IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 22, NO.

1, JANUARY 2007

Estimation and Measurement of Junction

Temperatures in a Three-Level
Voltage Source Converter
Thomas Brckner, Member, IEEE, and Steffen Bernet, Member, IEEE

AbstractThe design of a power converter must guarantee that

of all devices do not exceed
the operating junction temperatures
their limits under all specified operating conditions. Usually, this
is ensured by a simulative or analytical junction temperature estimation based on simple electrical and thermal models and semiconductor datasheet values. This paper discusses the difficulties
and quantifies the limitations of this approach on the example of
a three-level neutral point clamped voltage source converter (NPC
VSC) with insulated gate bipolar transistors. The calculations are
compared to the results of direct junction temperature measurements with an infrared camera. The paper also provides the experimental proof for the unequal loss and junction temperature
distribution in the three-level NPC VSC.
Index TermsInsulated gate bipolar transistors (IGBT), Junction temperature measurement, neutral point clamped voltage
source converter (NPC VSC), thermal modeling.

I. INTRODUCTION
OR the design of a power converter the maximum electrical
ratings and the thermal limitations of the semiconductor
devices play an equally important role. It must be guaranteed
of all devices do
that the operating junction temperatures
not exceed their limits under all specified operating conditions.
is an integral part of the conTherefore, the calculation of
verter design procedure. Nowadays, some semiconductor manufacturers provide online (internet) software tools for this purpose [1], [2]. Another recent development is the real-time calculation of the operating junction temperatures to prevent thermal
overload [3], [4].
Not only in these two cases but also in the standard power
calculation is commonly based on
converter design, the
datasheet values for both, the electrical behavior (conduction and switching losses) and the thermal characteristics
of the semiconductors. The level of the electrical modeling
varies from linear to third-order approximations. The semiconductors thermal characteristics are typically described by
component models [5]. While
one-dimensional (1-D)
the electrical data can be verified by the means of uncomplicated measