surface of radiation (m2 ) considered as a design-dependent factor.
It should be noted that for calculating the heat transfer coefficient of the radiation, the
transformer tank and the surrounding ambient are considered as a gray body and a black
body respectively. Therefore, the thermal resistance of the heat radiation is described as
follows [78]:

1
Rradiation = 2
. (3.31)
ς · ε · Arad · (θo−a + θa2 ) · (θo−a + θa )
38 3 Improved Dynamic Thermal Model

3.1.2.6 Equivalent Thermal Resistance of the Air-Side

The equivalent thermal resistance of the air-side corresponding to the thermal-electric anal-
ogy depicted in Fig. 3.2 is as follows:

1
Rth−air = 1 1 . (3.32)
Rair,convection
+ Rradiation

3.1.2.7 Equivalent Thermal Resistance

Finally, the thermal resistance between the oil at the top of the winding and the surrounding
ambient is as follows:

∆θto
Rth = (Rth−oil + Rth−air ), (3.33)
∆θto + ∆θw + ∆θBC

where ∆θto is equal to θc −θa in the thermal-electric analogy circuit of Fig. 3.2, ∆θBC is the
oil temperature drop between the top of the winding and the cooler inlet. According to the
thermal-electric analogy of Fig. 3.2, (3.33) indicates that if the temperature drop from top
of the winding to the ambient is equal to ∆θw + ∆θBC + ∆θto and the thermal resistance ex-
isting between these two points is equal to Rth−oil + Rth−air , the thermal resistance existing
between the oil in the cooler inlet with temperature of θc and the surrounding ambient with
temperature of θa is equal to Rth . If the thermal resistance of the air-side Rth−air is only
taken into consideration, without considering the Rth−oil , the inter-dependencies between
oil-side and air-side equations, and the effects of the air-side parameters on the oil-side pa-
rameters and winding temperature will not be taken into consideration. Substituting (3.33)
for Rth in (3.3) and (3.5) and using (3.2), the top-oil temperature is calculated in each time
step.

Moreover, Table 3.1 summarizes the parameters and variables used in (3.2). The input
variables are load factor (K), ambient temperature (θa ), number of running pumps (affects
Rth ), number of running fans (affects Rth ), and time between two sequential measurements
(Dt). While the constants of the model are PK and P0 ; and the output variable is the
top-oil temperature in each time step (θto (i)). Finally, the unknown parameters include
the empirical factors and design-dependent factors which are used in the proposed model
to calculate the thermal resistance (Rth ) and top-oil temperature (θto ). Since no design
detail of the transformers are available, the empirical factors and design-dependent factors
should be estimated [8] using data measured during normal operation of transformers. The
least-squares method is used for parameter estimation in this work, which is explained in
Appendix A.1.
3.2 Distribution Transformer Thermal Modeling 39

Table 3.1: Summary of the parameters, inputs, constants, and output of the proposed model
for outdoor power transformers.

Type of parameter Parameter Description

qoil , poil , qair , pair Empirical factors of natural convection
soil , sair Characteristic length of convection
Aoil , Aair Surface area of convection heat transfer
Arad Surface area of radiation heat transfer
gr Winding temperature gradient
Estimated Parameters y Winding exponent
Coil , m, n Empirical factors of forced oil convection
uoil , doil Proportionality constant, characteristic diameter of oil ducts
x′ Oil mixture weight
Cair , m′ , n′ Empirical factors of forced air convection
uair , dair Proportionality constant, characteristic diameter of air path
Dt Time step
K Load Factor
Input Variables NF ans Number of running fans
NP umps Number of running pumps
θa Ambient Temperature
P0 No load loss
Constants (Design-Dependent)
PK Short circuit loss
Output θto Top-oil temperature

3.2 Distribution Transformer Thermal Modeling

With the integration of the renewable energy-based distributed power plants, whose gen-
erated power is uncertain, the determination of the loading of distribution transformers
appears to be crucial. Therefore, the issue of the model-based online monitoring of dis-
tribution transformers may achieve acceptance, if a cost-benefit compromise is reached.
In this section, a new thermal model is proposed for calculation of the top-oil tempera-
ture of indoor distribution transformers using calculated tank temperature which accounts
for convection, radiation, conduction, and ventilation heat transfer phenomena. Usage of
the tank temperature instead of the top-oil temperature in thermal modeling provides the
ability for installation of low-cost thermal monitoring systems for distribution transform-
ers without removing them from service. Moreover, it provides the possibility to measure
and to calculate the top-oil temperature and to determine the hot-spot temperature for dis-
tribution transformers which have no temperature pocket for installing temperature sen-
sors.

3.2.1 Thermal Modeling Concept based on Tank Temperature

The difference equation to calculate the temperature (here, the tank temperature as the char-
acteristic temperature) for indoor distribution transformers corresponding to the principle
described in section 2.1.1 is as follows:
40 3 Improved Dynamic Thermal Model

Dt (i)
θtank (i + 1) = θtank (i) + (∆θtank (i) − θtank (i) + θa (i)) , (3.34)
τto (i)
∆Ttank (i) = Qtot (i) (i) · Rth (i), (3.35)
Qtot (i) (i) = P0 + PK · K(i)2 , (3.36)
τto (i) = Rth (i) · Cth . (3.37)

The most important parameters for accurate calculation of the tank temperature using (3.34)
is the calculation of the thermal resistance for each time step (Rth (i)), whose value depends
on the loading conditions, the ambient and station temperatures, and the ventilation of the
station.

3.2.2 Determination of the Thermal Resistance

The model proposed in this section for calculation of the top-oil temperature and the tank
temperature accounts for different thermal resistances corresponding to the natural con-
vection in the oil and air, the thermal radiation, the thermal conduction in the steel plates
of the transformer tank, and the natural ventilation in the station building. Moreover, the
temperature dependency of the thermal resistances and their mutual impacts are taken into
consideration.

Fig. 3.3 illustrates the schematic of an indoor distribution transformer. The generated
heat in the winding is transferred to the surrounding oil with the convective heat trans-
fer. According to the energy balance law, the oil heats up and enters the radiator with the
temperature assumed to be equal to the top-oil temperature at point 2 in Fig. 3.3 since
the temperature of the oil measured in the temperature pocket is assumed to be equal to
the temperature of the oil at the top height of the winding. The inner surface of the tank
wall has the same temperature as the oil. The heat is transferred to the outer surface of
the tank (point 3 in Fig. 3.3) with thermal conduction which causes a slight temperature
drop across the tank wall. The oil cools down in the radiator and enters the winding at the
bottom again. The heat is dissipated to the air inside the station with natural convection
of the air and thermal radiation. The air heats up and streams upward due to the natural
ventilation in the station and leaves the station from the top opening (point 4 in Fig 3.3).
Finally, cold air enters the station from the opening at the bottom of the station building
(point 5 in Fig 3.3). It is assumed that the heat conduction across the wall and door as well
as the heat radiation from the wall and the door of the transformer station are negligible.
The thermal network corresponding to the formerly described phenomena is depicted in
Fig. 3.4.
3.2 Distribution Transformer Thermal Modeling 41

1: Winding Temperature qw
2: Top-oil Temperature q to
4 3: Tank Temperature q tank
4: Air-outlet Temperature q s
5: Air-inlet Temperature q a

2
1 3

Figure 3.3: Schematic of an indoor distribution transformer.

Dq w Dqcond Dqtank
Rradiation
Roil ,convection Rcond Rvent
1 2 3 4 5

qw qto qtank
Rair,convection
Q tot
Cth qs qa

Roil , convection Convection of the oil R radiation Thermal Radiation

R cond Thermal Conduction Rair , convection Convection of the air

R vent Air ventilation C th Thermal capacity

qa Ambient temperature Qtot Total losses

qs Station temperature

Figure 3.4: Thermal-electric analogy of the proposed model for indoor distribution trans-
formers.
42 3 Improved Dynamic Thermal Model

3.2.2.1 Natural Convection in oil

The thermal resistance corresponding to the natural convection of the oil can be calculated
using (3.7)-(3.11).

3.2.2.2 Heat Conduction Through Tank Wall

According to Wiedemann-Franz Law, at all temperatures except very low temperatures, the
ratio of the thermal conductivity to the electrical conductivity of metals is proportional to
the temperature as follows [111]:

κsteel
= L · θtank , (3.38)
Σsteel (T )
1.45 × 106
Σsteel = , (3.39)
1 + 0.00094 · (θtank − 20)

where κsteel is the thermal conductivity of the steel plates (W/m · K) of the tank wall,
Σ is the temperature-dependent electrical conductivity of the steel plates of the tank wall
(1/ (Ω · m)), L is the Lorenz number equal to 2.45 × 10−8 W · Ω/K 2 [111]. The thermal
resistance of the heat conduction can be then written as follows:

l
Rcond = , (3.40)
κsteel · Acond

where l is the thickness of the tank wall, and Acond is the surface area of the thermal con-
duction.

3.2.2.3 Natural Convection in Air

The thermal resistance corresponding to the natural convection of the air can be calculated
using (3.20)-(3.23); however, the temperature gradient in the Grashof number should be
adjusted as follows:

g · βair (T ) · (θtank − θs )
Grair (T ) = s3air · . (3.41)
(νair (T ))2

Since in the radiator, the oil and in the same way the tank cool down from top to bottom of
the radiator and the air heats up (moves) from bottom to the top of the radiator, the reference
temperature for calculation of the thermal characteristics of the air-side is the logarithmic
mean of the tank temperature and the station temperature [78].
3.2 Distribution Transformer Thermal Modeling 43

3.2.2.4 Heat Radiation in Air

The thermal resistance corresponding to the heat radiation is calculated as follows:

1
Rradiation = 2
. (3.42)
ς · ε · Arad · (θtank + θs2 ) · (θtank + θs )

If the transformer is close to the station wall, the mutual radiation between the transformer
tank and the adjacent wall should be taken into consideration which affects the emissivity
factor in (3.42) as follows [78]:

1
ε= 1 1 , (3.43)
εtank
+ εwall
−1

where εtank and εwall are the emissitivity of the station wall and the transformer tank re-
spectively.

3.2.2.5 Natural Ventilation in Transformer Station

For a single cell transformer station with two identical openings at different heights, the
buoyancy-driven air flow can be written as follows [112]:
s
|∆P | · g · H
ψstation = Cd · Aopening · , (3.44)
Pinlet

where Cd is the discharge coefficient considered as a constant value equal to 0.611 for
shaped-edged orifices [112], Aopening is the surface area of the openings considered as a
design-dependent parameters, Pinlet is the pressure of the inlet opening, ∆P is the pressure
difference between the openings of different heights, and H is the height difference of two
openings considered as design-dependent parameter. By applying the gas law at constant
pressure [112], the relationship between the pressure and the temperature can be written as
follows: s s
|∆P | |θs − θa |
= , (3.45)
Pinlet θa + 273
where θa is the absolute value of the inlet opening temperature considered equal to the am-
bient temperature; and θs is the absolute value of the outlet opening temperature considered
equal to the station temperature. Finally, according to the energy conversion law, the ther-
mal resistance of the buoyancy-driven natural ventilation in the transformer station can be
written as follows:
1
Rvent = . (3.46)
ψstation · ρair (T ) · cpair (T )
44 3 Improved Dynamic Thermal Model

3.2.2.6 Equivalent Thermal Resistance

Since the thermal resistances of the proposed model are temperature-dependent, any
changes in the temperature of each node in Fig. 3.4 will affect the value of the thermal
resistances. Hence, their inter-relationships should be taken into consideration. The equiv-
alent thermal resistance between the tank wall and the surrounding ambient outside of the
transformer station is as follows [27]:

∆θtank
Rth = · Rth,total , (3.47)
∆θtank + ∆θcond + ∆θw
Rth,total = Roil,convection + Rcond + (Rradiation ||Rair,convection ) + Rvent , (3.48)

where ∆θtank is equal to θtank −θa in the thermal-electric analogy circuit of Fig. 3.4, ∆θcond
is the temperature drop across the tank wall. According to the thermal-electrical analogy of
Fig. 3.4, (3.47) indicates that, if the temperature drop between top of the winding and the
ambient is equal to ∆θw +∆θcond +∆θtank and the thermal resistance existing between these
two points is equal to Rth,total , the thermal resistance existing between the tank in the cooler
with temperature of θtank and the surrounding ambient with temperature of θa is equal
to Rth . If the thermal resistance of the air-side is only taken into consideration, without
considering the oil-side thermal resistances, the inter-dependencies between oil-side and
air-side equations, and the effects of the air-side parameters on the oil-side parameters and
winding temperature will not be taken into consideration. Using (3.34)-(3.37) and (3.47),
the tank temperature can be calculated in each time step.

3.2.3 Calculation of the Top-Oil Temperature using Tank Temperature

In order to calculate the top-oil temperature using calculated tank temperature, the temper-
ature drop across the tank wall should be added to the tank temperature. This can be written
as follows:

θto (i) = θtank (i) + ∆θcond (i) , (3.49)

2

∆θcond (i) = Rcond (i) · P0 + PK · K (i) . (3.50)

Nevertheless, if the top-oil temperature is used for parametrizing the model, (3.34), (3.35)
and (3.47) should be adopted for top-oil temperature as follows:
3.2 Distribution Transformer Thermal Modeling 45

Table 3.2: Categorization of the parameters, inputs, constants, and outputs of the proposed
model for indoor distribution transformers.

Type of parameter Parameter Description

qoil , poil , qair , pair Empirical factors of natural convection
soil , sair Characteristic length of convection
Aoil , Aair Surface area of convection heat transfer
Acond Surface area of the thermal conduction
l Thickness of the tank wall
Estimated Parameters
Arad Surface area of radiation heat transfer
Aopenings Surface area of the openings
H Height difference between two openings
gr Winding temperature gradient
y Winding exponent
Dt Time step
K Load Factor
Input Variables
θs Air-outlet Temperature
θa Air-inlet Temperature
P0 No load loss
Constants (Design-Dependent)
PK Short circuit loss
θto Top-oil temperature
Outputs
θtank Tank temperature

Dt (i)
θto (i + 1) = θto (i) + (∆θto (i) − θto (i) + θa (i)) , (3.51)
τto (i)
∆θto (i) = P0 + PK · K (i)2 · Rth,to , (3.52)

∆θtank + ∆θcond
Rth,to = · Rth,total , (3.53)
∆θtank + ∆θcond + ∆θw
τto (i) = Rth,to (i) · Cth . (3.54)

3.2.4 Summary of the Proposed Thermal Model for Distribution

Transformers

The proposed model is an empirical thermal model which has different unknown parameters
including the empirical parameters and design-dependent parameters. Table 3.2 categorizes
inputs, constants, outputs, and the unknown parameters of the model. In addition, the
thermal properties of the oil and the air as a function of temperature used in the governing
equations are listed in Tables A.1 and A.2.
 47

4 Hardware Implementation

"That the elements of the picture are combined with one another in a definite
way, represents that the things are so combined with one another."

Tractatus Logico-Philosophicus
L. Wittgenstein

The proposed thermal model in chapter 3 has been implemented on a hardware for
monitoring purposes. The tasks of the developed monitoring system are as fol-
lows:

• Online estimation of the unknown parameters of the model listed in Table 3.1 and
3.2 after obtaining a specified number of measurements. The specified number of
measurements is provided by the user.

• Online calculation of the loading capability for the specified number of measure-
ments and online calculation of the unknown parameters of the relationship between
the loading capability and the ambient temperature. The procedure is described in
sections 7.1 and 7.2.

• The calculation of the top-oil temperature according to the proposed model in section
3.2 after parametrizing of the model.

• The prediction of the hot-spot temperature based on the proposed model after
parametrizing of the model.

• The determination of the loading capability after obtaining the dependency of the
loading capability on the ambient temperature which will be described in section
7.2.1.
48 4 Hardware Implementation

Input Terminals

4x2 4 4 4 4

4 Channels for Current 4 Channels for Temperature

Measurements Measurements

4 4

8-channel USB and I2C for

Real-Time Clock A/D Converter Communication with
other Devices
SPI

I2C I2C

Microprocessor

Figure 4.1: Concept of the developed system.

4.1 Developed Monitoring System for ONAN

Transformers

The basic concept of the developed system for ONAN transformers is shown in Fig. 4.1.
The system has four analog inputs for current measurement and four analog inputs for tem-
perature measurement. The compatible temperature sensors are PT100 sensors. The analog
inputs are digitalized using a 12 bit analog/digital converter. The digitalized measured val-
ues are send to the microprocessor and can be saved on a memory card. The proposed
thermal model and the proposed algorithm for the determination of the loading capabil-
ity (described in section 7) are implemented into the microprocessor. Using the measured
values, the top-oil temperature and the loading capability of the transformer are calculated
online. Moreover, a real-time clock has been integrated into the hardware in order for the
microprocessor to log the temperature and current at time intervals specified by user. In ad-
dition, the system is capable of communicating with other devices for sending and receiving
the data.

Fig. 4.2a and Fig. 4.2b show the front side and back side of the prototype of the devel-
oped printed circuit board. Moreover, Fig. 4.3 shows the complete monitoring system for
ONAN transformers for calculation of the top-oil temperature, prediction of the hot-spot
temperature, and determination of the loading capability of transformers. The developed
4.1 Developed Monitoring System for ONAN Transformers 49

(a) (b)

Figure 4.2: Photos of the prototype of the developed printed circuit board for logging of
the current and temperature in which the proposed thermal model and the proposed
algorithm for determination of the loading capability are embedded. (a) Front side. (b)
Back side.

Data logger DC Power

with Built-in Supply
Temperature
Transducers
Current
Transducers

Main Fuses

Figure 4.3: Prototype of the developed monitoring system for ONAN transformers.

system consists of two main fuses, a DC power supply for energizing the data logger, three
current transducers for providing the galvanic isolation between the measuring system and
current transformers, and the constructed data logger with built-in temperature transducer
shown in Fig. 4.2a and Fig. 4.2b. The whole system is placed inside a chassis with the
degree of protection IP 66 which guarantees that the system can be installed in outdoor ap-
plication and the dust and powerful water jets have no harmful effects on the system [113].
The technical specifications of the proposed monitoring system is summarized in Table
4.1.
50 4 Hardware Implementation

Table 4.1: Technical specifications of the proposed monitoring system for ONAN trans-
formers.

Power Supply 230 V, 50 Hz

4 Channels Temperature Measurements
Input Channels
4 Channels Current Measurements
Compatible Temperature Sensors PT100 (2-wire, 3-wire, 4-wire)
Temperature -40 °C to 200 °C
Measuring Range
Current 4 to 20 mA
Communication Protocol MODBUS Master/Slave
Interfaces Ethernet, I2C
Temperature <0.1% of the full scale value
Measuring Error
Current <0.1% of the full scale value
Temperature 12 bit (0.058 °C)
Resolution
Current 12 bit (0.003 mA)
Real-Time Clock 32.768 kHz
Time Synchronization RTC, NTP
Protection Class IP66

Moreover, a simplified flowchart of the operation algorithm of the proposed system is

shown in Fig. 4.4. The tasks of the system before parametrization of the thermal model can
be mentioned as follows:

• The system measures the ambient temperature, the station temperature, the oil tem-
perature, the tank temperature, and the load of the transformer and saves the measured
value.

• Once the specified number of measurements are obtained, the model will be
parametrized and the unknown parameters of the model including the design-
dependent and the empirical factors are estimated.

• Once the model is parametrized, the top-oil temperature is calculated for the formerly
measured values and accordingly the loading capability of the transformer is deter-
mined for these measured values.

• Once the loading capability and the top-oil temperature are calculated, the depen-
dency of the loading capability on the ambient and station temperatures (in case of
indoor transformers) is determined. The procedure will be described in section 7.2

It is worthwhile to note that parametrization and training of the model and the lin-
ear dependency take approximately 10 minutes to be done by the developed sys-
tem.
4.1 Developed Monitoring System for ONAN Transformers 51

Start

i=1,
n=defined by
the user

Logging the current and

i=i+1
i=i+1 i=i+1
i=i+1
temperature

Save the measured

values
No Calculation
Calculation of of the
the top-oil
top-oil i=i+1
=i+1
temperature
temperatu t re and
and the
the loading
loading
capability
cap a ility corresponding
a ab corresponding to to the
the
last
last nn measured
measured Values
Values Yes i<n

Yes No

Parametrizing
Parametrizing of of the
the model
model Is
Is the
the model
model Saving
Saving the
the obtained
obtained
using
using the
the first
f rst nn measured
fi measured No already
already dependency
dependency
values
values parametrized?
p
parametrized? ?
New
New relationship
relationship for
f r
fo
dependency
dependency of of the
the loading
loading
capability
cap
a ab on the ambient
a ility on the ambient Saving
Saving the
the estimated
estimated
temperature
temperatu
te t re requested?
requested?? parameters
parameters of
of the
the Calculation
Calculation ofof the
the top-oil
top-oil temperature
t re and
temperatu and
model
model
Yes the
the loading
loading capability
cap
a aba ility corresponding
corresponding to
to
the
the first
f rst nn measured
fi measured Values
Values

No

New
New Parametrizing
Parametrizing of
of the
the No
parametrization
parametrization Yes model
model using
using the
the last
last Is
IIs the
the dependency
dependency of
of the
the loading
loading capability
cap a ility on
a ab on
requested?
requested? nn measured
measured values
values the
the ambient
ambient temperature
t re already
temperatu already obtained?
obtained?

Yes

Saving
Saving the
the Calculation
Calculation of
of the
the top-oil
top-oil and
and hot-spot
hot-spot
calculated
calculated values
values temperatures
temperatu
t res

Displaying
Displaying the
the measured
measured and
and Saving
Saving the
the
calculated
calculated values
values and
and sending
sending Determination
Determination of
of the
the loading
loading capability
cap
a ab
a ility
calculated
calculated values
values
them
them to
to the
the superior
u erior system
sup system

Figure 4.4: Flowchart of the operation algorithm of the developed system.

After parametrization, the tasks of the system are as follows:

• Calculation of the top-oil and hot-spot temperatures and determination of the loading
capability.

• Optionally, the loss of life and the aging rate of the transformer can be estimated.

• Displaying and saving the calculated temperatures and determined loading capability
for further investigations.

• In the cases in which the user requires new parametrization of the model and the load-
ing capability-ambient temperature dependency, the system loads the last n measured
data set and parametrizes the model again.
52 4 Hardware Implementation

50
Load Factor
40
30

/%
20
10
0
Measured Ambient Temperature
50
Measured Station Temperature
45 Measured Tank Temperature

40
35
/°C

30
25
20
15
10
23−Aug 28−Aug 02−Sep 07−Sep 12−Sep 17−Sep

Figure 4.5: Measured values for parametrization of the thermal model during normal oper-
ation of the indoor distribution transformer by the developed online monitoring system.

4.2 Field Test

The developed microprocessor-based low-cost monitoring system (shown in Fig. 4.3) has
been installed on an indoor distribution transformer whose characteristics are listed in Table
4.2. Although the transformer has a temperature pocket, it is used for another application.
The measured values are the tank temperature (point 3 in Fig. 3.3), temperature of the air
in the inlet and outlet openings (point 4 and 5 in Fig. 3.3) representing the ambient and
station temperatures respectively, and currents of each phase on the low voltage side of the
transformer.

Table 4.2: Characteristics of the indoor ONAN transformer under test.

Cooling System Power /kVA Nominal Voltage/kV Total Mass/T Mass of Oil/T
ONAN 1250 10/0.4 2.85 0.605

Fig. 4.5 shows the data measured during normal operation of the transformer within
one month (middle of August to middle of September) which have been used for
parametrization of the thermal model and the loading capability-ambient temperature de-
pendency. The procedure of the parametrization will be described in details in section
5.1.

Fig. 4.6 shows the measured ambient and station temperatures, the load, and the mea-
sured and calculated tank temperatures in October, November, and December. Since the
4.2 Field Test 53

transformer serves industrial loads, it experiences very rapid changes in the loading con-
ditions. As it can be seen, the developed thermal monitoring system is able to calculate
the tank temperature accurately even when the loading condition changes rapidly, which
is very important in case of urgent overloading of transformers. The mean absolute er-
ror and the maximum error of the temperature calculation are 0.38 K and 2.00 K respec-
tively.

During field test of the constructed thermal monitoring system, the linear dependency
of the loading capability on the ambient and station temperatures are investigated in the
parametrizing phase. The procedure is described in section 7.2.3 in detail. The yielded
linear dependency is as follows:

LC = 1.4520 − 0.0097 · θa − 0.0015 · θs . (4.1)

Fig. 4.7 shows the determined loading capability of the transformer under study (see Table
4.2) using proposed model and the linear equation (4.1). As it can be seen, the linear
equation can be a good representative of the thermal model for determination of the loading
capability, where the mean absolute deviation and the maximum deviation are 0.6% and
1.7%.
30 Ambient Temperature Station Temperature
20
/°C

10
0
−10
Load Factor
30
/%

20
10
0
30

25

20
/°C

15

10
Measured Tank Temperature Calculated Tank Temperature
5
17−Oct 27−Oct 06−Nov 16−Nov 26−Nov 06−Dec

Figure 4.6: Measured ambient and station temperatures, measured load, measured and
calculated tank temperatures using proposed model.
54 4 Hardware Implementation

30
Ambient Temperature Station Temperature
20

/°C
10

−10

Loading Capability Calculated using the Thermal Model

150
Loading Capability Calculated using the Linear Equation
145

140
/%

135

130

125

120
17−Oct 27−Oct 06−Nov 16−Nov 26−Nov 06−Dec

Figure 4.7: Ambient and station temperatures, and loading capability of the transformer
calculated using thermal model and linear equation (4.1).
 55

5 Validation of the Dynamic Thermal Model

"In order to discover whether the picture is true or false we must compare it
with reality."

Tractatus Logico-Philosophicus
L. Wittgenstein

In order to use thermal models for calculation of the loading capability of transformers, their
accuracy should be firstly investigated. If the top-oil temperature is calculated accurately, it
can be guaranteed that the hot-spot temperature and the loading capability can be predicted
and determined as accurate as possible.

In order to evaluate the accuracy of thermal models, CIGRÉ technical brochure 659 [6]
suggests to use the duration curve of the error in the calculation of the top-oil temperature.
The error duration curve is the sorted absolute error in the calculation of the temperature in
Kelvin against the time sorted by the error in percent representing the mean absolute error
and maximum error of calculation. The error duration curve indicates that for X% of the
studied time, the error is less than Y Kelvin which can be also used for comparing the per-
formance of different models. In addition, the accuracy of the proposed model is compared
with the accuracy of the conventional and widely used top-oil temperature models such
as IEC 60076-7 loading guide [8] described in section 2.1.1 as well as the linear model
described in 2.1.2.1. It is worthwhile to note that, the linear model is a simplified model
developed based on the IEEE C57.91 model whose accuracy is limited in most cases; how-
ever, due to its simple structure and its wide applications in the literature, a modified version
of (2.11) is used in this chapter in which two independent parameters and two regressors
are added to (2.11) as follows:

θto (i + 1) = k1 · θto (i) + k2 · θa (i) + k3 · K(i)2 + k4 · NP umps (i) + k5 · NF ans (i) + k6 , (5.1)

where k1 through k6 are the regressors. In the case of indoor distribution trans-
formers, one independent parameter and one regressor are added to (2.11) as fol-
lows:
θto (i + 1) = k1 · θto (i) + k2 · θa (i) + k3 · θs (i) + k4 · K(i)2 + k5 . (5.2)
56 5 Validation of the Dynamic Thermal Model

With these simple modifications, the accuracy of the linear model is significantly enhanced.
It is worthwhile to note that the performance of the linear model without modification is
not acceptable and will not be reported in this chapter. In the rest of this thesis, the linear
model is referred to the modified linear model.

In the rest of this chapter, the measured data during in-service operation of five power
transformers with different cooling systems and one indoor distribution transformer are
used for validation of the proposed model, the linear model (equation (5.1) or (5.2)), and
the IEC 60076-7 model [8]. The main characteristics of the transformers under study in
this thesis are listed in Table 5.1. The data sets of Tr1 through Tr6 are used in this chapter,
whereas the data set of Tr7 is used in section 6.2.3.

However, for the evaluation of the performance of the proposed thermal model in section
3, the unknown parameters of the model should be firstly determined. In other words,
the model should be parametrized. After parametrization of the model, it can be used for
calculation of the top-oil temperature.

5.1 Parametrization of the Thermal Model

The thermal model proposed in chapter 3 is an empirical thermal model utilizing design-
dependent and empirical factors which are normally not provided by the manufacturer of
the transformer. Therefore, they must be estimated using measured data during normal op-
eration of transformers. After the estimation of the unknown parameters of the model, they
can be used for calculation of the top-oil temperature; and the measured top-oil temperature
is used only for validation. It is a matter of fact that these estimated parameters should be
independent of the operating conditions of the transformer.

The parametrization of thermal models is an optimization problem aiming to determine

the unknown parameters φ of the thermal model f (X, φ) listed in Tables 3.1 and 3.2 with

Table 5.1: Characteristics of the transformers under study.

Transformer Tr1 Tr2 Tr3 Tr4 Tr5 Tr6 Tr7

Cooling System ODAF OFAF OFAN ONAF ONAN ONAN ONAN
Power in MVA 333 600 115 80 50 0.4 40
No-Load Loss in kW 47 300 67.52 10 18 0.43 14
Short Circuit Loss at 75°C in kW 510 1550 414.2 80 150 4.6 135
Total Mass in T 250 485 177.6 184 97.7 1.5 81.2
Mass of Oil in T 50 89 48.6 39 22.3 0.28 15.1
Number of Pumps 4 8 2 0 0 0 0
Number of Fans 8 8 0 6 0 0 0
5.1 Parametrization of the Thermal Model 57

the input data sets measured during normal operation of transformers denoted by X (see
Tables 3.1 and 3.2) in such a way that the sum of squares of the deviation of the calculated
top-oil temperature (using equation (3.2) or (3.49)) from the measured temperature Y is
minimized. This can be formulated as follows:

minimize kY − f (X, φ)k22 . (5.3)

As it can be seen from Fig. 5.1, for the estimation of the unknown parameters φ of the ther-
mal model f (X, φ), the input vector X consists of the measured ambient temperature, the
measured station temperature (if applicable), the load factor, the measured top-oil tempera-
ture, and number of running pumps and fans (if applicable), whereas the output of the model
Y consists of the measured top-oil temperature in the next time step.

Moreover, depending on the complexity of the model the parametrizing of the model re-
quires different number of measurements. Swift et al. in [13] used 480 measured values
corresponding to 24 hours for parametrizing of their proposed model. He et al. in [23]
used measurements corresponding to 5 days. Radakovic et al. in [29] used 1536 measured
corresponding to 16 days. Jauregui-Rivera et al. in [114] used 2880 measurement sets cor-
responding to 30 days. However, a big data set consisting of measurements during several
months does not lead necessarily to a significantly better accuracy in the calculation of the
top-oil temperature [115]. However, it has been shown in [116] that data sets consisting of
measurements during one month provides a good accuracy of the top-oil temperature cal-
culation. Moreover, as it can be seen from Fig. 5.2, data sets consisting of measurements

qa (i)
If applicable q s (i )

K (i ) Thermal Tto (i + 1) or Ttank (i + 1)

Model
If applicable N Pumps(i)
f ( X i ,f ) Y
i = 1,..., n
If applicable NFans(i) Output
Tto (i) or Ttank(i)

X
Inputs

Figure 5.1: Schematic of the parametrization.

58 5 Validation of the Dynamic Thermal Model

Figure 5.2: Effect of parametrization period on the accuracy of the thermal model [116].

more than one months do not lead to a significantly better accuracy. Moreover, the accu-
racy of the model parametrized using data set measured during one week and two weeks
is limited. Furthermore, a good data set should be a representative of different operating
conditions of transformers. As an example, if the load factor of the transformer during one
month is always equal to 30% and the data set of this month is used for parametrization of
thermal models, the accuracy of thermal models, specially semi-physical models, will be
limited when the load of the transformer is increased significantly. However, the models
developed based on physical phenomena (physical models) can easily extrapolate over the
range of parametrization data set. Hence, the range of the input parameters may not be a
crucial case for physical models.

Hence, in this thesis data set consisting of values measured during one month are used for
parametrization of the proposed model. Moreover, in order to evaluate the performance of
thermal models under different operating conditions, the data sets used for parametrization
of the model are not used for validation of the model. Table 5.2 shows the data sets used for
parametrization and validation of the proposed thermal models for the transformers listed
in Table 5.1.

In this thesis, the unknown parameters of the proposed models listed in Tables 3.1 and 3.2
are estimated using non-linear least squares method described in appendix A.1. This esti-
mated unknown parameters with the parametrization data sets (see Table 5.2) are provided
in Table 5.3.
5.1 Parametrization of the Thermal Model 59

Table 5.2: The months during which the data sets have been measured. The data sets are
measured during normal operation of transformers and are used for parametrization and
validation of the models for the transformers under study.

Parametrization data set Validation data set

Transformer measured during measured during
Tr1 (ODAF) June August, December
Tr2 (OFAF) July August
Tr3 (OFAN) April July
Tr4 (ONAF) December from January to May
Tr5 (ONAN) December from January to March
Tr6 (ONAN) March from June to August

Table 5.3: Unknown parameters of the proposed thermal models estimated using least-
squares method for the transformers under study.

Tr1 (ODAF) Tr2 (OFAF) Tr3 (OFAN) Tr4 (ONAF) Tr5 (ONAN) Tr6 (ONAN)
qoil n.a. 1.6 0.1 1.06 1.9 0.32
poil n.a. 0.71 0.33 0.3 0.4 1.40
qair n.a. n.a. 0.62 n.a. 3 1.09
pair n.a. n.a. 0.28 n.a. 0.08 0.26
soil 0.16 0.03 0.1 0.49 0.32 0.26
sair 2.86 3.07 2.91 2.67 2.21 6.12
Aoil 1.82 1.08 1.78 0.49 0.9 0.21
Aair 78.4 76.14 79.53 82.08 82.48 8.39
Arad 51.62 99.48 48.01 68.67 99.17 20.47
uoil × doil 0.23 0.32 0.13 0.48 n.a. n.a.
uair × dair 1.14 0.002 n.a. 0.48 n.a. n.a.
Coil 0.45 0.01 0.49 n.a. n.a. n.a.
m 0.88 0.002 0.46 n.a. n.a. n.a.
n 1.12 0 0.2 n.a. n.a. n.a.
Cair 0.53 0.56 n.a. 0.31 n.a. n.a.
m′ 0.58 0.86 n.a. 0.18 n.a. n.a.
n′ 0.34 0 n.a. 0.6 n.a. n.a.
gr 17.34 20.52 25 15.08 19.14 15
y 1.6 1.12 2 2 1.6 1.96
x′ 0.15 0.48 0.47 0.11 0.14 0.19
Aopenings n.a. n.a. n.a. n.a. n.a. 0.50
H n.a. n.a. n.a. n.a. n.a. 3.01
l/Acond n.a. n.a. n.a. n.a. n.a. 0.001
60 5 Validation of the Dynamic Thermal Model

5.2 ODAF Transformers

In ODAF transformers, as it can be seen from Fig. 3.1c, the oil entering the cooler is approx-
imately at maximum temperature since the oil is circulated by the pumps. Consequently,
compared with other cooling methods, the velocity of the oil in the winding of ODAF trans-
formers is higher. Therefore, the magnitude of the heat transfer coefficient increases inside
the winding and the heat flux per unit transfer area can be increased without exceeding the
average temperature rise permissible in the winding. But there are also economic and tech-
nical limits imposed on increasing the heat flux per unit transfer area.

The proposed model in section 3.1 is applied on Tr1 whose characteristics are listed in Ta-
ble 5.1. Tr1 is a single-phase auto-transformer with four coolers each of which consists
of one pump and two fans. It should be noted that Tr1 has 33% more cooling capacity
than it needs in order to have sufficient cooling capacity during overloading of the trans-
former.

In order to parametrize the proposed model, the measured data during in-service operation
of the transformer in June has been used. As formerly stated, the measured data include
the load factor, numbers of running pumps and fans, as well as the ambient and top-oil
temperatures. After parametrizing of the model, it can be used to calculate the top-oil
temperature. The estimated parameters of the proposed model are listed in Table 5.3. The
measured top-oil temperature is used for validation of the model.

5.2.1 Performance of the Model in High Loading Regimes

Investigation of the performance of thermal models under high loading and high ambient
temperature regimes is of crucial tasks. Fig. 5.3 shows the inputs of the model, namely the
load factor, the ambient temperature, numbers of running pumps and fans, and the measured
and calculated temperatures during August where the load of the transformer is near its
nameplate rating and the ambient temperature is also high between 12th and 19th of August.
As it can be seen, the calculated temperature is in good agreement with the measured top-
oil temperature. The mean absolute error and the maximum error in this case are 1.1 and
5.7 Kelvin respectively. Moreover, the performance of the improved model is not affected
by the intense increase of the load factor which is not the case in the conventional models
introduced in section 2.1.

Moreover, Fig. 5.4 shows the error duration curves of the improved model, the linear
model, and the IEC model. As it can be perceived, the error in the calculation of the top-
oil temperature using improved model is significantly lower than that of the conventional
5.2 ODAF Transformers 61

models. The performance of the IEC 60076-7 model is limited for ODAF transformers
since it does not account for the contribution of fans and pumps. On the contrary, the
modified linear model (equation (5.1)) which is named in the rest of this section as the linear
model shows a better performance compared with the IEC model. The mean absolute error
and the maximum error in the calculation of the top-oil temperature using each model for
this scenario are listed in Table 5.4.

100 Load Factor

80
/%

60
40
20
40 Ambient Temperature
30
/°C

20
10
Number of Running Pumps Number of Running Fans
8
6
/1

4
2
0
60 Measured Top−Oil Temperature Calculated Top−Oil Temperature
55
50
/°C

45
40
35
30
12−Aug 17−Aug 22−Aug 27−Aug 31−Aug

Figure 5.3: The measured and calculated top-oil temperature during in-service operation
of Tr1 where the load factor is near the nameplate load and the ambient temperature is
high. The model has been parametrized by data measured in June.

IEC 60076−7
Linear Model
8 Proposed Model
Sorted Absolute Error/K

0
0 10 20 30 40 50 60 70 80 90 100
Time/%

Figure 5.4: Duration curve of the error in the calculation of the top-oil temperature using
different models for Tr1 during August when the load factor is near the nameplate rating
and the ambient temperature is high.
62 5 Validation of the Dynamic Thermal Model

Table 5.4: Indices of error in the calculation of the temperature during August for Tr1 using
the IEC 60076-7, the linear model, and the proposed model.

Model Mean Absolute Error/K Maximum Error/K

IEC 60076-7 4.0 8.2
Linear Model 1.3 5.8
Proposed Model 1.1 5.7

5.2.2 Performance of the Model at Low Ambient Temperatures

As discussed in section 2.3, the accuracy of the conventional models are limited at low
ambient temperature [60] since they do not account for the viscosity of the oil. It is shown in
this scenario that the proposed model provides also an accurate performance at low ambient
temperatures.

Fig. 5.5 shows the input of the proposed thermal model as well as the measured and cal-
culated top-oil temperatures in December. As stated formerly, the data measured during
in-service operation of the transformer in June have been used for parametrizing the model.
As it can be seen, the proposed model can calculate the top-oil temperature accurately es-
pecially when the ambient temperature is very low (between 19th and 25th of December
when the ambient temperature falls below -10°C). Moreover, since the model has been
parametrized in June, it has a good long-term performance in the calculation of the top-
oil temperature. The mean absolute error and the maximum error in the calculation of the
top-oil temperature corresponding to Fig. 5.5 are 1.3 and 4.2 Kelvin respectively. Fig. 5.6
shows the duration curve of the error in the calculated top-oil temperature using different
thermal models. As it can be seen, the accuracy of the conventional models is limited when
the ambient temperature is low. The maximum error in the calculation of the top-oil tem-
perature using conventional models in this case, which is approximately 8 Kelvin, occurs
when the ambient temperature falls below -10 °C. It explicitly means that the conventional
models are not able to extrapolate beyond the training data since they consider the thermal
resistance of the transformer as a constant value. Therefore, they can not provide reliable
results under different operating conditions. The quantitative investigation of the error for
this scenario is listed in Table 5.5.
5.2 ODAF Transformers 63

80
60 Load Factor

/%
40
20
10 Ambient Temperature

/°C
0
−10
−20
6 Number of Running Pumps Number of Running Fans
4
/1

2
40
35 Measured Top Oil Temperature
Calculated Top Oil Temperature
30
25
/°C

20
15
10
5
01−Dec 07−Dec 13−Dec 19−Dec 25−Dec 30−Dec

Figure 5.5: The measured and calculated top-oil temperatures for Tr1 when the ambient
temperature is very low. The model has been parametrized by data measured in June.

IEC 60076−7
Linear Model
8
Proposed Model
Sorted Absolute Error/K

0
0 10 20 30 40 50 60 70 80 90 100
Time/%

Figure 5.6: Duration curve of the error in the calculation of the top-oil temperature using
different models for Tr1 during December when the ambient temperature is very low.
The model has been parametrized by data measured in June.

Table 5.5: Indices of error in the calculation of the temperature during December for Tr1
using IEC 60076-7, linear model and proposed model.

Model Mean Absolute Error/K Maximum Error/K

IEC 60076-7 2.3 8.2
Linear Model 1.9 8.9
Proposed Model 1.3 4.2
64 5 Validation of the Dynamic Thermal Model

5.3 OFAF Transformers

Due to the high oil velocity in the cooler of OF power transformers, a good oil-side heat
transfer coefficient is obtained. In such a cooling system, within the windings, the oil flows
practically in the natural way and the presence of the oil pumps affects mainly the radiator
performance. This method of cooling is characterized by the mixing of the oil heated up in
the winding and the cold oil streaming along the tank wall through the shunt path illustrated
in Fig. 3.1b. The temperature of the oil entering the cooler is lower than the maximum oil
temperature.

The proposed model in section 3.1 is applied on Tr2 whose characteristics are listed in
Table 5.1. Tr2 has 8 pumps and 8 fans in which all pumps are always in-service and the
fans are switched depending on their control algorithm.

Fig. 5.7 shows the inputs of the model as well as the measured and calculated top-oil
temperatures for Tr2 in August. The model is parametrized using data set measured during
in-service operation of the transformer in July. As it can be seen, the model can calculate
the top-oil temperature very well. The mean absolute error and the maximum error of the
calculation corresponding to Fig. 5.7 is 1.6 and 5.8 Kelvin respectively. Moreover, the
model represents a good performance for OFAF transformers when the load factor is near
the nameplate rating of the transformer between August the 27th and September the 1st when
the load factor of the transformer is about 90% at the ambient temperature of approximately
35 °C.

100
80 Load Factor
60
/%

40
20
0
40
30
/°C

20
10 Ambient Temperature

8
6
/1

4
2
Number of Running Pumps Number of Running Fans
55

50
/°C

45

40
Measured Top−Oil Temperature Calculated Top−Oil Temperature
02−Aug 07−Aug 12−Aug 17−Aug 22−Aug 27−Aug 31−Aug

Figure 5.7: The measured and calculated top-oil temperatures for Tr2. The model has been
parametrized by data measured in July.
5.3 OFAF Transformers 65

Fig. 5.8 shows the duration curve of the error in the calculation of the top-oil temperature
using proposed model and the conventional models during August. As it can be perceived,
the conventional models especially the IEC model has no acceptable performance when
it comes to OFAF transformers even in short-term investigations. However, the modified
linear model has the privilege to represent a good performance in this case compared with
the IEC model since it accounts for the number of running fans and pumps. The mean
absolute error and the maximum error of the temperature calculation using the proposed
model and the conventional models are listed in Table 5.6.

14
IEC 60076−7
Linear Model
12 Proposed Model

10
Sorted Absolute Error/K

0
0 10 20 30 40 50 60 70 80 90 100
Time/%

Figure 5.8: Duration curve of the error in the calculation of the top-oil temperature us-
ing different models for Tr2 during August. The model has been parametrized by data
measured in July.

Table 5.6: Indices of error in the calculation of the temperature during August for Tr2 using
the IEC 60076-7, the linear model and the proposed model.

Model Mean Absolute Error/K Maximum Error/K

IEC 60076-7 6.0 13.7
Linear Model 2.2 7.1
Proposed Model 1.6 5.8
66 5 Validation of the Dynamic Thermal Model

5.4 OFAN Transformers

With the same amount of heat to be dissipated, the temperature of the oil entering the
windings at the bottom in OFAN transformers is higher compared with OFAF transformers
since the temperature rise of the oil in the cooler at the middle height is higher than the
OFAF cooling system. But the temperature change of the oil in the windings remains
unchanged. Therefore, it is expected that OFAN transformers are operating at higher oil
temperature compared with OFAF transformers. Moreover, in this cooling mode there is no
way to force the oil temperature to remain under a certain value or a specified temperature
range which can be achieved in AF cooling systems by switching of the fans using a proper
switching algorithm.

The proposed model in section 3.1 is applied on Tr3 with OFAN cooling system whose
characteristics are listed in Table 5.1 and equipped with two pumps that are always in-
service.

Fig. 5.9 shows the inputs of the model, namely the ambient temperature, the load factor,
and number of running pumps as well as the measured and calculated temperatures in July
for Tr3. The model is parametrized using the data measured during in-service operation of
the transformer in April. As it can be seen, the oil temperature rise is approximately 35 °C
whereas in the case of OFAF transformers depicted formerly in Fig 5.7, the oil temperature
rise is approximately 15 °C. Moreover, the calculated temperature is in a good agreement
with the measured temperature where the mean absolute error and the maximum error in
the calculation of the temperature are 1.1 and 7.7 Kelvin respectively. It is worthwhile to
note that the maximum error and the mean absolute error are affected severely by the time
interval shown by dashed rectangular in Fig. 5.9 where there is no clear reason for sudden
decrease of the oil temperature. This may be due to external factors such as wind. A high
wind velocity with the proper direction can be considered as the forced convection and
can affect the heat transfer coefficient of the air-side resulting in the enhancement of the
radiator performance and decreasing of the oil temperature. However, since the calculated
temperature is higher than the measured temperature, this deviation may not lead to the
underestimation of the hot-spot temperature.

Fig. 5.10 shows the error duration curve of the proposed model as well as the conventional
model. The proposed model has more or less the same performance as the linear and IEC
Model. In the case of Tr3, since the pumps are always in operation and there is no fan to
be switched frequently, the continuous operation of pumps can be perceived as a constant
negative top-oil temperature rise over ambient temperature at rated load; in addition, the
loading condition of the transformer remains approximately constant over time. Hence, the
5.4 OFAN Transformers 67

linear and IEC models can represent a good performance in this case. The mean absolute
error and the maximum error of the temperature calculation using different models are listed
in Table 5.7.

100
80
60
/%

40 Load Factor
20
0
40
30
/°C

20
10
Ambient Temperature

Number of Running Pumps

/1

80
Measured Top−Oil Temperature Calculated Top−Oil Temperature
75
70
/°C

65
60
55
50
03−Jul 08−Jul 13−Jul 18−Jul 23−Jul 28−Jul

Figure 5.9: The measured and calculated top-oil temperatures for Tr3. The model has been
parametrized by data measured in April.

IEC 60076−7
Linear Model
8
Proposed Model
Sorted Absolute Error/K

0
0 10 20 30 40 50 60 70 80 90 100
Time/%

Figure 5.10: Duration curve of the error in the calculation of the top-oil temperature using
different models for Tr3 during July. The model has been parametrized by data measured
in April.
68 5 Validation of the Dynamic Thermal Model

Table 5.7: Indices of error in the calculation of the temperature during July for Tr3 using
the IEC 60076-7, the linear model and the proposed model.

Model Mean Absolute Error/K Maximum Error/K

IEC 60076-7 1.1 5.8
Linear Model 1.1 8.2
Proposed Model 1.1 7.7

5.5 ONAF Transformers

ONAF transformers are characterized by a higher heat transfer coefficient due to the oper-
ation of fans in comparison with ONAN transformers. However, the improvement of the
heat transfer coefficients are related to the air-side; but the oil-side heat transfer coefficient
is not affected significantly. Moreover, under the same oil temperature rise and winding
temperature gradient, ONAF transformers can serve higher loads compared with ONAN
transformers.

The proposed model in section 3.1 is applied on Tr4 with ONAF cooling system whose
characteristics are listed in Table 5.1 and equipped with two groups of fans. Fig. 5.11 shows
the inputs of the model, namely the ambient temperature, the load factor, and number of
running fans as well as the measured and calculated temperatures from the beginning of
January to the end of May for Tr4. The model is parametrized using data measured during
in-service operation of the transformer in December. The load factor of the transformer
is approximately 90% of the nameplate loading whereas the ambient temperature changes
in a wide range from -10 °C to 30 °C. As it can be seen, the calculated temperature is in
agreement with the measured temperature where the mean absolute error and the maximum
error is 0.7 and 5.8 Kelvin respectively.

Fig. 5.12 shows the error duration curve in the calculation of the top-oil temperature using
different models corresponding to the time interval of Fig. 5.11. As it can be seen, the
linear model has a large maximum error equal to 27 Kelvin. This error occurs where two
fans are switched off at the end of January. This large maximum error is due to the nature
of the linear model. In other words, the temperature changes in a transformer is non-linear;
and hence, the linear model is a non-physical model which is not reliable in all cases. At
the same time, the IEC model has also a limited accuracy with a maximum absolute error
of 8.8 Kelvin. However, the IEC and linear models present a good performance where no
fan is switched. In this case the continuous operation of fans is perceived by the models
as a constant negative oil temperature rise over ambient temperature. The constant loading
5.5 ONAF Transformers 69

condition may also be a reason for the good performance of the linear and IEC models. The
mean absolute error and the maximum error of the temperature calculation using different
models are listed in Table 5.8.

100
80
60
/%

40 Load Factor
20
0
30 Ambient Temperature
20
/°C

10
0
−10

6
/1

Number of Running Fans

0
70
65
60
55
50
/°C

45
40
35 Measured Top−Oil Temperature
30 Calculated Top−Oil Temperature
02−Jan 22−Jan 11−Feb 03−Mar 23−Mar 12−Apr 02−May 22−May

Figure 5.11: The measured and calculated top-oil temperatures for Tr4. The model has
been parametrized by data measured in December.

28
IEC 60076−7
26 Linear Model
24 Proposed Model

22
20
Sorted Absolute Error/K

18
16
14
12
10
8
6
4
2
0
0 10 20 30 40 50 60 70 80 90 100
Time/%

Figure 5.12: Duration curve of the error in the calculation of the top-oil temperature using
different models for Tr4 from January to May. The model has been parametrized by data
measured in December.
70 5 Validation of the Dynamic Thermal Model

Table 5.8: Indices of error in the calculation of the temperature from January to May for
Tr4 using the IEC 60076-7, the linear model and the proposed model.

Model Mean Absolute Error/K Maximum Error/K

IEC 60076-7 1.0 8.8
Linear Model 0.8 27.0
Proposed Model 0.7 5.8

5.6 ONAN Transformers

In the closed cooling circuit of transformers with ON cooling system, the oil is kept in
circulation by gravitational buoyancy. The heat developing in active parts transfers into the
surrounding oil. The oil warms up, its specific gravity decreases and flows upward, while
the ascending oil is replaced by the incoming cold oil. The hot oil dissipates its heat along
the colder walls of the coolers, its specific gravity increases and it flows downwards. In the
coolers a flow of downward direction is produced which is opposed to the prevailing flow
along the active parts [75].

The proposed model in section 3.1 is applied on Tr5 with ONAN cooling system whose
characteristics are listed in Table 5.1. Tr5 is a wind farm transformer which experiences
severe load changes and overload conditions during winter.

Fig. 5.13 shows the inputs of the model, namely the load factor, the ambient temperature
as well as the measured and calculated temperatures from the end of January to the end
of March for Tr5. The model is parametrized using the data measured during in-service
operation of the transformer in December. The transformer experiences a volatile loading
condition which changes suddenly from very low loading to the overloading regime. The
recorded load of the transformer in this time period varies from 0% to 121%. Moreover,
the ambient temperature varies in a wide range from -10 °C to 20 °C. As it can be seen,
the proposed model can calculate the top-oil temperature accurately where the mean abso-
lute error and maximum error of temperature calculation are 1.1 and 6.9 Kelvin. Fig 5.14
shows the duration curve of the error in the calculation of the top-oil temperature using
different models. As it can be seen, the accuracy of the proposed model is approximately
two times better than the conventional models. Moreover, in the case of the ONAN trans-
formers, the linear and IEC models represent the same performance for calculation of the
temperature.

For comparing the models quantitatively, the mean absolute error and the maximum ab-
solute error in the temperature calculation using different models are listed in Table
5.9.
5.7 ONAN Indoor Distribution Transformers 71

120 Load Factor

100
80

/%
60
40
20
0
20 Ambient Temperature
10
/°C

0
−10
50
45 Measured Top−Oil Temperature
40 Calculated Top−Oil Temperature
35
/°C

30
25
20
15
10
5
0
22−Jan 01−Feb 11−Feb 21−Feb 03−Mar 13−Mar 23−Mar

Figure 5.13: The measured and calculated top-oil temperatures for Tr5. The model has
been parametrized by data measured in December.

IEC 60076−7
Linear Model
8
Proposed Model
Sorted Absolute Error/K

0
0 10 20 30 40 50 60 70 80 90 100
Time/%

Figure 5.14: Duration curve of the error in the calculation of the top-oil temperature using
different models for Tr5 from end of January to end of March. The model has been
parametrized by data measured in December.

5.7 ONAN Indoor Distribution Transformers

According to the IEC 60076-7 loading guide [8], in case of the indoor transformers, the
temperature of the enclosure should be used instead of the ambient temperature. As an
alternative, correction can be added to the ambient temperature. As an example, for a
72 5 Validation of the Dynamic Thermal Model

Table 5.9: Indices of error in the calculation of the temperature from January to March for
Tr5 using the IEC 60076-7, the linear model and the proposed model.

Model Mean Absolute Error/K Maximum Error/K

IEC 60076-7 2.2 8.5
Linear Model 2.0 8.5
Proposed Model 1.1 6.9

building with natural ventilation and one installed 500 kVA transformer, 8 Kelvin should
be added to the ambient temperature. However, it is expressively mentioned in the IEC
loading guide [8] that "the temperature corrections have been estimated for typical substa-
tion loading conditions using representative values of transformer losses. They are based
on the results of a series of natural and forced cooling tests in underground vaults and sub-
stations and on random measurements in substations and kiosks." However, it is shown in
the rest of this section that the consideration of the enclosure ventilation has a significant
impact on the accuracy of the thermal models.

The proposed model in section 3.2 is applied on Tr6 which is an indoor distribution
transformer with ONAN cooling system whose characteristics are listed in Table 5.1.
This transformer serves domestic loads; and photovoltaic systems are also connected to
it.

Fig. 5.15 shows the measured ambient and station temperatures, the load factor, and the top-
oil temperature, in addition to the calculated top-oil temperature using (3.49) in June, July,
and August using the unknown parameters estimated by use of measured data in March dur-
ing in-service operation. It is worthwhile to note that the top-oil temperature is calculated
using calculation of the tank temperature and the model is parametrized using the measured
tank temperature. As it can be seen, the calculated top-oil temperature is in a good agree-
ment with the measured top-oil temperature where the mean absolute error and the maxi-
mum error are 0.3 and 3.1 Kelvin respectively. The maximum error of the temperature cal-
culation occurred where there was a measurement gap in the calculation process due to the
maintenance of the monitoring system leading to increasing of the Dt in (3.34) and, conse-
quently, a significant calculation error. Due to the good performance of the model, it is even
applicable for the transformers which do not have temperature pocket as well as for trans-
formers whose temperature pocket is used for other applications. Moreover, measurement
of the tank temperature is easier than the measurement of the top-oil temperature and it is
not necessary to remove the transformer from service. According to the IEC loading guide
60076-7 [8], the temperature sensor for the measurement of the tank temperature should be
installed at the top of the tank or near the temperature pocket.
5.7 ONAN Indoor Distribution Transformers 73

40
Load
30

/%
20
10
0
Ambient Temperature Station Temperature
40
30
/°C

20
10
50
Measured Top−Oil Temperaute Calculated Top−Oil Temperature
45

40
/°C

35

30

25
09−Jun 29−Jun 19−Jul 08−Aug

Figure 5.15: The measured and calculated top-oil temperatures for Tr6. The top-oil tem-
perature has been calculated using the calculation of the tank temperature. The model
has been parametrized by data measured in March.

Hence, the tank temperature can be a good representative of the top-oil temperature. De-
pending on the desired accuracy, it can be directly used instead of the top-oil temperature, or
to achieve the high accuracy, the top-oil temperature can be calculated based on the model
parametrized using the tank temperature.

On the contrary, the conventional models do not provide the ability to calculate the top-
oil temperature using the tank temperature. In order to investigate the accuracy of the
conventional models, instead of the tank temperature, the top-oil temperature is used to
parametrize the conventional models. Fig. 5.16 shows the duration curve of the error in the
calculation of the top-oil temperature using different models. As it can be seen, although
the top-oil temperature is calculated using the tank temperature by the proposed model, it
represents a significantly better performance compared with conventional models. It can
be perceived that the modifications applied to the linear model in (5.2) lead to enhancing
the accuracy of the linear model since it provides better accuracy compared with the IEC
model. However, due to the non-linear nature of temperature change, the accuracy of the
linear model (even after modification) is still limited for the indoor distribution transform-
ers.

For comparing the models quantitatively, the mean absolute error and the maxi-
mum error in the temperature calculation using different models are listed in Table
5.10.
74 5 Validation of the Dynamic Thermal Model

IEC 60076−7
Linear Model
6 The Proposed Model

Sorted Absolute Error/K

4

0
0 10 20 30 40 50 60 70 80 90 100
Time/%

Figure 5.16: Duration curve of the error in the calculation of the top-oil temperature using
different models for Tr6 in June, July, and August. The model has been parametrized by
data measured in March.

Table 5.10: Indices of error in the calculation of the temperature from June to August for
Tr6 using the IEC 60076-7, the linear model and the proposed model.

Model Mean Absolute Error/K Maximum Error/K

IEC 60076-7 3.7 7.1
Linear Model 0.8 5.5
Proposed Model 0.3 3.1

5.8 Summary of the Validation

It is worthwhile to summarize the main points of the validation of the proposed model
and the conventional models. The quantitative accuracy evaluations for different studied
cases in this section is listed in Table 5.11 from which the following items can be per-
ceived:

• The accuracy of the proposed model is always better than the conventional models.

• The cooling method of air-cooled transformers does not significantly affect the per-
formance of the proposed model which is not the case in the conventional models.
5.8 Summary of the Validation 75

Table 5.11: Indices of error in the calculation of the temperature using the IEC 60076-7,
the linear model and the proposed model.

Mean Absolute Error/K Maximum Error/K

Cooling Corresponding IEC Linear Proposed IEC Linear Proposed
Transformer System to 60076-7 Model Model 60076-7 Model Model
Tr1 ODAF Fig. 5.4 4.0 1.3 1.1 8.2 5.8 5.7
Tr1 ODAF Fig. 5.6 2.3 1.9 1.3 8.2 8.5 4.2
Tr2 OFAF Fig. 5.8 6.0 2.2 1.6 13.7 7.1 5.8
Tr3 OFAN Fig. 5.10 1.1 1.1 1.1 5.8 8.2 7.7
Tr4 ONAF Fig. 5.12 1.0 0.8 0.7 8.8 27.0 5.8
Tr5 ONAN Fig. 5.14 2.2 2.0 1.1 8.5 8.5 6.9
Tr6 ONAN Fig. 5.16 3.7 0.8 0.3 7.1 5.5 3.1

• The modified linear model provides a better performance compared with the IEC
loading guide since the loading guide does not directly account for running pumps
and fans.

• The accuracy of the conventional models are limited in most cases since they consider
the thermal resistance of transformers as a constant value which is not affected by
temperature and operating condition of transformers.

• In case of indoor transformers, modeling of the ventilation method in the enclosure

plays a crucial role in the calculation of the oil temperature.

• The measurement of the tank temperature contains every information that the mea-
surement of the top-oil temperature can provide.
 77

6 Monitoring of the Cooling System

"In every inference there is a proposition that serves as a ground, and another,
namely the conclusion, that is drawn from the former, and finally the inference
(consequence) according to which the truth of the conclusion is connected un-
failingly with the truth of the first proposition"

The Critique of Pure Reason

E. Kant

I
N transformers with air-forced cooling system, due to the high air velocity, the air-side
heat transfer coefficient is increased; thus, compared to the natural air cooling, a smaller
temperature rise on the air-side will suffice for transferring an equal quantity of heat [75].
In other words, in the case of failure of a group of fans, the air-side heat transfer coefficient
decreases which causes an excessive oil temperature rise and a consequent reduction of the
loading capability of the transformer [27]. Therefore, without failure detection of fans, the
temperatures of the oil and the winding may exceed the allowable temperatures under a
specific ambient temperature resulting in lifetime reduction and, in some cases, production
of bubbles and gases in the oil [22]. Hence, the performance of the cooling fans plays
an important role to keep the temperatures of the oil and the windings under a certain
value.

In order to detect the malfunction of the cooling system in power transformers, two different
online algorithms are presented in this chapter which can be applied on every transform-
ers. Both algorithms use an accurate model for calculating the top-oil temperature which
has been proposed in section 3.1. The first algorithm benefits from the deviation of the
calculated top-oil temperature from the measured top-oil temperature. The second algo-
rithm is based on the detection of changes in the unknown parameters of the model listed
in Table 3.1 which are supposed to be constant or slightly variable in time. Both algorithms
are validated through two case studies using data measured during in-service operation of
Tr1 and Tr2, whose characteristics are listed in Table 5.1. Both transformers have been
equipped with monitoring systems, which measure the load factor, the ambient temper-
ature, the top-oil temperature, and numbers of running pumps and fans during operation
of the transformer. These measured values are used for estimating the design-dependent
78 6 Monitoring of the Cooling System

and the empirical factors (listed in Table 3.1) of the thermal model using non-linear least
squares method described in appendix A.1 and for recognizing healthy or faulty operations
of fans and pumps.

Both transformers experienced failures in part of their cooling systems due to a failure of
power supply of the fans which were not in accordance with the information supplied by the
control systems. In the case of Tr1, the control system aimed to switch on two additional
fans; however, due to the failure in the power supply of the fans, they did not start to run.
The site check revealed later that two fans failed to operate while the pump related to the
failed fans kept in operation. According to IEEE C57.91 [9] Annex H, the failure of two
fans with the pump still in operation is not a severe failure for ODAF transformer. In the
case of Tr2 which is an OFAF transformer, the control system of the fans switched on two
additional fans, in order to control the temperature under the specified loading condition.
Nevertheless, due to the failure in energy supply, which was later revealed by site check,
three out of six fans failed to operate which was not in accordance with the data provided
by the monitoring system. This is a severe failure for an OFAF transformer [9] that leads to
an oil temperature increase.

6.1 Monitoring by the Error in the Calculation of the

Top-Oil Temperature

The error in the calculation of the top-oil temperature can be used for detection of failures in
the operation of the cooling system. If an accurate model is used for calculation of the top-
oil temperature, the deviation of the calculated temperature from the measured temperature
will be small. In case of failure in the cooling system, this deviation increases. Moreover,
the severity of the failure affects the deviation of the calculated top-oil temperature from
the measured top-oil temperature.

Fig. 6.1 shows the measured load factor, the ambient temperature, numbers of running
pumps and fans, the measured top-oil temperature, as well as the calculated top-oil temper-
ature during normal and faulty operations of Tr1. As it can be seen, before the occurrence
of the failure on December 21st at 6:00, the calculated temperature is in a good agreement
with the measured temperature. But after the failure occurrence, the calculated tempera-
ture starts to deviate rapidly from the measured temperature indicating that a failure has
occurred. In the same way, Fig. 6.2 shows the measured and calculated top-oil temperature
during faulty and healthy operations of the cooling system for Tr2. The failure in cooling
6.1 Monitoring by the Error in the Calculation of the Top-Oil Temperature 79

system occurred on 21st of January and cleared on 25th of February. As it can be seen, al-
though the ambient temperature decreases and the average load over time remains constant,
the temperature of the oil increases rapidly after failure of three fans. Moreover, before the
occurrence and after the clearance of the failure, the calculated top-oil temperature is in a
good agreement with the measured temperature; nevertheless, during faulty operation of the
fans, the calculated temperature deviates from the measured temperature. The early detec-
tion of these failures might be possible with an online algorithm. This deviation should be
detected by a model-based monitoring system in such a way that the algorithm is applicable
on every transformer.

It is worthwhile to note that the thermal model should be firstly parametrized using the
measured values obtained during in-service operation of the transformer under study. The
parametrized model for Tr1 and Tr2 in section 5 are used in this section. The data set
measured during in-service operations in June and July are used to parametrize the thermal
model for Tr1 and Tr2 respectively.

6.1.1 An Algorithm for Analyzing the Error for Detection of Failures

in the Cooling System

The proposed algorithm for malfunction detection of the cooling system in power trans-
formers using online thermal monitoring is based on the determination of the error in the

60
/%

30
Load Factor
20
Ambient Temperature
10
/°C

0
−10
Number of Running Pumps
Actual Number of Fans
6
Recorded Number of Fans 21st Dec; 6:00
/1

4
2
40
30
20
/°C

10
Measured Top−Oil Temperature
0
Calculated Top−Oil Temperature
−10
01−Dec 07−Dec 13−Dec 19−Dec 25−Dec 31−Dec

Figure 6.1: The measured and calculated top-oil temperatures for Tr1 before and after
occurrence of the failure.
80 6 Monitoring of the Cooling System

60

/%
30
Load Factor
20 Ambient Temperature

/°C
10
0

8
6
/1

4 21−Jan; 14:00 25−Feb;10:00

2 Number of Pumps Recorded Number of Fans Actual Number of Fans

40

30
/°C

20

10 Measured Top Oil Temperature Calculated Top Oil Temperature

20−Jan 03−Feb 16−Feb 02−Mar 15−Mar

Figure 6.2: The measured and calculated top-oil temperatures for Tr2 before and after
occurrence of the failure.

calculated top-oil temperature. If the top-oil temperature is calculated accurately, this error
will be low during normal operation of the transformer. If a failure arises, the deviation
of the calculated temperature from the measured temperature will increase. This deviation
should be detected by the online monitoring system. In this section, the error in the calcu-
lated top-oil temperature is standardized and mapped into the interval of (−2, 2). Thereby,
the malfunction of cooling system can be detected and the alarm signal can be triggered
easily.

For online monitoring purposes, the recursive error (w) in the calculation of the temperature
is determined as follows [117]:

Tto,i − θto,i
wi = q , (6.1)
1 + f ′ i · (F ′ i−1 · Fi−1 )−1 · fi
i = p + 1, ..., m, (6.2)

where Tto is the measured top-oil temperature, θto is the calculated top-oil temperature us-
ing (3.1), f is the Jacobian vector corresponding to the measurement in the current step (i),
F is the Jacobian matrix corresponding to all of the previous steps, m is the number of ob-
servations, and p is the number of empirical or design-dependent parameters in the thermal
model. The inverse of the information matrix (F ′ i−1 · Fi−1 )−1 can be calculated recursively
6.1 Monitoring by the Error in the Calculation of the Top-Oil Temperature 81

for online purposes in an economical manner as follows [117]:

−1 −1 (F ′ i−1 · Fi−1 )−1 · fi · f ′ i · (F ′ i−1 · Fi−1 )−1

(F ′ i · Fi ) = (F ′ i−1 · Fi−1 ) − . (6.3)
1 + f ′ i · (F ′ i−1 · Fi−1 )−1 · fi

The recursive error w is a normally distributed random variable with zero mean and standard
variance w ∼ N(0, σ 2 ); therefore, the plot of the recursive error against time is expected to
be as a random scattering of the points above and below the line w = 0 being within the
band (−2σ, 2σ) corresponding to the 95% confidence interval [118] which is shown in Fig.
6.3a.

It has been shown in [119] that if the w in the calculated temperature

q is distributed normally
|w |
with zero mean and standard deviation w ∼ N(0, Iσ 2 ), σi
i
is distributed normally with
mean of 0.82218 and deviation of 0.34914. Therefore, the standardized recursive error
is distributed normally with zero mean and unit deviation (Z ∼ N(0, 1)) calculated as
follows:
r
wi
σi
− 0.82218
Zi = , (6.4)
0.34914
σi = σi−1 + wi 2 , (6.5)

where σ is the standard deviation of the error. The acceptable pattern of Z against
time is shown in Fig. 6.3b. This pattern can be used to detect the significance of the
deviation of each single point from normal distribution with zero mean and unit vari-
ance.

However, exceeding this band does not mean a catastrophic failure and the unnecessary
triggering of the alarm should also be avoided. As an example, the unnecessary alarm signal
may be triggered when external factors (e.g. sun intensity or wind speed) are not considered

w Z

+2σ +2

Time Time

-2σ -2
(a) (b)

Figure 6.3: Allowable band for recursive error (a) defined by (6.1), (b) defined by (6.4).
82 6 Monitoring of the Cooling System

in the thermal model which lead to a slight increase or decrease in the temperature rise of the
oil. Therefore, an appropriate procedure is needed to properly trigger the alarm signal. In
this paper, if the error exceeds the allowable band −2 ≤ Zi ≤ 2 for a series of consecutive
times, the cumulative sum of the deviation from the band is calculated. The advantage of
using cumulative sum is to increase the sensitivity of the method by combining the evidence
of the current standardized error with that of previous errors. Such a procedure enables us
to consider, not only the values of the last error but also the value of the last n errors for any
n [119]. In the case of a failure, the cumulative sum starts to increase for the upper value
of the band and to decrease for the lower value of the band. The mathematically equivalent
scheme of this procedure is as follows:

γ0 + = γ0 − = 0, (6.6)
γi + = max(0, γi−1
+
+ Zi − 2), (6.7)
−
γi − = min(0, γi−1 + Zi + 2), (6.8)

where γi + and γi − are the cumulative sum of the w in the case of exceeding and falling be-
low the allowable band respectively. The alarm signal is triggered if:

γi + > 2 ∨ γi − < −2. (6.9)

It should be noted that the upper and lower values of the band are chosen accord-
ing to the 95% confidence interval. This formulation leads to an easy automation of
the malfunction detection of the cooling system as a part of an online monitoring sys-
tem.

The flowchart of the proposed method for monitoring of the cooling system is illustrated
in Fig. 6.4. Firstly, the top-oil temperature should be calculated. Afterwards, the error in
the calculated top-oil temperature is calculated followed by the calculation of the recursive
error leading to a normal distribution with zero mean and standard variance. Thereafter, the
distribution of error is mapped into the normal distribution with zero mean and unit vari-
ance. Finally, the criteria for triggering the alarm signal are checked.

6.1.2 Validation of the Proposed Algorithm

Through four different scenarios the following points are shown and validated in this sec-
tion:
6.1 Monitoring by the Error in the Calculation of the Top-Oil Temperature 83

Start

g 0+ = g 0- = 0
i = p +1

Calculation of the top-oil

temperature

Calculation of the error between the

measured and the calculated top-oil
temperature

i=i+1 i=i+1
(Next Calculation of w
(Next
Measurement) Measurement)

wi
- 0.82218
si
Zi =
0.34914

g i + = max(0, g i+-1 + Z i - 2)
g i - = min( 0, g i--1 + Z i + 2)

g i+ > 2
Alarm=0 No g i - < -2 Yes Alarm=1

Figure 6.4: Flowchart of the proposed method for malfunction detection of the cooling
system.

• The algorithm recognizes the healthy operation of the cooling system. In other words,
it will not trigger false alarm signal.

• The algorithm detects failures in fan and pump operations.

• The performance of the algorithm is investigated through two case studies.

• The loading guides are not suitable for this algorithm since their accuracy is limited.
84 6 Monitoring of the Cooling System

6.1.2.1 Healthy Operation of the Cooling System

Provided that an accurate thermal model is used in the online algorithm for detection of
failures in the operation of the cooling system, the normal operation of the transformer
will not be led to triggering the alarm signal even in the case of the sudden changes in
the operation conditions of the transformer. The proposed algorithm is applied on Tr4
with ONAF cooling system whose characteristics is listed in Table 5.1. The measured and
calculated temperatures using the proposed thermal model shown in Fig. 6.5. As it can
be seen, in the time period between January 22nd and February 11th , the load factor of the
the transformer decreases rapidly for a short period of time and the number of running
fans is reduced by the control system of the fans. Without an accurate thermal model,
the calculated top-oil temperature will deviate from the measured temperature in this time
period which is the case in the linear model as well as the IEC model (see Fig. 5.12).
Consequently, this leads to an increase of the standardized error in the calculation of top-
oil temperature (Zi ) and a trigger of the alarm signal which is not the case when using an
accurate thermal model. Fig. 6.6 shows the calculated standardized error in the time period
corresponding to Fig. 6.5. As it can be seen, the proposed algorithm recognizes the healthy
operation of the fans for Tr4 even in a long period of time.

100
80
60
/%

40 Load Factor
20
0
30 Ambient Temperature
20
/°C

10
0
−10
6
/1

Number of Running Fans

0
70
65
60
55
50
/°C

45
40
35 Measured Top−Oil Temperature
30 Calculated Top−Oil Temperature
02−Jan 22−Jan 11−Feb 03−Mar 23−Mar 12−Apr 02−May 22−May

Figure 6.5: The measured and calculated top-oil temperatures from January to May for Tr4
when the operation of the cooling system is healthy. The model has been parametrized
in December.
6.1 Monitoring by the Error in the Calculation of the Top-Oil Temperature 85

Standardized Error (Z)

2

0
/1
-1

-2

02-Jan 22-Jan 11-Feb 03-Mar 23-Mar 12-Apr 02-May 22-May

Figure 6.6: Standardized error Zi for Tr4 from January to May when the cooling system is
operated properly.

6.1.2.2 Detection of Fan and Pump Failures

The performance of the algorithm in the malfunction detection of pumps and fans is in-
vestigated by applying artificial failures (manipulation of the input data set) through three
different scenarios as follows:

• Reduction of the cooling capacity due to failure of one pump.

• Reduction of the cooling capacity due to failure of one Fans.
• Reduction of the cooling capacity due to pollution or contamination leading to reduc-
tion of the fan speed.

Failure of one pump: In this scenario, an artificial failure is applied on the input data set
of Tr1 measured during August. It is assumed that on 22nd of August at 10:15 o’clock
one pump fails which is corresponding to 25% of the cooling capacity provided by pumps,
whereas all of the cooling capacity provided by fans are available. The top-oil temperature
is calculated using the thermal model in section 3.1 with unknown parameters listed in
Table 5.3. After this artificial failures, as it can be seen from Fig. 6.7, the calculated top-
oil temperature starts to deviate from the measured top-oil temperature. Consequently, the
standardized error exceeds the allowable band prescribed by (6.9) leading to triggering of
the alarm signal on 22nd of August at 14:15 o’clock. Hence, the algorithm is able to detect
the malfunction of pumps.

Failure of one fan: In this scenario, an artificial failure is applied on the input data set
of Tr4 (ONAF) measured during April. It is assumed that one fan fails to operate which
is corresponding to failure of 16.7% of the total cooling capacity. The top-oil tempera-
ture is calculated using the thermal model in section 3.1 with unknown parameters listed
in Table 5.3. After this artificial failures, as it can be seen from Fig. 6.8, the calculated
top-oil temperature deviates significantly from the measured top-oil temperature. Conse-
quently, the standardized error exceeds the allowable band prescribed by (6.9) leading to
86 6 Monitoring of the Cooling System

100 Load Factor

/%
50
Ambient Temperature
30

/°C
20
10
Actual Number of Pumps
Recorded Number of Pumps 22−Aug; 10:15
/1 Actual Number of Fans
8
4
2 3
Meausred Top−Oil Temperature Calculated Top−Oil Temperature
60
/°C

50
40
30
22−Aug; 14:15
2
/1

0
−2 Standardized Error
Alarm Signal
15−Aug 19−Aug 23−Aug 27−Aug 31−Aug

Figure 6.7: The measured and calculated top-oil temperatures as well as the standardized
error Zi for Tr1 when one pump fails to operate.

100
80
/%

60 Load Factor
40 Ambient Temperature
/°C

20
0
6
/1

5
Actual Number of Fans Recorded Number of Fans
65
60
/°C

55
50
45 Measured Top−Oil Temperature Calculated Top−Oil Temperature
Standardized Error
Alarm Signal
2
/1

0
−2

02−Apr 07−Apr 12−Apr 17−Apr 22−Apr 27−Apr

Figure 6.8: The measured and calculated top-oil temperatures as well as the standardized
error Zi for Tr4 when one fan fails to operate.

triggering of the alarm signal. Hence, the algorithm is able to detect the malfunction of
fans.

Reduction of the effective cooling capacity due to pollution or contamination: In this sce-
nario, an artificial failure is applied on the input data set of Tr4 (ONAF) measured during
April. It is assumed that the effective cooling capacity degrades gradually during a few days
and reaches finally 90% of the installed cooling capacity and remains constant for several
days (see third subfigure of Fig. 6.9). The top-oil temperature is calculated using the ther-
6.1 Monitoring by the Error in the Calculation of the Top-Oil Temperature 87

100

/%
75 Load Factor
50
30 Ambient Temperature

/°C
20
10
0
100
/% 95 97%
90
Actual Cooling Capacity Recorded Cooling Capacity
70 Measured Top−Oil Temperature Calculated Top−Oil Temperature
65
/°C

60
55
50

2
/1

0
−2 Standardized Error
Alarm Signal
02−Apr 07−Apr 12−Apr 17−Apr 22−Apr 27−Apr

Figure 6.9: The measured and calculated top-oil temperatures as well as the standardized
error Zi for Tr4 when the effective cooling system is degraded gradually.

mal model in section 3.1 with unknown parameters listed in Table 5.3. As it can be seen
from Fig. 6.9, at the beginning, the gradual degradation does not lead to rapid deviation
of the calculated top-oil temperature from the measured top-oil temperature since it has no
significant impact on the thermal behavior of the transformer. However, the gradual degra-
dation affects the thermal behavior when the actual cooling capacity falls below 97% of the
total cooling capacity leading to triggering of the alarm signal indicating that a malfunction
has been occurred although it is not a severe malfunction. Afterwards, due to the intense
rising of the ambient temperature, the deviation of the calculated top-oil temperature from
the measured top-oil temperature decreases and the standardized error returns back to the
allowable band. When the actual cooling capacity falls bellow approximately 95% of the
installed cooling capacity, the deviation between the calculated and the measured top-oil
temperature increases significantly and the standardized error exceeds the allowable band;
and consequently, the alarm signal is triggered again. After the clearance of the failure,
the deviation of the calculated top-oil temperature from the measured top-oil temperature
is negligible. Hence, the algorithm is able to detect the reduction of the cooling capacity
due to contamination or pollution.

6.1.2.3 Two Case Studies

The control system of the cooling system of Tr1 aimed to switch on two additional fans on
December 21st at 6:00; however, due to the failure in the power supply of the fans, they did
not start to run. The site check revealed later that two fans failed to operate while the pump
related to the failed fans kept in operation.
88 6 Monitoring of the Cooling System

Fig. 6.10 shows the standardized error Zi for Tr1 in December calculated using (6.4) at each
time based on the information of previous time, the number of running fans in December
according to the monitoring system, and the alarm signal triggered using the proposed
algorithm. As it can be seen, before the occurrence of the failure, Zi is in accordance
with the pattern depicted in Fig. 6.3. On the contrary, after failure of the two fans, the
calculated temperature starts to deviate from the measured temperature indicating that a
failure occurred in the cooling system. The standardized error Zi exceeds the maximum
allowable value and the criteria prescribed by (6.9) leads to triggering the alarm signal on
December 21st at 9:00. In other words, the failure in the cooling system is detected three
hours (three steps) after the occurrence.

For Tr2 which is an OFAF transformer, the control system of the fans switched on two addi-
tional fans, in order to control the temperature under specified loading condition on January
21st at 14:00. Nevertheless, due to the failure in energy supply, which was later revealed by
site check, three out of six fans failed to operate. Fig. 6.11 illustrates the standardized error
Zi , the number of the running fans according to the monitoring system, and the alarm signal
for Tr2 from middle of January to middle of March. Before the failure, the standardized
error Zi lies in the allowable band −2 ≤ Zi ≤ 2; however, after the switching of the fans
on January 21st at 14:00, the error in the calculation of the top-oil temperature increases
resulting in the exceeding of the standardized error Zi from the allowable band indicating
that a failure has occurred in the fan operation. As it can be seen in Fig. 6.11, the failure of
fans is detected one hour (one step) after the occurrence.

Standardized Error (Z )
i
6 Alarm Signal
Number of Running Fans

21−Dec; 6:00
4
/1

0
21−Dec; 9:00

−2

05−Dec 10−Dec 15−Dec 20−Dec 25−Dec 30−Dec

Figure 6.10: Standardized error Zi and alarm signal triggered using the proposed algorithm
with the proposed thermal model for Tr1 when two fans fail to operate due to energiza-
tion failure.
6.2 Changes of the Estimated Parameters for Detection of Fan Failures 89

7
21−Jan; 14:00
6

3
/1

2
21−Jan; 15:00
1
Standardized Error (Z )
0 i
Alarm Signal
−1 Number of Running Fans

−2

14−Jan 24−Jan 03−Feb 13−Feb 23−Feb

Figure 6.11: Standardized error Zi and alarm signal triggered using the proposed algo-
rithm with the proposed thermal model for Tr2 when three fans fail to operate due to
energization failure.

6.1.2.4 Incapability of the Loading Guide

Finally, it should be shown that the conventional models such as IEC 60076-7 model have
some limitations for malfunction detection of the cooling system. The limitation of the con-
ventional models roots back to the limited accuracy in the calculation of the top-oil temper-
ature. Fig. 6.12 illustrates the standardized error Zi calculated according to the proposed
algorithm in Fig. 6.4 using IEC 60076-7 model for Tr1 in the same operating situation as
Fig. 6.10. As it can be seen, the standardized error lies in allowable band after December
21st at 6:00 and no alarm signal has been triggered. Because before failure occurrence,
the big deviation between measured and calculated temperatures leads to recognition of the
error after failure occurrence as normal error, and consequently the standardized error lies
in allowable band. Moreover, a false alarm signal has been triggered when no failure has
occurred. Therefore, an accurate thermal model is a necessary part of an online monitoring
system.

6.2 Changes of the Estimated Parameters for Detection of

Fan Failures

Although the previous proposed algorithm can rapidly detect failure of the cooling sys-
tem, the algorithm is not applicable during parametrization phase. As it is stated in sec-
90 6 Monitoring of the Cooling System

6 Standardized Error (Z ) using IEC 60076−7

i
Alarm Signal using IEC 60076−7
Number of Running Fans 21−Dec; 6:00

/1 4

−2

05−Dec 10−Dec 15−Dec 20−Dec 25−Dec 30−Dec

Figure 6.12: Standardized error Zi and alarm signal triggered using the proposed algorithm
and the IEC 60076-7 model for Tr1.

tion 5.1, to parametrize the thermal model, a data set consisting of one month of mea-
sured data should be available. Therefore, it is worthwhile to provide an algorithm for
the failure detection of cooling system which is able to be used during parametrization
phase.

In the proposed model in section 3.1, there are some parameters whose values vary slightly
and in some cases do not change with the operating condition of the transformer. However,
in case of failures, the gradient of the changes in the estimated parameters (second term
in the right side of (A.4)) will be high since the error in the calculation of the temperature
is high. On the one hand, the severity of changes depends on the severity of failure. On
the other hand, slight changes due to different operating conditions should not be led to
triggering of an unnecessary alarm signal. Therefore, an algorithm should be used for
detection of changes in the estimated parameters which should be not only simple enough
in order to integrate into an online monitoring system but also accurate enough in order to
be able to detect failures as soon as possible.

As an instance, Fig. 6.13 shows the changes in the estimated parameter uair in (3.27) before
and after the occurrence of a failure in fan operation of the Tr1 on December 21st at 6:00. As
it can be seen, the estimated parameter uair varies slightly before the failure; nonetheless,
when the failure occurs, the value of the estimated parameter changes suddenly. Defining
a threshold value is not applicable for every transformer. Therefore, this sudden change
should be detected using an online algorithm in such a way that the algorithm can be used
for every transformer with AF cooling system.
6.2 Changes of the Estimated Parameters for Detection of Fan Failures 91

4.9

uair

4.85

4.8
/m/s

4.75

21−Dec; 6:00
4.7

4.65
10−Dec 15−Dec 20−Dec 25−Dec 30−Dec

Figure 6.13: Changes in an estimated parameter of the air convection for Tr1.

6.2.1 Hypothesis Testing

Assuming that the values of the estimated parameters before and after failure are known
and equal to uair,0 and uair,1 respectively, the idea of the online detection of changes in
the estimated parameters is based on taking samples with fixed size N and computing a
decision rule at the end of each sample to test between two following hypotheses about the
parameter uair [120]:

H0 : uair = uair,0 , (6.10)

H1 : uair = uair,1 . (6.11)

As long as the decision is taken in favor of H0 (called null hypothesis), the sampling and
test continue indicating that no failure has occurred. The alarm signal will be triggered
after the first sample of observations for which the decision is taken in favor of H1 (called
alternative hypothesis) [120]. If the probability of the alternative hypothesis is less than or
equal to the probability of the null hypothesis, the failure is detectable [120]. Let p0 and p1
are the probability of the null and alternative hypotheses respectively. For testing between
H0 and H1 , there exists a test ζ and a constant ξ such that [120]:
(
p1
1, when ≥ξ
ζ= p0
p1
(6.12)
0, when p0
<ξ
92 6 Monitoring of the Cooling System

The constant value ξ should be determined in such a way that the test in (6.12) can be
applied on every transformer. The online algorithm is responsible for calculation of the
probabilities of each hypothesis and testing (6.12).

6.2.2 The Proposed Algorithm for Detection of Changes in the

Estimated Parameters

The proposed algorithm for detection of changes in the estimated parameters is based on
the calculation of the probability of the estimated parameters at each time and comparing
them using the Logarithm of the Likelihood Ratio (LLR).

The estimated parameters by least-squares method are distributed normally with the mean
µi and the variance σi . The probability of the occurrence of the estimated parameters is
calculated as follows [120]:
  2 
1  φ̂i − µi
 
p φ̂i = p exp − , (6.13)

2π (σ 2 )i 2 (σ 2 )i
 2  
σ 2 i = E φ̂i − E 2 φ̂i , (6.14)


 
µi = E φ̂i , (6.15)

where φ is the estimated parameters, i is the current time step, E (·) denotes the expected
value, and superscript ˆ indicates a certain estimated parameters.

The LLR is a strong index for detection of changes in the estimated parameters [120].
It compares the logarithm of the probability of the occurrence of two consecutive sets of
estimated parameters which can be calculated as follows:
  

p φ̂i
LLRi = 2 × log    . (6.16)
p φ̂i−1

For sufficiently large number of measurements, the LLR is a random variable with the
distribution function very close to the asymptotically chi-squared (χ2 ) distribution with
the degree of freedom equal to the dimension of the estimated parameters [121] (in this
context equals to 1), according to which the allowed band for the LLR can be defined. The
threshold value for the LLR in the algorithm is considered equal to 3.8 corresponding to
the 95% confidence interval of the χ2 distribution. It explicitly means that if the LLR of the
6.2 Changes of the Estimated Parameters for Detection of Fan Failures 93

estimated parameters is greater than 3.8 or lower than -3.8, it will be considered as a faulty
condition. This can be mathematically formulated as follows:

LLRi > 3.8 ∨ LLRi < −3.8. (6.17)

It should be noted that for reaching a wider confidence interval, a smaller threshold value
may be used which may lead to unnecessary triggering of alarm signal; on the con-
trary, if a lower confidence interval is considered, the non-severe failures may not be de-
tected. Therefore, depending on the application, a suitable confidence interval should be
selected.

The flowchart of the proposed algorithm for detection of changes in the estimated param-
eters of a thermal model is depicted in Fig. 6.14 according to which the following steps
should be taken:

Start

i = k +1

First measurement

Calculation of the
top-oil temperature
for other applications
Parametrizing of the model

Calculation of the likelihood of

Next each parameter (P(i)) Next
measurement measurement
(i=i+1) (i=i+1)
Calculation of log-likelihood ratio

LLRi = 2 ´ log
( )
p fˆi
p (fˆ )
i -1

Alarm=0 No LLRi > 3.8 Yes Alarm=1

Figure 6.14: Flowchart of the proposed algorithm for detection of changes in unknown
parameters of thermal models.
94 6 Monitoring of the Cooling System

1. Parametrizing the thermal model. In this step, the design-dependent and empirical
factors of the thermal model (listed in Table 3.1) are estimated. The required values
for parametrization of the thermal models are the measured input parameters, namely
the load factor; the ambient temperature; number of running pumps and fans; the
measured top-oil temperature; the constant parameters; namely no-load and short cir-
cuit losses; and the estimated parameters of the previous time step (φi in (A.4)). The
estimated parameters can be further used for other applications e.g. the calculation of
the hot-spot temperature, the determination of the loading capability, and etc.
2. Calculating the likelihood of the estimated parameters. The inputs of this step are the
estimated parameters of the air-side heat transfer in the current time step as well as
their means and variances corresponding to all previous time steps. The involved
equations in this step are (6.13)-(6.15). It is worthwhile to note that the oil-side
heat transfer has impacts on the air-side heat transfer which has been described in
section3.1.2.
3. Calculating the LLRs. The inputs of this step are the likelihood of the current and
previous time steps. The involved equation in this step is (6.16).
4. Checking the allowed band. If the allowed band in (6.17) is violated, the alarm signal
will be triggered and the remedial actions should be carried out and the algorithm
waits for the next measurements. If it is not the case, the algorithm waits for the next
measurements.
5. Go to step 1.

6.2.3 Validation of the Proposed Algorithm through Two Case Studies

The proposed algorithm is applied on Tr1 and Tr2 whose characteristics are listed in Table
5.1. The proposed online algorithm in this section is responsible for detection of the fans
failures and the time of the occurrence of each failure. It is worthwhile to note that for
determination of the number of the failed fans, (3.2) should be solved for the number of
operative fans in (3.27) which is an optimization problem and can be easily done using the
non-linear least-squares method described in section A.1.

Fig. 6.15 illustrates the estimated parameters of the convection in the oil and the air during
December for Tr1, when a failure occurred in the cooling system of the transformer. The
two upper rows of the sub-figures correspond to the estimated parameters of the air-side
convection, whereas the two lower rows correspond to the oil-side convection parameters.
As it can be seen, when a group of fans fails to operate, although the temperature of the oil
changes, the estimated parameters of the oil-side remain constant. On the contrary, accord-
ing to Fig. 6.15, some estimated parameters of the air-side react rapidly to the undesirable
6.2 Changes of the Estimated Parameters for Detection of Fan Failures 95

5 85.05 2

85 0.2
4.8 84.95 1.9
0.1
84.9
4.6 84.85 1.8 0
u A p q
air
air air air
3 1 0.8 1

0.8
2.5 0.5 0.7
0.6

2 0 0.6 0.4
s C mair n
air air air
1.1 0.092 4.05
2.05

1 0.091 4 2

0.9 0.09 3.95 1.95

uoil Aoil poil noil

1.55 0.15
0.6 0.6

1.5 0.5 0.5 0.1

1.45 0.4 0.4 0.05

5−Dec 13−Dec 21−Dec 30−Dec 5−Dec 13−Dec 21−Dec 30−Dec 5−Dec 13−Dec 21−Dec 30−Dec 5−Dec 13−Dec 21−Dec 30−Dec
soil Coil moil noil

Figure 6.15: Qualitative representation of changes in the estimated parameters of the pro-
posed thermal model, when a failure occurs in the cooling system of Tr1.

change which leads to increase the second term of the right side of (A.4). This change
should be detected automatically which will not be possible without online monitoring sys-
tem.

In order to better explain the method for calculation of the LLR, the proposed procedure in
section 6.2.2 is described with an example of one estimated parameter for Tr1. Assuming
that the estimated parameter uair as a representative of φ is φi−1 = 4.8265 before the
occurrence of the failure (in time step i − 1) and the corresponding mean value (µ in (6.15))
and the variance (σ in (6.14)) for all of previous time steps are equal to µi−1 = 4.8658
and σi−1 = 0.0258 respectively. The standardized likelihood of this situation is p (φi−1 ) =
0.1254 according to (6.13). If the same procedure is carried out for the faulty situation, the
standardized likelihood is p (φi ) = 7.3468 × 10−6 . Hence, the LLR can be calculated using
(6.16) as follows:

7.3468 × 10−6
 
LLRi = 2 × log = −8.4644.
0.1254

Therefore, the calculated LLR violates the allowed band in (6.17) leading to triggering of
the alarm signal.

Fig. 6.16 shows the calculated LLR of the estimated parameters of the air-side for Tr1.
As it can be seen, the changes of the parameters, which are related directly to the forced
convection of the air (in (3.25)) especially uair (proportionality factor of the fans) and sair
96 6 Monitoring of the Cooling System

(characteristic length of the air convection) are significant after the occurrence of the failure
in the fan and exceeding the allowed band (|LLR| < 3.8) leading to triggering of the alarm
one time step after the occurrence of the failure. Moreover, it is expected that the empirical
factors of the forced convection of the air change significantly after the failure. But, Tr1
has 33% more cooling capacity that it needs which explicitly means that this failure is not
a severe failure for Tr1. However, the algorithm is able to detect the failure, even if the
failure is not severe.

Furthermore, the reason for the fluctuations in the LLR of the estimated parameters of the
air-side which are observable at the beginning of each sub-figures of Fig. 6.16 can be
mentioned as follows:

– The estimated parameters by least-squares method are distributed normally with the
mean µi and the variance σi . The true value of the mean and the variance at the starting
point of the algorithm is unknown; however, they will be updated with new measure-
ments. In other words, the likelihood of the normal distribution can be calculated if
sufficiently large number of measurements are available.
– The distribution of the LLRs are very close to asymptotically chi-squared (χ2 ) distri-
bution for sufficiently large number of measurements.

In this case, approximately 50 time steps are needed for online determination of the true
values of the mean and variance as well as for the distribution of the LLRs to be close
to the asymptotically chi-squared (χ2 ) distribution. This may be perceived as the time
for providing a reliable result by the proposed algorithm. Yet, the fluctuations do not
exceed the allowed band implying that the algorithm does not trigger false alarm sig-
nals.

Fig. 6.17 illustrates the LLRs of the estimated parameters of the air-side for Tr2. As
it can be seen, due to the severity of the failure, all of the estimated parameters of the
air-side related to the forced convection react rapidly to the failure leading to triggering
of the alarm signal one time step after the failure. Before the occurrence of the failure
no violation from the allowed band is observable from the LLR of the estimated parame-
ters; on the contrary, after the failure, the measured and calculated temperature are not in
agreement leading to increasing of the second term on the right side of the (A.4) and con-
sequently increasing of the LLR of the parameters and violation of the prescribed allowed
band (|LLR| < 3.8).

Meanwhile, it is worthwhile to summarize the results of the application of the proposed

algorithm on both transformers under study in this section. Table 6.1 summarizes the cal-
culated LLRs of the parameters which are affected by the failure in the fan operation. As
it can be seen from the highlighted values in Table 6.1, all of the parameters related to the
6.2 Changes of the Estimated Parameters for Detection of Fan Failures 97

3.8 3.8 3.8 3.8

0 0 0 0
LLR/1

−3.8 −3.8 −3.8 −3.8

21−Dec 7:00

−10 −10 −10 −10

u A p q
air air air air

3.8 3.8 3.8 3.8

0 0 0 0
LLR/1

−3.8 −3.8 −3.8 −3.8

−10 −10 −10 −10

5−Dec 13−Dec 21−Dec 30−Dec 5−Dec 13−Dec 21−Dec 30−Dec 5−Dec 13−Dec21−Dec 30−Dec 5−Dec 13−Dec 21−Dec 30−Dec
s C m n
air air air air

Figure 6.16: The calculated LLR of the estimated parameters of the air-side for Tr1.

3.8 3.8 3.8 3.8

0 0 0 0
LLR/1

−3.8 −3.8 −3.8 −3.8

−10 −10 −10 −10

21−Jan 15:00

−15 −15 −15 −15

u A p q
air air air air

3.8 3.8 3.8 3.8

0 0 0 0
LLR/1

−3.8 −3.8 −3.8 −3.8

−10 −10 −10 −10

−15 −15 −15 −15

21−Jan 30−Jan 7−Feb 21−Jan 30−Jan 7−Feb 21−Jan 30−Jan 7−Feb 21−Jan 30−Jan 7−Feb
s C m n
air air air air

Figure 6.17: The calculated LLR of the estimated parameters of the air-side for Tr2.

forced convection of the air except Aair are affected seriously by the fans failure in such a
way that their LLRs exceed the allowed band (|LLR| < 3.8); therefore, it is suggested to
use all of them for detection of failure in fan operation, even though some of them are less
affected compared with the other. Furthermore, among the LLRs listed in Table 6.1, the
most important one is the proportionality factor uair which is directly related to the number
of operative fans according to (3.27).
98 6 Monitoring of the Cooling System

Table 6.1: Summary of the application of the proposed algorithm on both transformers
under study.

Tr1 Tr2
Actual number of operative fans 4 3
Recorded number of operative fans 6 6
Failure occurred on December 21st at 6:00 January 21st at 14:00
Failure detected on December 21st at 7:00 January 21st at 15:00
before failure 0.0170 -1.2931
LLR of uair
at the time of the failure -8.4644 -5.0303
before failure 0.0086 0.0266
LLR of Aair
at the time of the failure 1.2473 1.2433
before failure 0.0017 0.2263
LLR of pair
at the time of the failure -0.2050 1.0142
before failure 0.0027 0.0950
LLR of qair
at the time of the failure -0.4321 -1.7204
before failure 0.0155 0.0527
LLR of sair
at the time of the failure -5.6426 -1.2044
before failure 0.0312 -0.8735
LLR of Cair
at the time of the failure -0.6660 -6.9808
before failure -0.0815 0.8071
LLR of mair
at the time of the failure 0.4817 -13.6450
before failure 0.0165 -1.0413
LLR of nair
at the time of the failure -1.7793 -6.9381

The healthy frequent switching of the fans should not lead to abrupt changes in the esti-
mated parameters of the air-side equations. Although the parameters of the air-side equa-
tions change with the healthy switching of the fans, these changes do not lead to the LLRs
to exceed the allowed band. Fig 6.18a shows the healthy operation of fans for Tr1 during
July. This frequent switching affects the estimated parameters of the air-side, whereas the
parameters of the oil-side equations remain constant. This leads to variation of the calcu-
lated LLR corresponding to each parameters of the air-side over time. However, as it can
be seen from Fig. 6.18b, the LLRs do not exceed the allowed band |LLR| < 3.8 indicating
that the proposed algorithm does not trigger false alarm signal.

As it is stated before, the algorithm is able to detect the non-severe failure too. It should
however be noted that a severe change in the loading condition of the transformer will not
be perceived by the algorithm as a failure. The severe changes in the loading condition
occurs often in wind farm transformers or in the cases in which the transformer on a bay is
taken out for maintenance (or due to an outage) and the whole load will be served by the
transformer in a neighboring bay.
6.2 Changes of the Estimated Parameters for Detection of Fan Failures 99

The severe load changes of transformers affect the parameters of the oil convection which
do not change in case of fan failures (see Fig. 6.15). In order to validate this claim,
the real-time measured data of a 40 MVA ONAN wind farm transformer (Tr9), which is
subjected to the severe load changes, are used. Fig. 6.19 shows the measured ambient
temperature, the measured top-oil temperature, and the measured load factor of the above
mentioned transformer in March. As it can be seen, the load of the transformer changes
rapidly from 0% to 80% and vice versa which leads to the rapid changes in oil tempera-
ture.

10

8
Number of Running Fans/1

0
03−Jul 08−Jul 13−Jul 18−Jul 23−Jul 28−Jul

(a)

3.8 3.8 3.8 3.8

LLR/1

0 0 0 0

−3.8 −3.8 −3.8

−3.8
pair qair
uair Aair

3.8 3.8 3.8

3.8
LLR/1

0 0 0 0

−3.8 −3.8 −3.8 −3.8

08−Jul 18−Jul 28−Jul 08−Jul 18−Jul 28−Jul 08−Jul 18−Jul 28−Jul 08−Jul 18−Jul 28−Jul
s C mair n
air air air

(b)

Figure 6.18: Validation of the algorithm in the case of the healthy operation of fans for
Tr1. (a) Frequent Switching of the fans. (b) Qualitative representation of changes in the
estimated parameters of the proposed thermal model, when the switching of the fans are
healthy.
100 6 Monitoring of the Cooling System

Since the transformer is an ONAN transformer, the parameters of the forced convection of
the air and the oil do not play roles. These parameters are namely C, m, n and u of the air-
side and the oil-side in Table 3.1. Fig. 6.20 shows the estimated parameters of the natural
convection of the oil and the air corresponding to the measured data shown in Fig. 6.19.
As it can be seen, the parameters of the oil convection change, if the load varies extremely;
nevertheless, the parameters of the air convection remain constant and the corresponding
calculated LLRs are always equal to zero. However, the LLR of the oil-side parameters
change correspondingly. Therefore, since the parameters of the air convection are used for
detection of failures in fan operations, the algorithm does not perceive the load changes as
failure.

It is however worthwhile to note that the proposed algorithm for detection of changes in
the estimated parameters of a thermal model is based on the regression theory having the
demerit that for the detection of the clearance of the failure, the algorithm should deviate
somehow from the pure theory. In other word, the probability of the occurrence of two
sequential events are compared as shown
 in (6.16); however, after the failure occurrence,
the probability of each point p φ̂i should be compared with the probability of the last
 
point before occurrence of the failure p φ̂bef ore . Another demerit concerning the online
algorithm for detection of the changes in the estimated parameters of the thermal model
is that in comparison with the algorithm proposed in section 6.1 it requires significantly
higher computational efforts since the model should be parametrized for each time step and
the controlling of the estimated parameters is not a trivial task. Moreover, the detection of
the pump failure is not possible with this algorithm.

100
Load Factor
80
60
/%

40
20
0
60
50
40
30
/°C

20
10
0
−10 Measured Ambient Temperature
Measured Top−Oil Temperature
03−Mar 08−Mar 13−Mar 18−Mar 23−Mar 28−Mar

Figure 6.19: Measured values during normal operation of the Tr9.

6.2 Changes of the Estimated Parameters for Detection of Fan Failures 101

1 4.02

0.5 4

0 3.98
soil Aoil

2.1 2

2 1.5

1.9 1
qoil poil

4.22 77.9

4.21 77.8

4.2 77.7
sair Aair

2 −1

1.5 −1.5

1 −2
06−Mar 16−Mar 26−Mar 06−Mar 16−Mar 26−Mar
qair pair

Figure 6.20: Qualitative representation of changes in the estimated parameters of the pro-
posed thermal model for Tr9 under severely volatile loading conditions.
 103

7 Assessment of the Loading Capability

"The picture contains the possibility of the state of affairs which it represents."

Tractatus Logico-Philosophicus
L. Wittgenstein

O
VERLOADING of power transformers can become necessary with the integration of
renewable energy technologies due to economic reasons or simply to make the best
use of the existing assets. However, without consideration of technical issues, indirect costs
may be imposed and consequently affect the benefit of the operation due to the accelerated
aging and damages during overloading [122]. The winding temperature [35] and the perfor-
mance of the cooling system [22, 123] are important limiting factors for loading of power
transformers.

If a transformer is overloaded, the temperatures of the winding and the oil are near or,
in some cases, over the allowable temperatures corresponding to unity relative aging rate
determined by their thermal classes [22,58]. Hence, for prevention of the lifetime reduction
of a transformer, it should be ensured that the heat generated in the winding is properly
transferred to the surrounding ambient through the oil inside and the air outside of the
transformer. The transfer of the heat to the surrounding ambient can be investigated using
a thermal model which determines how much a transformer can be overloaded without
suffering accelerated thermal aging [8].

The loading of transformers is mainly limited by their windings hot-spot temperature [9].
The prescribed winding temperature is equal to 98 °C for non-thermally upgraded papers
[8] and 110 °C for thermally-upgraded papers [8, 9] which are based on a reference ambi-
ent temperature which is equal to 30 °C (yearly average ambient temperature) according to
IEEE loading guide [9] or 40 °C (maximum ambient temperature) according to IEC loading
guides [8, 21]. Hence, the ambient temperatures lower than the reference ambient tempera-
ture may offer the planned loading of transformers beyond the nameplate rating subject to
the condition that the hot-spot temperature should not exceed the maximum prescribed tem-
perature. This way of assessment of the loading capability of transformers guarantees the
unity relative aging rate of the transformer in every periods of time.
104 7 Assessment of the Loading Capability

Moreover, the unity relative aging rate over a year may be achieved by the compensation
of the sacrificed life of transformers in that periods of time where the ambient temper-
ature or the transformer load is low. In this way, although transformers experience ac-
celerated thermal aging in some periods of time, the yearly relative aging rate is equal
to one. In this method, there are some times, in which the hot-spot temperature exceeds
98 °C. This method provides the possibility for the planned loading over nameplate rat-
ing.

7.1 Loading Beyond Nameplate Rating

Power transformers’ life expectancy at various operating temperatures is not accurately

known, but the information given regarding loss of insulation life at elevated temperature is
the best that can be produced from present knowledge of the subject.

In view of the present tendency to overload the transformer at low ambient temperatures,
it seems desirable to determine the overload values which are safe, rather than let the sit-
uation drift along with the result that in some cases unsafe values may be used. To allow
the operator to assign the rating to an expensive apparatus depending upon the operating
conditions is a very important matter and anything that can be done to make it safe is worth-
while.

To provide a guidance on risk associated with higher operating temperature, three different
loading conditions beyond nameplate have been defined as follows:

• Normal Cyclic Loading: Loading beyond nameplate rating may be applied on trans-
formers during part of the load cycle which can be lead to sacrificing of the lifetime
of transformers. But this sacrificing is compensated during rest of the load cycle
specially in low ambient temperatures or low loading regime [8]. In normal cyclic
loading, the hot-spot temperature and the daily loss of life are the limiting factor for
the dynamic loading [6]. This principle can be extended for long periods of time for
the purpose of the planned loading beyond the nameplate rating.

• Long-Time Emergency Loading: This type of dynamic loading is referred to the load-
ing resulting from prolonged outages of some system elements that will not be recon-
nected before the transformer reaches a new and higher steady-state temperature [8].
In the long-time emergency loading, the loss of life and the bubbling inception tem-
perature are the limiting factors of the loading [6].
7.1 Loading Beyond Nameplate Rating 105

• Short-Time Emergency Loading: This type of dynamic loading is referred to unusu-

ally heavy loading of a transient nature (less than 30 min) due to the occurrence of
one or more unlikely events which seriously disturb normal system loading [8]. The
bubbling inception temperature is the limiting factor for the short-time emergency
loading of transformers [6].

During overloading of transformers the following aspects should be taken into account
[8, 9]:

• The temperatures of windings, cleats, leads, insulation and oil should not reach an
unacceptable level [8].

• The leakage flux density outside the core increases causing additional eddy-current
heating in metallic parts linked by the leakage flux. This may lead to the generation
of the gas bubble at the surfaces of heavy metallic parts heated by the leakage flux or
be produced by super-saturation of the oil [9].

• As the temperature changes during overloading of transformers, the moisture and gas
contents in the insulation and in the oil should not exceed the critical values [8, 9].

• Bushings, tap-changers, cable-end connections and current transformers will also be

exposed to higher stresses whose values should not exceed their design margin [8, 9].

• Due to the possible presence of the gas bubbles, the dielectric strength in the region of
high electrical stress will be reduced. These bubbles are likely to occur when the hot-
spot temperature exceeds 140 °C for a transformer with a winding insulation moisture
content of about 2% [8].

• Temporary deterioration of the mechanical properties at higher temperatures could

reduce the short-circuit strength [8, 9].

• The expansion of the oil could cause overflow of the oil in the conservator [8, 9].

• Breaking of excessively high currents in the tap-changer could be hazardous [8, 9].

• Pressure build-up in the bushings may result in a failure due to oil leakage. Gassing in
condenser type bushings may also occur if the temperature of the insulation exceeds
about 140 °C [8, 9].

• The gasket materials in the transformer may become more brittle as a result of ele-
vated temperatures [8, 9].
106 7 Assessment of the Loading Capability

Moreover, a large transformer could be more vulnerable to loading beyond nameplate rating
than a smaller one. In addition, the consequences of a transformer failure are more severe
for larger sizes than for smaller units. Moreover, the cooling capacity of transformer plays
an important role during overloading periods.

Therefore, every effort towards the accurate determination of the loading capability of trans-
former is worthwhile. With development of online transformer monitoring applications and
inclusion of dynamic thermal models, real-time dynamic loading and rating of transformers
by the hot-spot temperature instead of the nominal current is viable. The determination of
the dynamic loading capability of transformers offers network operators the chance to take
full advantage of the transformer thermal performance, enables safe overloading, and keeps
the transformer lifetime consumption under control.

Dynamic thermal model enables prediction (calculation) of the transformer hot-spot tem-
perature from the measured ambient temperature, cooling and load profiles [6]. The deter-
mination of the dynamic loading capability on the other hand, represents a reversed dynamic
thermal modeling process, i.e. determination of the load factor limit that allows the trans-
former to be operated below the chosen hot-spot temperature limit. The difference between
this type of overloading and the formerly described planned loading beyond nameplate rat-
ing is that in the dynamic overloading, the hot-spot temperature will not exceed the maxi-
mum temperature prescribed by the thermal class of the insulation; therefore, the lifetime of
the transformer will not be sacrificed during a part of the load cycle. In other word, the rela-
tive aging rate of the transformer is always equal to or less than one.

Transformers may be loaded above rating for any period of time provided that the hot-spot
temperature is not greater than 98 °C [6, 8, 9].

The hot-spot temperature of the windings is calculated as follows [9]:

θhs = θto + HF · gr · K y . (7.1)

The hot-spot factor HF = 1.3 can be considered accurate enough for medium and
large power transformer whereas for standard distribution transformers with a short-circuit
impedance less than 8% the value of HF = 1.1 is accurate enough for loading considera-
tions [8]. The dynamic loading capability of transformers under each operating condition
can be yielded by solving (7.1) for K subject to θhs = 98 °C. Since the top-oil temperature
is dependent on the load factor too, solving this non-linear equation is not straightforward
but an optimization problem. In this work, the non-linear least squares method described in
7.2 Dependency of the Loading Capability on the Ambient Temperature 107

section A.1 is used to solve (7.1) for K as follows:

min k98 °C − θhs (K)k22 . (7.2)

The dynamic loading capability depends fundamentally on three variables: ambient tem-
perature, intensity of transformer cooling (cooling stage), and the chosen criterion for the
maximum load factor (in this case θhs = 98 °C). Therefore, it is assumed that the whole
cooling capacity is available during overloading of the transformer. In this context, it is very
important to monitor the condition of the cooling system using an online monitoring system
before the transformer is overloaded. Thus, the loading capability is a function of the ambi-
ent temperature. Using the benefit of the low ambient temperature, assuming that the condi-
tion of the cooling system of the transformer is observable, and considering the design lim-
itation of other components e.g. the bushing, etc., the transformer can be overloaded safely
without sacrificing its lifetime or increasing its relative aging rate.

7.2 Dependency of the Loading Capability on the Ambient

Temperature

For integration of the calculation of the dynamic loading capability into online monitoring
systems, the procedure of the assessment may be simplified since a model-based online
monitoring system has, in some cases, just 10 minutes to estimate the parameters, to cal-
culate the top-oil and the hot-spot temperatures, to determine the loading capability, and
to estimate the loss-of-life. In some cases, the time for dealing with these tasks increases
significantly. Moreover, it is assumed that the unknown parameters used in the proposed
model are design-dependent parameters or empirical factors whose values do not change
under different in-service conditions. Therefore, the loading capability is only dependent on
the ambient temperature. In the case of indoor transformers, in addition to the ambient tem-
perature, the station temperature affects the loading capability too.

According to the CIGRÉ Brochure 659 [6], there is a linear dependency between
the loading capability of outdoor transformers and the ambient temperature as fol-
lows:
LC = a + b · θa , (7.3)

where a and b are empirical factors which can be estimated based on the loading capability
calculated using (7.2) during parametrization phase.
108 7 Assessment of the Loading Capability

However, in the rest of this section, it is shown that the linear dependency is not valid for
all cooling systems under all operating conditions. In this section, a quadratic function is
considered to describe this dependency as follows:

LC = a + b · θa + c · θa2 , (7.4)

where a, b, and c are empirical factors which can be estimated based on the loading capa-
bility calculated using (7.2) during parametrization phase.

Moreover, this dependency may be assumed to be linear for indoor distribution transformers
with a good approximation as follows:

LC = a + b · θa + c · θs , (7.5)

where a, b, and c are empirical factors which can be estimated based on the loading capa-
bility calculated using (7.2) during parametrization phase.

Once these factors are estimated, they are valid for all operating conditions. It is shown in
the rest of this section that (7.4) and (7.5) are good approximations for determination of the
loading capability.

7.2.1 Investigation of the Concavity of the Dependency

Fig. 7.1 shows the loading capability of Tr1 as a function of the ambient temperature. The
colored points are the loading capability determined using the proposed model which is
parametrized in June. During parametrization of the proposed model, the loading capabil-
ity of the transformer as a function of the ambient temperature can be derived. As it can be
seen from the red points in Fig. 7.1 which represent the loading capability of Tr1 in June
determined using the training data set, the dependency of the loading capability on the am-
bient temperature is more or less concave. Two lines are fitted to the red points to provide
simple equations which may be applicable for different ambient temperatures. These equa-
tions are the linear and quadratic equations described by (7.3) and (7.4) respectively. The
solid and the dashed red lines in Fig. 7.1 are the linear and quadratic fitted curves to the red
points respectively. In other words, the linear and the quadratic equations are determined
by the ambient temperatures in range of approximately 10 °C and 30 °C. For this trans-
former in this temperature range, the dependency of the loading capability on the ambient
temperature is not highly concave; therefore, in this temperature range, the linear equation
does not deviate from the quadratic equation. However, as it can be seen from the green
points in Fig. 7.1, the linear equation shows a deviation from the quadratic equation at very
7.2 Dependency of the Loading Capability on the Ambient Temperature 109

220
1.8960−0.0138×θ
a
−5 2
1.8960−0.0126×θ −3.1399e ×θ
a a
200
Proposed Model in June
Proposed Model in August

Loading Capability/%
Proposed Model in December
180

160

140
133

−20 −10 0 10 20 30 40 50
Ambient Temperature/°C

Figure 7.1: Loading capability as a function of the ambient temperature for Tr1 determined
using the data set measured during in-service operation in June.

low ambient temperatures. The same deviation exists at very high ambient temperatures
(near 40 °C). It should be noted that if the parameters of the linear equation are estimated
using high ambient temperatures (between 30 °C and 40 °C), the deviation at low ambient
temperature will be higher. In other words, firstly, the accuracy of the linear equation de-
pends on the parametrization data set; and secondly, the linear equation overestimates the
loading capability of the transformer which may lead to the accelerated aging of the trans-
former. However, for this transformer, the concavity of the dependency leads to a deviation
of approximately 3% which may not be problematic for Tr1.

It should be noted that due to the excessive cooling capacity of Tr1, it has a high loading
capability. Tr1 has 33% more cooling capacity than it needs. However, limitations concern-
ing other parts of a transformer, e.g., bushings and on-load tap changer during overloading
periods, should also be kept in mind.

Moreover, Fig. 7.2 shows the ambient temperature, current load factor, and calculated
loading capability of Tr1 using the proposed model, the linear and quadratic equations
in December with the parameters estimated using the data set measured during normal
operation of the transformer in June. The calculated loading capability using the proposed
model is considered as the reference value in the rest of this section. As it can be seen, the
loading capability determined using the linear equation is always higher than that of the
quadratic equation and the proposed model which may lead to increasing the relative aging
rate of the transformer. This deviation is higher at very low ambient temperature. The mean
absolute and the maximum deviations of the linear equation from the proposed model are
110 7 Assessment of the Loading Capability

20
Ambient Temperature
10

/°C
0
−10
−20
60
/% 40
20
Current Load Factor

210 Overload Capability using the Proposed Model

Overload Capability using the Linear Equation
Overload capability using the Quadratic Equation
200
/%

190

180

170
05−Dec 10−Dec 15−Dec 20−Dec 25−Dec 30−Dec

Figure 7.2: Loading capability of Tr1 in December determined using the proposed model,
the linear and quadratic equations.

0.4% and 2.3% respectively.

Moreover, if a wider range of ambient temperatures is used for parametrizing of the linear
dependency of the loading capability on the ambient temperature, the deficiency of the
linear equation will not be resolved. Fig. 7.3 shows the loading capability as a function of
the ambient temperature for Tr3 which is an OFAN transformer. In order to parametrize the
linear and quadratic equations, the loading capability of the transformer calculated using
the proposed model in April has been used. The training data set contains a wide range of
ambient temperatures between 0 °C and 30 °C. In addition, Tr3 experienced very high
ambient temperatures in summer seasons and very low ambient temperatures in winter
seasons. As it can be seen, for very high and very low ambient temperatures the linear
equation cannot be a good representative of the loading capability-ambient temperature
dependency.

However, if the parameters of the linear model are estimated by high ambient temperatures
the deviation of the linear equation from the thermal model will be as high as approximately
10% which may lead to accelerated thermal ageing of Tr3. In other words, the accuracy
of the linear equation depends on the parametrization data set which is not the case for
quadratic equation. Moreover, the quadratic equation is accurate enough for this purpose.
Furthermore, Fig. 7.4 shows the loading capability of Tr3 during July, when the maximum
deviation of the linear equation from the quadratic equation and the proposed model oc-
curs. The maximum deviation reaches 2.6% which may affect the relative aging rate of the
transformer seriously. In addition, once the ambient temperature exceeds the 30 °C, the rate
7.2 Dependency of the Loading Capability on the Ambient Temperature 111

180
1.5642−0.0139×θ
a
170 1.5583−0.0127×θ −4.2015e−5×θ2
a a
Proposed Model in April
160 Proposed Model in July
Proposed Model in Januray
Loading Capability/%
150

140

130

120

110

100

90
−20 −10 0 10 20 30 40 50
Ambient Temperature/°C

Figure 7.3: Loading capability in April, July, and January as a function of ambient tem-
perature for Tr3 determined using the data set measured during in-service operation in
April.

40
30
/°C

20
Ambient Temperature
10
100
80
60
/%

40 Current Load Factor

20
0
Overload Capability using the Proposed Model
135 Overload Capability using the Linear Equation
130 Overload Capability using the Quadratic Equation
125
120
/%

115
110
105
100
13−Jul 14−Jul 15−Jul 16−Jul 17−Jul

Figure 7.4: Loading capability of Tr3 in July determined using the proposed model, the
linear and quadratic equations.

of the deviation of the linear equation from the reference value increases which may lead to
more sacrificing of the lifetime of the transformer. In the case of emergency overloading,
the sacrificed life time of the transformer should be compensated at lower ambient temper-
ature in order to have a yearly relative aging rate equal to unity.
112 7 Assessment of the Loading Capability

7.2.2 Loading Capability of Wind Farm Transformers

The determination of the loading capability of wind farm transformers is a crucial task
since they have a highly volatile loading profile varying suddenly from very low loading
to overloading conditions. Without dynamic determination of the loading capability, their
lifetime may be sacrificed in periods in which the transformer is loaded beyond nameplate
rating.

Tr5 whose characteristics are listed in Table 5.1 is a wind farm ONAN transformer with
a very volatile loading condition. As described in section 5.6, the proposed model can
calculate the top-oil temperature of the transformer very well, even in case of overload-
ing conditions. This guarantees that the loading capability determined using the proposed
model is reliable. The proposed thermal model in section 3.1 has been parametrized with
the data set measured during normal operation of the transformer in December. Using this
data set, the loading capability of the transformer can also be determined and the depen-
dency of the loading capability on the ambient temperature can be investigated for this
transformer.

Fig. 7.5 shows the loading capability of Tr5 as a function of ambient temperature. The
red points illustrate the loading capability determined using the proposed model in De-
cember (The data set for parametrizing of the model and the loading-ambient temperature-
dependency). The solid red line is the linear equation fitted to the red points whereas the
dashed red line is the fitted quadratic equation to the red points. The blue points are the
loading capability of the transformer from the end of January to the end of March deter-
mined using the proposed model parametrized in December. The dotted line is the linear
equation fitted to the blue points. In the same way, the green points are the loading capa-
bility determined in the same manner on the data set measured and calculated during June.
As it can be seen, the fitted linear equation (the solid line) overestimates the capability of
the transformer to carry the load.

Nonetheless, the quadratic equation which is determined using the data set measured in
December can extrapolate above the range of ambient temperatures used for parametrizing
of the quadratic equation. In other words, as an example at an ambient temperature of 30
°C, what the linear equation determines as the loading capability is approximately 8% more
than the capability of the transformer. This may be much more crucial in the case of the
determination of the loading capability in summer season using the model parametrized
in winter season and vice versa. As an instance, according to the dashed line which has
been determined by data measured in December, the transformer can carry the load equal
to 112% at the ambient temperature of 40 °C. At this ambient temperature, it is expected
that the transformer can carry a load equal to 100%.
7.2 Dependency of the Loading Capability on the Ambient Temperature 113

Moreover, if a wider range of ambient temperatures (the dotted line) is used for the
parametrization of the linear equation, the accuracy of the linear equation will be better
at the high ambient temperatures. In other words, the accuracy of the linear equation de-
pends on the parametrization data set, whereas the parameters of the quadratic equation are
independent of the parametrization data set.

Moreover, Fig. 7.6 shows the loading capability of Tr5 using the proposed model as well as
the linear and quadratic equations in June with the model parametrized by the measured data
in December. The transformer carries the load equal to 122.2% of the nameplate loading
between the 11th and the 16th of June where the ambient temperature is 21.4 °C. The loading
capability of the transformer determined using the proposed model under this condition is
120.4% of the nameplate rating which indicates that the current loading of the transformer
is above the dynamic rating of the transformer which would lead to approximately 5%
increase in the relative aging rate if the hot-spot temperature exceeds the 98 °C which did
not happen for this transformer in this period of time. However, at this ambient temperature,
the loading capability determined using the linear approximation is equal to 125.4%. If the
linear approximation had been used for determination of the loading capability, the relative
aging rate (calculated by (2.26)) could have reached the 115% or 15% more than the unity
relative aging.

7.2.3 Loading Capability of Indoor Distribution Transformers

For indoor distribution transformers, the loading capability of transformers depends not
only on the ambient temperature but also on the station temperature. The proposed method

150

140
Loading Capability/%

130
1.3772-0.0063×θa
120
1.3776-0.0067×θa-6.2364e-5×θ2a
Using the Proposed Model in December
110
Using the Proposed Model in
End of January to End of March
100 1.376-0.007×θa
Using the Proposed Model in June

-20 -10 0 10 20 30 40 50
Ambient Temperature/°C

Figure 7.5: Loading capability of Tr5 from December to March as a function of the ambient
temperature.
114 7 Assessment of the Loading Capability

40
30 Ambient Temperature
21.4 °C

/°C
20
10
0
140
120 Current Load Factor
100 122.2 %
/% 80
60
40
20
0

130
/%

120
120.4 %

110 Loading Capability using the Proposed Model

Loading Capability using the Linear Equation
Loading Capability using the Quadratic Equation
100
06−Jun 11−Jun 16−Jun 21−Jun 26−Jun

Figure 7.6: Loading capability of Tr5 in June calculated using the proposed model, the
linear and quadratic equations.

is applied on Tr6 for which the top-oil temperature was calculated in section 5.7 where
the data set measured during normal operation of the transformer in March was used for
parametrization of the model.

The measured data in March can be used not only for parametrization of the thermal
model but also for estimation of the empirical factors of the linear equation (7.5). In other
words, the top-oil temperature and the loading capability can be calculated for this data set
too.

Fig. 7.7 illustrates the measured ambient and station temperatures as well as the calculated
loading capability of Tr6 in March using the unknown parameters estimated in the same
month. As it can be seen, when the ambient and station temperatures increase, the loading
capability decreases. In order to find an equation for the linear dependency of the loading
capability on the ambient and station temperatures, (7.5) is fitted to the loading capability
in Fig. 7.7. The resulted equation is as follows:

LC = 1.5430 − 0.0121 · θa − 0.0015 · θs . (7.6)

The coefficient of determination (R2 ) as a metric for goodness of fit in this case is equal to
0.99. Equation (7.6) is in a good agreement with the recommendation of the IEEE C57.91
loading guide clause 6 [9], where it is stated that for each degree of ambient temperature
lower than the reference ambient temperature the load of the transformer can be increased
conservatively by 1%. The station temperature does not taken into account directly accord-
7.2 Dependency of the Loading Capability on the Ambient Temperature 115

25
Ambient Temperature Station Temperature
20
15

/°C
10
5
0

160

150
/%

140

130 Loading Capability using the Prposed Model

01−Mar 06−Mar 11−Mar 16−Mar 21−Mar 26−Mar 31−Mar

Figure 7.7: Loading capability of Tr6 determined using the proposed model in March
which has been used to estimated the parameters of the linear dependency.

ing to IEEE [8, 9] and IEC [8] loading guides; it is considered as an extra temperature rise.
Nevertheless, according to (7.6), it can be increased by 1.2%.

Moreover, (7.6) should be valid for other operating conditions. In other words, it is expected
that if (7.6) is applied for other operation conditions, i.e. other ambient and station tem-
peratures, the calculated loading capability should be in a good agreement with the loading
capability calculated using the thermal model and (7.1).

Fig. 7.8 shows the ambient and station temperatures in June, July, August as well as the
loading capability calculated with the thermal model using (7.1) and (7.6), whose factors
are estimated using the calculated loading capability in March (parametrizing data set). As
it can be seen, the deviation between two calculated loading capabilities is small, where the
mean absolute deviation in the calculation of the loading capability is 0.3% and the max-
imum deviation is 1.48%. Moreover, it can be perceived from Fig. 7.8 that although the
station temperature reaches 40 °C in some cases, the ambient temperature is still under the
reference temperature; hence, the loading capability of the transformer stays over name-
plate value. Finally, the linear dependency of the loading capability of indoor distribution
transformers on the ambient and station temperatures is a simplification very close to the
the proposed model.
116 7 Assessment of the Loading Capability

40

30

/°C
20

10 Ambient Temperature Station Temperature

140

130
/%

120

110
Loading Capability Calculated using the Thermal Model
Loading Capability Calculated using the Linear Equation
100
09−Jun 19−Jun 29−Jun 09−Jul 19−Jul 29−Jul 08−Aug 18−Aug

Figure 7.8: Ambient and station temperatures, and the loading capability of Tr6 calculated
using the thermal model and linear equation (7.6) in June, July, and August.

7.3 Summary of the Loading Capability

It is worthwhile to conclude the dependency of the loading capability on the ambient tem-
perature. For power transformers, for definition of the loading capability-ambient temper-
ature dependency, the accuracy of the linear equation depends strongly on the parametriza-
tion data set. If high ambient temperatures are used for parametrization of the linear equa-
tion, the accuracy of the linear equation may be affected at low ambient temperatures and
vice versa; moreover, if a wide range of ambient temperatures (as an example in range of
-10 °C to 30 °C) is used, the accuracy of the linear equation will be higher; however such
a data set is rarely available. On the contrary, the loading capability may be considered
as a quadratic function of the ambient temperature, although in some cases the concavity
of the quadratic equation is not significant at normal ambient temperatures (as an example
in the case of the Tr1). In some cases (as an example in the case of Tr5), the deviation
between the linear representation of the dependency and its quadratic representation is high
at very low or very high ambient temperatures. The accuracy of the quadratic equation
does not dependent on the parametrization data set. In other word, if the quadratic equa-
tion is parametrized using high ambient temperatures, low ambient temperatures, or normal
ambient temperatures, its accuracy will not be changed.

Fig 7.9 shows the loading capability as a function of ambient temperature for Tr1, Tr2, Tr3,
Tr4, Tr5 determined using their training data sets. As it can be seen, except Tr5, all of the
studied transformers have approximately the same loading capability. However, the curves
are more concave in some cases e.g. Tr5 compared with the other cases. Moreover, Tr1
7.3 Summary of the Loading Capability 117

has 33% extra cooling capacity leading to significantly higher loading capability compared
with other transformers under study.
220
Tr5 ONAN Tr1 (ODAF)
200 1.3776-0.0067×θa-6.2364e-5×θ2a 1.8873-0.0129×θa-1.5158e-5×θ2a
Overload Capability/%

Tr2 (OFAF)
180
1.5354-0.0125×θa-1.9197e-5×θ2a
160

140
133
Tr3 (OFAN)
120
1.5583-0.0127×θa-4.2015e-5×θ2a
100 Tr4 (ONAF)
1.4580-0.0117×θa-2.8116e-5×θ2a
80
-10 0 10 20 30 40
Ambient Temperature/°C

Figure 7.9: Loading capability as a function of ambient temperature for Tr1, Tr2, Tr3, Tr4,
and Tr5 determined using the training data sets.

It should be further noted that as a quick approximation according to clause 6 of IEEE

C57.91 loading guide [9] for each degree ambient temperature lower than the average am-
bient temperature (30 °C), the load can be increased for ONAN transformers by 1% of the
rating and for other cooling methods by 0.75% of the rating capacity. These values are
valid for ambient temperature range between -30 °C to 50 °C. Nevertheless, according to
the results presented in this section, the rate of increasing or decreasing of the load at the
ambient temperatures other than the average ambient temperature depends on the ambient
temperature. As an example, assume that Tr5 can carry a load equal to 138% of nameplate
rating at 0 °C. According to the quadratic equation of Tr5, at the ambient temperature of
10 °C, the loading capability is 130%, i.e. 8% lower than that at 0 °C. But at the ambient
temperature of 20 °C, the loading capability is 122% or 12% lower than that at 0°C. If the
dependency had been linear, this value should have been 16%.
 119

8 Conclusions and Future Works

"If something is to be called a method, it must be a procedure in accordance

with principles"

The Critique of Pure Reason

E. Kant

After the decentralizing of power systems and undiscriminating policy of system opera-
tors in addition to the demand growth, system operators should utilize their assets over the
name plate rating. Although it may increase the availability of transformers and the eco-
nomic benefits, without consideration of technical issues, indirect costs (e.g. the cost of the
unplanned replacing of transformers) may be imposed and affect the benefit of operation.
Online thermal monitoring of transformers is used continuously during operation and offers
possibilities to predict the hot-spot temperature, to determine the loading capability, to esti-
mate the lifetime, and to monitor the condition of the cooling system.

8.1 Conclusions

This thesis presents a new concept for thermal monitoring of transformers for calculating
the top-oil and hot-spot temperatures, determining the loading capability, and monitoring
of the cooling system. The main contributions and achievements of this dissertation can be
described as follows:

• A comprehensive review on the dynamic thermal modeling of transformers was pre-

sented from a neutral but critical point of view, in which the origin of the loading
guides and the conventional models were discussed in detail and their advantages and
disadvantages and boundaries of applications were introduced. Moreover, the ad-
vanced dynamic thermal models trying to overcome the demerits of the conventional
models have been presented.

The affecting factors which should be taken into consideration during the develop-
ment or the use of thermal models have been discussed in detail, including: viscosity
120 8 Conclusions and Future Works

of the medium, harmonic contamination of the load current, solar irradiation, wind
speed, tap changer position, geomagnetically induced and direct currents, and mois-
ture content of the solid insulation. Finally, some of the applications of the thermal
models have been reviewed including: prediction of the hot-spot temperature, esti-
mation of the loss-of-life, assessment of the overload capability, online monitoring
of the cooling system, and estimation of the loss-of-life with high penetration of the
photovoltaics and plug-in electric vehicles.

• In order to calculate the top-oil temperature, a new empirical-based thermal model

was proposed in this thesis which uses different heat transfer modes, namely: the
natural and forced convection, the heat radiation, the heat conduction, and the natural
ventilation in transformer building (if applicable) with consideration of their inter-
dependencies. Different heat transfer modes have been connected to each other using
a thermal-electric analogy.

The main advantages of the proposed model is that the model accounts for the
temperature-dependent thermal resistances and the inter-dependencies between dif-
ferent affecting factors. The proposed model has been implemented on a hardware
and validated using measured data during normal operation of transformers with dif-
ferent cooling systems. A variation of the top-oil temperature model has been pro-
posed for indoor transformers. Moreover, another variation of the model has been
proposed and presented which offers the calculation of the top-oil temperature using
the tank temperature. In this procedure, there is no need to measure the top-oil tem-
perature and the installation of the monitoring system can be done during in-service
operation of the transformer. Moreover, using this model, the determination of the
top-oil temperature on transformers having no available temperature pocket is possi-
ble too.

• In order to validate the performance of the proposed models, the unknown parameters
of the models should be firstly estimated. The parameters of the models have been es-
timated by the least squares method using the data measured during normal operation
of the transformers under study. The data sets for parametrization contain load fac-
tor, ambient temperature, number of running pumps and fans (if applicable), station
temperature (in the case of indoor transformers), and measured top-oil temperature
measured during one month. After parametrization or estimation of the unknown pa-
rameters of the models, the model should be validated and its performance should
be evaluated. The results of the validations have shown that the accuracy of the im-
proved models are significantly higher than that of the conventional models and it is
independent of the cooling method.
8.1 Conclusions 121

• In order to monitor the condition of the cooling system, two online algorithms have
been proposed in this thesis. Both algorithms use the proposed dynamic top-oil tem-
perature model as their cores.

The first algorithm is based upon the error in the calculation of the top-oil temperature.
If the standardized error exceeds an allowable band, an alarm signal will be triggered
indicating that a failure has occurred in the cooling system. The allowable band of
this algorithm can be utilized for every transformer with ODAF, OFAF, and ONAF
cooling systems.

The second algorithm, however, is based on the detection of changes in the estimated
parameters of the proposed top-oil temperature model. The changes are detected
based on the calculation of the probability of the estimated parameters followed by
the calculation of the logarithm of likelihood ratio for which an allowed band has
been suggested in this thesis that is applicable on every transformer with AF cooling
system. Moreover, this algorithm provides the possibility of monitoring of the cooling
system during parametrization phase.

Moreover, both algorithms have been validated using the measured data during nor-
mal and faulty operations of two power transformers with ODAF and OFAF cooling
systems. It is noteworthy that although the second algorithm is faster than the first
one, it has some demerits as follows:

– It is more complicated than the first algorithm and needs more computational
efforts.

– The second algorithm can not detect the condition of pumps.

• A method has been proposed for determination of the dynamic loading capability of
transformers. This method is based on the reversed procedure of the calculation of
the hot-spot temperature.

In the calculation of the hot-spot temperature, using a specified loading condition

and ambient temperature, the hot-spot temperature will be determined; whereas in
the determination of the loading capability, the maximum load of the transformer
under the current ambient temperature and the state of the cooling system is calculated
subject to the condition that the hot-spot temperature does not exceed a prescribed
value. In other words, the ambient temperatures lower than the reference ambient
temperature may offer planned loading of transformers beyond nameplate rating.

The method can be afterward reduced to a simplified relationship between the loading
capability and the ambient temperature. It has been shown that the dependency of
122 8 Conclusions and Future Works

the loading capability on the ambient temperature can be described accurately by a

quadratic equation. On the contrary, for indoor transformers, a linear relationship
between the loading capability and the ambient and station temperatures contains
every information that is necessary for determination of the loading capability.

8.2 Recommendations for Future Works

The main focus of this thesis was to propose a dynamic top-oil temperature model for
use in an online monitoring system to determine the loading capability and to monitor the
condition of the cooling system. The proposed model should use the least possible input
data and be almost independent of the transformer under study and at the same time accurate
enough for these purposes. These conditions lead to the limitation of accuracy of the model.
The following recommendations may increase the reliability and accuracy of the proposed
model and algorithms.

• As formerly mentioned in section 2.3.3, the external factors (e.g. sun intensity) may
affect the accuracy of the thermal model. Although there are some computational and
simple ways [16,72–75,77] to take the sun intensity into account, they do not account
for the current environmental conditions at the site of transformers. The best practice
can be to use sun intensity measurement sensors and include the measured values into
the top-oil temperature model as an independent input. The same expression is also
valid for the effects of wind and its direction on the thermal behavior of transformers.

• It is worthwhile to investigate the thermal behavior of water-cooled transformers with

consideration of the water flow rate and water temperature to control the temperature
of the oil.

• The position of the tap-changer affects the short circuit loss and consequently the
total losses of a transformer. Accounting for tap-changer position as an additional
input may increase the accuracy of the top-oil temperature model. Although different
approaches have been proposed [79, 81], no significant results for online monitoring
purposes have been reported so far.

• In this thesis, only one set of empirical and design-dependent factors have been used
in the proposed model since the model is a physical model. However, for semi-
physical models such as the conventional models, different sets of unknown parame-
ters may be used in the thermal model which may increase the accuracy of the con-
ventional models significantly. However, the optimal number of unknown parameters
8.2 Recommendations for Future Works 123

sets and a compromise between the accuracy and the computational efforts should be
investigated.

• For monitoring of cooling system, other algorithms may be proposed which can over-
come the deficiencies of the proposed algorithms in this dissertation. These deficien-
cies are namely that the first algorithm can not be used during parametrization phase
while the second algorithm can not detect the malfunction of pumps.

• The algorithms for monitoring of cooling systems in this dissertation have been val-
idated with the detection of the fans failures. It is worthwhile to validate the perfor-
mance of the algorithms through different case studies to detect the failure of pumps
and speed reduction of fans due to contamination. However, the measured data for
these types of failures are not available easily. In these cases, the performance of
the algorithms can be further validated by applying artificial failures of these types
on an in-service power transformer and measurements during the faulty and healthy
operations.

• In order to strengthen the methods proposed in this dissertation for the determination
of the loading capability, it will be helpful to include other inputs into the top-oil
temperature model. These inputs may include the amount of moisture and gases in
the oil since they will limit the loading capability of transformers.

• A coupled thermal monitoring of transformers, bushings, and busbars may lead to a

more realistic determination of the loading capability of transformers.
 125

Bibliography

[1] M. Wang, A. J. Vandermaar, and K. D. Srivastava, “Review of condition assessment

of power transformers in service,” IEEE Electrical Insulation Magazine, vol. 18,
no. 6, pp. 12–25, November/December 2002.

[2] H. A. Aldhubaib and M. M. A. Salama, “A novel approach to investigate the effect

of maintenance on the replacement time for transformers,” IEEE Transactions on
Power Delivery, vol. 29, no. 4, pp. 1603–1612, August 2014.

[3] S. Tenbohlen, T. Stirl, and M. Roesner, “Benefit of sensors for on-line monitoring
systems for power transformers,” in Matpost, 2003.

[4] S. Tenbohlen, J. Jagers, G. Bastos, D. Desai, and et al., “Transformer reliability

survey,” Cigre Working Group A2.37, Tech. Rep. 642, December 2015.

[5] A. Bossi, J. Dind, J. Frisson, U. Khoudiakov, H. F. Light, D. V. Narke, Y. Tournier,

and J. Verdon, “An international survey on failures in large power transformers in
service,” Electra, no. 88, pp. 21–48, May 1983.

[6] J. Lapworth, P. Picher, J. Channet, and et al., “Transformer thermal modelling,” Cigre
Working Group A2.38, Tech. Rep. 659, 2016.

[7] S. Tenbohlen, T. Stirl, and M. Stach, “Assessment of overload capacity of power

transformers by on-line monitoring systems,” in IEEE Power Engineering Society
Winter Meeting, vol. 1, 2001, pp. 329–334.

[8] IEC 60076-7: Power Transformers- Loading Guide for Oil-Immersed Power Trans-
fomators, International Electrotechnical Commission (IEC) Std., 2005.

[9] IEEE std C57.91: IEEE Guide for Loading Mineral-Oil- Immersed Transformers,
American National Standard Std., 1995.

[10] B. C. Leisture, W. H. Hagman, and J. L. Kirtley, “An improved transformer top oil
temperature model for use in an on-line monitoring and diagnostic system,” IEEE
Transactions on Power Delivery, vol. 12, no. 1, pp. 249–256, January 1997.
126 Bibliography

[11] D. Susa, M. Lehtonen, and H. Nordman, “Dynamic thermal modelling of power

transformers,” IEEE Transactions on Power Delivery, vol. 20, no. 1, pp. 197–204,
January 2005.

[12] G. Swift, T. Molinski, R. Bray, and et al., “A fundamental approach to transformer

thermal modeling. part i: Theory and equivalent circuit,” IEEE Transactions on
Power Delivery, vol. 16, no. 2, pp. 171– 175, April 2001.

[13] G. Swift, T. Molinski, R. Bray, and et al., “A fundamental approach to transformer

thermal modeling. part ii: Field verification,” IEEE Transactions on Power Delivery,
vol. 16, no. 2, pp. 176–180, April 2001.

[14] J. F. Lindsay, “Temperature rise of an oil-filled transformer with varying load,” IEEE
Transactions on Power Apparatus and Systems, vol. PAS-103, no. 9, pp. 2530–2536,
September 1984.

[15] D. Kalic, Z. Radakovic, Z. Lazarevic, and et al., “On the determination of charac-
teristic temperatures in power oil transformers during transient states,” Archiv fuer
Elektrotechnik, vol. 76, pp. 457–468, 1993.

[16] G. L. Alegi and W. Z. Black, “Real-time thermal model for an oil-immeresed forced-
air coolded transformer,” IEEE Transactions on Power Delivery, vol. 5, no. 2, pp.
991–999, April 1990.

[17] A. Skillen, A. Revell, H. Iacovides, and et al., “Numerical prediction of local hot-
spot phenomena in transformer windings,” Applied Thermal Engineering, vol. 36,
pp. 69–105, 2012.

[18] N. E. Wakil, N. C. Chereches, and J. Padet, “Numerical study of heat transfer

and fluid flow in a power transformer,” International Journal of Thermal Sciences,
vol. 45, no. 6, pp. 645–626, June 2006.

[19] M. A. Tsili, E. I. Amoiralis, A. G. Kladas, and et al., “Power transformer thermal

analysis by using an advanced coupled 3d heat transfer and fluid flow fem model,”
International Journal of Thermal Sciences, vol. 53, pp. 188–201, March 2012.

[20] S. Tenbohlen, N. Schmidt, C. Breuer, S. Khandan, and R. Lebreton, “Investigation of

thermal behavior of an oil directed cooled transformer winding,” IEEE Transactions
On Power Delivery, vol. PP, no. 99, 2017.

[21] IEC 60354- Loading Guide for Oil-Immeresed Power Transformers, International
Elechtrotechnical commission Std., 1991.
Bibliography 127

[22] IEC 60076-2: Power Transformers- Temperature rise, International Electrotechnical

Commission (IEC) Std., 1993.

[23] Q. He, J. Si, and D. Tylavsky, “Prediction of top-oil temperature for transformers
using neural networks,” IEEE Transactions on Power Delivery, vol. 15, no. 4, pp.
1205–1511, October 2000.

[24] M. Hell, P. J. Costa, and F. Gomido, “Recurrent neurofuzzy network in thermal mod-
eling of power transformers,” IEEE Transactions on Power Delivery, vol. 22, no. 2,
pp. 904–910, April 2007.

[25] M. Hell, P. J. Costa, and F. Gomido, “Participatory learning in power transformers

thermal modeling,” IEEE Transactions on Power Delivery, vol. 23, no. 4, pp. 2058–
2067, October 2008.

[26] L. M. Souza, A. P. Lemos, W. M. Caminhas, and W. C. Boaventura, “Thermal model-

ing of power transformers using evolving fuzzy systems,” Engineering Applications
of Artificial Intelligence, vol. 25, no. 5, pp. 980–988, August 2012.

[27] M. Djamali and S. Tenbohlen, “Malfunction detection of the cooling system in air-
forced power transformers using online thermal monitoring,” IEEE Transactions on
Power Delivery, vol. 32, no. 2, pp. 1058–1067, April 2017.

[28] D. Susa and M. Lehtonen, “Dynamic thermal modeling of power transformers: Fur-
ther development-part i,” IEEE Transactions on Power Delivery, vol. 21, no. 4, pp.
1961–1970, October 2006.

[29] Z. Radakovic, D. Jacic, J. Lukic, and S. Milosavljevic, “Loading of transformers in

conditions of controlled cooling system,” International Transactions on electrical
Energy Systems, vol. 24, no. 2, pp. 203–214, 2014.

[30] W. H. Tang, Q. H. Wu, and Z. J. Richardson, “Equivalent heat circuit based power
transformer thermal model,” IEE Proceedings - Electric Power Applications, vol.
149, no. 2, pp. 87–92, March 2002.

[31] R. Goldschmidt, “Temperature curves and the rating of electrical machinery,” Jour-
nal of the Institution of Electrical Engineers, vol. 34, no. 172, pp. 660–691, May
1905.

[32] W. H. Cooney, “Predetermination of self-cooled oil-immersed transformer tempera-

tures before conditions are constant,” Transactions of the American Institute of Elec-
trical Engineering, vol. 44, pp. 611–618, January 1925.
128 Bibliography

[33] V. M. Montsinger, “Effect of load factor on operation of power trasnformers by tem-

perature,” Transactions of the American Institute of Electrical Engineering, vol. 59,
no. 11, pp. 632–636, November 1940.

[34] V. M. Montsinger, “Effect of barometric pressure on temperature rise of self-cooled

stationary induction apparatus,” Transactions of the American Institute of Electrical
Engineering, vol. 35, no. 1, pp. 599–633, January 1916.

[35] V. M. Montsinger, “Loading transformers by temperature,” Transactions of the Amer-

ican Institute of Electrical Engineering, vol. 49, no. 2, pp. 776–790, April 1930.

[36] A. E. Kennelly, “The thermal time constants of dynamo-electric machines,” Ameri-

can Institute of Electrical Engineers, vol. 44, no. 2, pp. 142–149, February 1925.

[37] The British Electrical and Allied Industries Research Association, “Thermal transfer-
ence in transformer oils,” Journal of the Institution of Electrical Engineers, vol. 67,
no. 388, pp. 527–537, April 1929.

[38] V. M. Montsinger and W. H. Cooney, “Temperature rise of stationary electrical ap-

paratus as influenced by radiation, convection and altitude,” American Institute of
Electrical Engineers, vol. 43, no. 9, pp. 803–812, September 1924.

[39] E. T. Noris, “The thermal rating of transformers,” Journal of the Institution of Elec-
trical Engineers, vol. 66, no. 380, pp. 841–854, August 1928.

[40] IEEE std. C57.91.81: IEEE guide for loading mineral oil immersed overhead and
pad-mounted distribution transfomators rated 500kVA and less with 65°C or 55°C
average winding , American National Standrad Std., 1981.

[41] H. Nordman, N. Rafsback, and D. Susa, “Temperature responses to step changes

in the load current of power transformers,” IEEE Transactions On Power Delivery,
vol. 18, no. 1, pp. 107–112, January 2003.

[42] D. Tylavsky, Q. He, G. A. Mcculla, and J. R. Hunt, “Sources of error in substation

distribution transformer dynamic thermal modeling,” IEEE Transactions on Power
Delivery, vol. 15, no. 1, pp. 178– 185, January 2000.

[43] D. Tylavsky, X. Mao, and G. A. McCulla, “Transformer thermal modeling: Improv-

ing reliability using data quality control,” IEEE Transactions on Power Delivery,
vol. 21, no. 3, pp. 1357–1366, July 2006.

[44] L. Jauregui-Rivera and D. Tylavsky, “Acceptability of four transformer top-oil

thermal models-part i: Defining metrics,” IEEE Transactions On Power Delivery,
vol. 23, no. 2, pp. 860–865, April 2008.
Bibliography 129

[45] D. Susa, J. Palola, M. Lehtonen, and M. Hyvaerinen, “Temperature rises in an ofaf

transformer at ofan cooling mode in service,” IEEE Transactions on Power Delivery,
vol. 20, no. 4, pp. 2517–2525, October 2005.

[46] D. Susa and M. Lehtonen, “Dynamic thermal modeling of power transformers: Fur-
ther development-part ii,” IEEE Transactions on Power Delivery, vol. 21, no. 4, pp.
1971–1980, October 2006.

[47] G. Bach, “Ueber die erwaermung des n-koerper-systems,” Archiv fuer Elektrotech-
nik, vol. 11, pp. 749–760, 1933.

[48] Z. Radakovic and K. Feser, “A new method for the calculation of the hot-spot tem-
perature in power transformers with onan cooling,” IEEE Transactions on Power
Delivery, vol. 18, no. 4, pp. 1284–1292, October 2003.

[49] W. H. Tang, Q. H. Wu, and Z. J. Richardson, “A simplified transformer thermal

model based on thermal-electric analogy,” IEEE Transactions on Power Delivery,
vol. 19, no. 3, pp. 1112–1119, July 2004.

[50] Z. Radakovic and S. Maksimovic, “Non-stationary thermal model of indoor trans-

former stations,” Electrical Engineering, vol. 84, pp. 109–117, 2002.

[51] M. Z. Degfa, R. J. Millar, M. Lehtonen, and O. Hyvoenen, “Dynamic thermal mod-

eling of mv/lv prefabricated substations,” IEEE Transactions on Power Delivery,
vol. 29, no. 2, pp. 786–793, April 2014.

[52] A. M. Lockie and D. K. Whirlow, “Thermal performance of distribution transformers

in underground vaults i-the problem,” IEEE Transactions on Power Apparatus and
Systems, vol. PAS-87, no. 9, pp. 1741–1745, September 1968.

[53] D. K. Whirlow and A. M. Lockie, “Thermal performance of distribution transform-

ers in underground vaults: Ii - mathematical model,” IEEE Transactions on Power
Apparatus and Systems, vol. PAS-87, no. 9, pp. 1745–1754, September 1968.

[54] J. P. A. Sandraz, F. de Leon, and J. Cultrera, “Validated transient heat-transfer model

for underground transformer in rectangular vault,” IEEE Transactions On Power De-
livery, vol. 28, no. 3, pp. 1770–1779, July 2013.

[55] D. Susa and M. Lehtonen, “Iec 60076Ű7 loading guide thermal model constants
estimation,” European Transactions On Electrical Power, vol. 23, no. 7, pp. 946–
960, February 2012.
130 Bibliography

[56] R. Vilaithong, S. Tenbohlen, and T. Stirl, “Improved top-oil temperature model for
unsteady-state conditions of power transformers,” in 14th International Symposium
on High Voltage Enginnering, 2005.

[57] Guides for Operation of Transformers, Regulators, and Reactors, AIEE Transformer
Subcommittee Std., 1945.

[58] IEEE Guide for Loading Mineral- Oil-Immersed Transformers and Step-Voltage
Regulators, IEEE Power & Energy Society Std., 2011.

[59] C. Fortescue, “Correction on transformer temperatures for variation in room temper-

ature taking into account both copper and iron losses,” American Institute of Electri-
cal Engineers, vol. 32, no. 2, pp. 607–616, February 1913.

[60] J. Aubin and Y. Langhame, “Effect of oil viscosity on transformer loading capability
at low ambient temperature,” IEEE Transactions on Power Delivery, vol. 7, no. 2,
pp. 516–524, April 1992.

[61] M. V. Thaden, P. Mehta, S. Tuli, and R. L. Grubb, “Temperature rise tests on a

forced-oil-air cooled (foa) (ofaf) core-form transformer, including loading beyond
nameplate,” IEEE Transactions on Power Delivery, vol. 10, no. 2, pp. 913–923, April
1995.

[62] J. Aubin, R. Bergeron, and R. Morin, “Distribution transformer overloading capa-

bility under cold-load pickup conditions,” IEEE Transactions on Power Delivery,
vol. 5, no. 4, pp. 1883–1891, October 1990.

[63] R. L. Grubb, H. Hudis, and A. R. Traut, “A transformer thermal duct study of vari-
ous insulating fluids,” IEEE Transactions on Apparatus and Systems, vol. PAS-100,
no. 2, pp. 466–473, February 1981.

[64] Y. Langhame, J. Castonguay, and J. P. Crine, “Physical behavior of paraffinic oils at

low temperatures,” IEEE Transactions on Electrical Insulation, vol. EI-20, no. 3, pp.
629–638, 1985.

[65] D. Martin, J. Wijaya, N. Lelekakis, D. Susa, and N. Heyward, “Thermal analysis

of two transformers filled with different oils,” IEEE Electrical Insulation Magazine,
vol. 30, no. 1, pp. 39–45, January 2014.

[66] D. Kweon, K. Koo, J. Woo, and Y. Kim, “Hot spot temperature for 154 kv trans-
former filled with mineral oil and natural ester fluid,” IEEE Transactions on Di-
electrics and Electrical Insulation, vol. 19, no. 3, pp. 1013–1020, June 2012.
Bibliography 131

[67] H. Al-Amin, J. O’Brien, and M. Lashbrook, “Synthetic ester transformer fluid: A

total solution to windpark transformer technology,” Renewable Energy, vol. 49, pp.
33–38, January 2013.

[68] A. E. Emanuel and M. Wang, “Estimation of loss of life of power transformers sup-
plying nonlinear loads,” IEEE Transactions on Power Apparatus and Systems, vol.
PAS-104, no. 3, pp. 628–636, 1985.

[69] IEEE std C57.110 IEEE Recommended Practice for Establishing Liquid-Filled and
Dry-Type Power and Distribution Transformer Capability When Supplying Nonsinu-
soidal Load Currents, IEEE Std., 2008.

[70] H. Turker, D. Chatroux, and A. Halby, “Low-voltage transformer loss-of-life as-

sessments for a high penetration of plug-in hybrid electric vehicles (phevs),” IEEE
Transactions On Power Delivery, vol. 27, no. 3, pp. 1323–1331, July 2012.

[71] IEEE C57.119: IEEE Recommended Practice for Performing Temperature Rise Tests
on Oil-Immersed Power Transformers at Loads Beyond Nameplate Ratings, Power
Engineering Society Std., March 2002.

[72] D. Tylavsky, Q. He, J. Si, G. A. Mcculla, and J. R. Hunt, “Transformer top-oil

temperature modeling and simulation,” IEEE Transcations on Industry Applications,
vol. 36, no. 5, pp. 1219–1225, September/ October 2000.

[73] B. Gorgan, P. V. Notingher, J. M. Wetzer, H. F. A. Verhaart, P. A. A. F. Wouters,

and A. van Schijndel, “Influence of solar irradiation on power transformer thermal
balance,” IEEE Transactions on Dielectrics and Electrical Insulation, vol. 19, no. 6,
pp. 1843–1851, December 2012.

[74] E. Hajidavalloo and M. Mohamadianfard, “Effect of sun radiation on the thermal be-
havior of distribution transformer,” Applied Thermal Engineering, vol. 30, pp. 1133–
1139, 2010.

[75] K. Karsai, D. Kerenyi, and L. Kiss, Large Power Transformers. Elsevier, 1987.

[76] V. M. Montsinger and I. Wetherill, “Effect of color of tank on the temperature of

self-cooled transformers under service conditions,” Transactions of the American
Institute of Electrical Engineering, vol. 49, no. 1, pp. 41–50, 1930.

[77] A. E. B. Abu-Elanien and M. M. A. Salama, “A monte carlo approach for calculating

the thermal lifetime of transformer insulation,” Electrical Power and Energy Systems,
vol. 43, no. 1, pp. 481–487, December 2012.
132 Bibliography

[78] Y. A. Cengel, Heat Transfer: A Practical Approach, ser. Mechanical engineering.

McGraw-Hill, 2003.

[79] Z. Radakovic and S. Tenbohlen, “Thermal model of oil power transformers with a tap
changer,” Turkish Journal of Electrical Engineering & Computer Sciences, vol. 24,
pp. 3293–3308, April 2015.

[80] Z. Radakovic and M. Sorgic, “Basics of detailed thermal-hydraulic model for thermal
design of oil power transformers,” Ieee Transactions On Power Delivery, vol. 25,
no. 2, pp. 790–802, April 2010.

[81] M. Djamali and S. Tenbohlen, “A dynamic top oil temperature model for power
transformers with consideration of the tap changer position,” in The 19th Interna-
tional Symposium on High Voltage Engineering (ISH), Pilsen, 2015.

[82] V. V. Vakhnina, V. A. Shapovalov, V. N. Kuznetsov, and D. A. Kretov, “The influence

of geomagetic stroms on the thermal processes in the tank of a power transformer,”
IEEE Transactions on Power Delivery, vol. 30, no. 4, pp. 1702 – 1707, August 2015.

[83] L. Marti, A. Rezaei-Zare, and A. Narang, “Simulation of transformer hotspot heating

due to geomagnetically induced currents,” IEEE Transactions on Power Delivery,
vol. 28, no. 1, pp. 320–327, January 2013.

[84] R. S. Girgis and K. B. Vedante, “Impact of gics on power transformers: Overheat-

ing is not the real issue.” IEEE Electrification Magazine, vol. 3, no. 4, pp. 8–12,
December 2015.

[85] M. Beltle and S. Tenbohlen, “Power transformer diagnosis based on mechanical os-
cillations due to ac and dc currents,” IEEE Transactions on Dielectrics and Electrical
Insulation, vol. 23, no. 3, pp. 1515–1522, June 2016.

[86] S. Tenbohlen, S. Coenen, M. Djamali, A. Müller, M. H. Samimi, and M. Siegel,

“Diagnostic measurements for power transformers,” Energies, vol. 9, no. 5, p. 347,
2016.

[87] R. Linan, D. Ponce, E. Betancourt, and G. Tamez, “Optimized models for over-
load monitoring of power transformers in real time moisture migration model,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 20, no. 6, pp. 1977–1983,
December 2013.

[88] Y. Cui, H. Ma, T. Saha, C. Ekanayake, and D. Martin, “Moisture dependent thermal
modelling of power transformer,” IEEE Transactions on Power Delivery, vol. 31,
no. 5, pp. 2140–2150, May 2016.
Bibliography 133

[89] L. W. Pierce, “Predicting liquid filled transformer loading capability,” IEEE Trans-
actions on Industry Applications, vol. 30, no. 1, pp. 170–178, 1994.

[90] Cigre working group 09 of study committee 12, “Heat-run test procedure for power
transformers,” Electra, vol. 129, pp. 37–48, 1990.

[91] D. Feng, Z. Wang, and P. Jarman, “Evaluation of power transformers effective hot-
spot factors by thermal modeling of scrapped units,” IEEE Transactions on Power
Delivery, vol. 29, no. 5, pp. 2077–2085, October 2014.

[92] Z. Radakovic, U. Radoman, and P. Kostic, “Decomposition of the hot-spot factor,”

IEEE Transactions On Power Delivery, vol. 30, no. 1, pp. 403–411, February 2015.

[93] A. M. Emsley and G. C. Stevens, “Review of chemical indicators of degradation of

cellulosic electrical paper insulation in oil-filled transformers,” IEE Proceedings -
Science, Measurement and Technology, vol. 141, no. 5, pp. 324–334, 1994.

[94] A. M. Emsley, X. Xiao, R. J. Heywood, and M. Ali, “Degradation of cellulosic in-

sulation in power transformers. part 3: Effects of oxygen and water on ageing in
oil,” IEE Proceedings - Science, Measurement and Technology, vol. 147, no. 3, pp.
115–119, 2000.

[95] D. Martin, Y. Cui, C. Ekanayake, H. Ma, and T. Saha, “An updated model to de-
termine the life remaining of transformer insulation,” IEEE Transactions on Power
Delivery, vol. 30, no. 1, pp. 395–702, February 2015.

[96] N. Lelekakis, D. Martin, and J. Wijaya, “Ageing rate of paper insulation used in
power transformers part 1: oil/paper system with low oxygen concentration,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 19, no. 6, pp. 1999–2008,
December 2012.

[97] N. Lelekakis, D. Martin, and J. Wijaya, “Ageing rate of paper insulation used in
power transformers part 2: Oil/paper system with medium and high oxygen concen-
tration,” IEEE Transactions on Dielectrics and Electrical Insulation, vol. 19, no. 6,
pp. 1999–2008, December 2012.

[98] L. E. Lundgaard, W. Hansen, D. Linhjell, and T. Painter, “Aging of oil-impregnated

paper in power transformers,” IEEE Transactions on Power Delivery, vol. 19, no. 1,
pp. 230–239, January 2004.

[99] M. Djamali and S. Tenbohlen, “Assessment of the dynamic overload capability of

distribution transformers,” in The 19th International Symposium on High Voltage
Engineering (ISH), 2015.
134 Bibliography

[100] AIEE Transformer subcommittee, “Interim report on guides for overloading trans-
formers and voltage regulators,” Transactions of the American Institue of Electrical
Engineering, vol. 61, no. 9, pp. 692–694, September 1942.

[101] S. Tenbohlen, T. Stirl, G. Bastos, and et al., “Experienced-based evaluation of eco-

nomic benefits of on-line monitoring systems for power transformers,” in Cigre Ses-
sion, Paris, 2002.

[102] W. Zhan, A. E. Goulart, M. Falahi, and R. Preethi, “Development of a low-cost self-

diagnostic module for oil-immerse forced-air cooling transformers,” IEEE Transac-
tions on Power Delivery, vol. 30, no. 1, pp. 129–137, February 2015.

[103] M. Djamali and S. Tenbohlen, “A validated online algorithm for detection of fan
failures in oil-immersed power transformers,” International Journal of Thermal Sci-
ences, vol. 116, no. C, pp. 224–233, June 2017.

[104] K. Qian, C. Zhou, and Y. Yuan, “Impacts of high penetration level of fully electric
vehicle charging loads on the thermal ageing of power transformers,” International
Journal of Electrical Power and Energy Systems, vol. 65, pp. 102–112, 2015.

[105] M. A. Awadallah, T. Xu, B. Venkatesh, and B. N. Singh, “On the effects of solar
panels on distribution transformers,” IEEE Transactions on Power Delivery, vol. 31,
no. 3, pp. 1176–1185, June 2016.

[106] S. F. Abdelsamad, W. G. Morsi, and T. S. Sidhu, “Probabilistic impact of transporta-

tion electrification on the loss-of-life of distribution transformers in the presence of
rooftop solar photovoltaic,” IEEE Transactions on Sustainable Energy, vol. 6, no. 4,
pp. 1565–1573, September 2015.

[107] A. D. Hilshey, P. D. H. Hines, P. Rezaei, and J. R. Dowds, “Estimating the impact

of electric vehicle smart charging on distribution transformer aging,” IEEE Transac-
tions on Smart Grid, vol. 4, no. 2, pp. 905–913, June 2013.

[108] Q. Gong, S. Midlam-Mohler, E. Serra, and V. Marano, “Pev charging control con-
sidering transformer life and experimental validation of a 25 kva distribution trans-
former,” IEEE Transactions on Smart Grid, vol. 6, no. 2, pp. 648–656, March 2015.

[109] K. R. Shah and D. P. Sekulic, Fundamental of Heat Exchanger Design. John Wiley
& Sons, INC., 2003.

[110] VDI, VDI Heat Atlas, 6th ed. Springer, 2006.

[111] A. Bejan and A. D. Kraus, Heat Transfer Handbook. John Wiley & Sons, INC.,
2003.
Bibliography 135

[112] D. Etheridge, Natural Ventilation of Buildings: Theory, Measurement and Design.

John Wiley & Sons Ltd, 2012.

[113] IEC 60529: Degrees of Protection provided by enclosures (IP Code), International
Electrotechnical Commission (IEC) Std., 2005.

[114] L. Jauregui-Rivera and D. Tylavsky, “Acceptability of four transformer top-oil ther-

mal models-part ii: Defining metrics,” IEEE Transactions On Power Delivery,
vol. 23, no. 2, pp. 866–872, April 2008.

[115] R. Vilaithong, “Models for thermal and mechanical monitoring of power trans-
former,” Ph.D. dissertation, University of Stuttgart, 2011.

[116] N. Schmidt, “Temperature prediction and estimation of noncritical overload in oil

cooled transformers based on simplified thermal modelling,” 2010, internal Presen-
tation at University of Stuttgart.

[117] R. L. Brown, J. Durbin, and J. M. Evans, “Techniques for testing the constancy of
regression relationships over time,” Journal of the Royal Statistical Society. Series B
(Methodological), vol. 37, no. 2, pp. 149–192, 1975.

[118] J. O. Rawling, S. G. Pantula, and D. A. Dickey, Applied Regression Analysis: A

Research Tool. Springer, 1998.

[119] J. S. Galpin and D. M. Hawkins, “The use of recursive residuals in checking model
fit in linear regression,” The American Statistician, vol. 38, no. 2, pp. 94–105, May
1984.

[120] B. Basseville and I. V. Nikiforov, Detection of abrupt changes: theory and applica-
tion. Prentice Hall, 1993.

[121] S. S. Wilks, “The large-sample distribution of the likelihood ratio for testing com-
posite hypotheses,” The Annals of Mathematical Statistics, vol. 9, no. 1, pp. 60–62,
March 1938.

[122] M. Humayun, M. Z. Degfa, A. Safdarian, and M. Lehtonen, “Utilization improve-

ment of transformers using demand response,” IEEE Transactions on Power Deliv-
ery, vol. 30, no. 1, pp. 202–210, Feb 2015.

[123] Z. Radakovic, M. Sorgic., W. V. der Veken, and G. Claessens, “Ratings of oil power
transformer in different cooling modes,” IEEE Transactions on Power Delivery,
vol. 27, no. 2, pp. 618–625, April 2012.

[124] G. A. F. Seber and C. J. Wild, Nonlinear Regression. John Wiley & Sons, INC.,
2003.
Appendix
A.1 Nonlinear Least-Squares Method for Parameter Estimation 139

A.1 Nonlinear Least-Squares Method for Parameter

Estimation

According to the nonlinear regression theory, n observations yielded based on measure-

ments using an online monitoring system (xi , yi ) from a fixed-regressor non-linear ther-
mal model with a known functional relationship f (equation (3.2)) are considered [124].
Hence:
Yi = f (Xi , φ∗ ) + ǫi , (A.1)

where Xi is the input matrix of the model measured by the online monitoring system dur-
ing normal operation of the transformer, Yi is the measured top-oil temperature, the true
value φ∗ of φ which is known to belong to Φ -a subset of real numbers ℜ- is the vector
of the parameters to be estimated (Table 3.1), and ǫi is the vector of the error between the
calculated and the measured top-oil temperature whose expected value is zero (E[ǫi ] = 0).
The least-squares estimation of φ∗ , denoted by φ̂, minimizes the sum of squares of the error
over φ ∈ Φ as follows [124]:
n h i2
X 2
S(φ̂) = Yi − f (Xi , φ̂) = Y − f (X, φ̂) . (A.2)
2
i=1

Gauss-Newton algorithm uses a Taylor series of f (Yi , φ̂) around the current estimation of
the parameters (φ(a) ). Using this approximation, the minimization problem described in
(A.2) is converted to a linear least-squares problem as follows [124]:
  2
(a) (a) (a)
min r −F · φ̂ − φ , (A.3)
2

where r (a) = Y − f (X, φ(a) ) and F (a) is the Jacobian matrix in the current iteration.
Substituting in (A.2) leads to estimation of the parameters in the next iteration as fol-
lows [124]:  ′
 −1 ′
φ(a+1) = φ(a) + F (a) · F (a) · F (a) · r (a) . (A.4)

This provides an iterative scheme for estimation of the parameters based on measured data.
This equation can also be used recursively. If a new input-output set (xi , yi ) is available,
the previously estimated parameters φ can be updated using (A.4) [124]. However, for
online monitoring purposes, due to the huge number of elements in F , the inversion of the
(Fi ′ · Fi ) may be time consuming and in some cases not possible; therefore, it should be
calculated recursively using the inverse matrix of the previous time steps as follows [117]:
140

−1 −1 (F ′ i−1 · Fi−1 )−1 · fi · f ′ i · (F ′ i−1 · Fi−1 )−1

(F ′ i · Fi ) = (F ′i−1 · Fi−1 ) + . (A.5)
1 + f ′ i · (F ′ i−1 · Fi−1 )−1 · fi
A.2 Thermal Properties of Oil and Air 141

A.2 Thermal Properties of Oil and Air

The thermal properties of oil, air used in this work are functions of the fluid temperature
listed in Table A.1 and A.2 respectively.

Table A.1: Temperature-dependent thermal properties of oil [63].

Property Used in Equation

Thermal conductivity (W/(m·K)) 3.8, 3.10 κoil (T ) = 0.124 − 1.525 · 10−4 · T
24.736 ·(T +273)−4.163
Kinematic viscosity (m /s)
2
3.10, 3.9,3.14 νoil (T ) = (ee − 0.7) · 10−6
Fluid density (kg/m )
3
3.10 ρoil (T ) = 887 − 0.659 · T
Specific heat (W·s/(kg·K)) 3.10 cpoil (T ) = 1960 + 4.005 · T
Volumetric thermal expansion coefficient (1/K) 3.9 βoil = 8.6 · 10−4

Table A.2: Temperature-dependent thermal properties of air [110].

Property Used in Equation

Thermal conductivity (W/(m·K)) 3.21, 3.23 κair (T ) = 0.02382 + 6.478 · 10−5 · T
Kinematic viscosity (m /s)
2
3.23, 3.22, 3.26 νair (T ) = (2.25 · T 2 + 0.10 · T + 13.68) · 10−6
Fluid density (kg/m )
3
3.23 ρair (T ) = 1.094 · 10−5 · T 2 − 0.004 · T + 1.278
Specific heat (W·s/(kg·K)) 3.23 cpair (T ) = 4.12 · 10−4 · T 2 + 0.02 · T + 1006
Volumetric thermal expansion coefficient (1/K) 3.22 βair (T ) = 0.004 − 1.3 · 10−5 · T
 142

A.3 Publications

• M. Djamali, S. Tenbohlen, E. Junge, M. Konermann; ”Real-Time Evaluation of the

Dynamic Loading Capability of Indoor Distribution Transformers”; Accepted for
publication in IEEE Transactions on Power Delivery.

• M. Djamali, S. Tenbohlen; ”Hundred Years of Experiences in Dynamic Thermal

Modeling of Power Transformers”; IET Generation, Transmission & Distribution,
Vol.11, Issue. 11, August 2017.

• M. Djamali, S. Tenbohlen; ”A Validated Online Algorithm for Detection of Fan Fail-

ures in Oil-Immersed Power Transformers”; International Journal of Thermal Sci-
ences, 116C (2017) pp. 224-233.

• M. Djamali, S. Tenbohlen; ”Malfunction Detection of the Cooling System in Air-

Forced Power Transformers Using Online Thermal Monitoring”; IEEE Transactions
on Power Delivery, vol.32, No.2, pp. 1058-1067, April 2017.

• S. Tenbohlen, M. Djamali; ”Assessment of Overload Capability of Power Transform-

ers by Online Monitoring”; Cigré VIII Workspot; Recife 2016.

• S. Tenbohlen, S. Coenen, M. Djamali, A. Müller, M. H. Samimi, M. Siegel; ”Diag-

nostic Measurements for Power Transformers”; Energies 2016, 9(5).

• S. Tenbohlen, M. Djamali; ”A Dynamic Top-Oil Temperature Model for Online As-

sessment of Overload Capability of Power Transformers”; Cigré SC A2 Colloquium,
Shanghai 2015.

• M. Djamali, S. Tenbohlen; ”A Dynamic Top Oil Temperature Model for Power Trans-
formers with Consideration of the Tap Changer Position”; 19th International Sympo-
sium on high Voltage Engineering (ISH); Pilsen 2015.

• M. Djamali, S. Tenbohlen, G. Walker; ”Assessment of the Dynamic Overload Capa-

bility of Distribution Transformers”; 19th International Symposium on high Voltage
Engineering (ISH); Pilsen 2015.

• M. Djamali, S. Tenbohlen; ”A New Top-Oil Temperature Model based on IEC60076-

7 for OFAF and ONAF Transformers”; VDE-ETG-Fachtagung Diagnostik elek-
trischer Betriebsmittel; Berlin 2014.

• M. Djamali, S. Tenbohlen; ”A Priori Metrics to Select Best Training Dataset of Top-

Oil Temperature Models of Power Transformers”; International Conference on High
Voltage Engineering and Application (ICHVE); Poznan 2014.

View publication stats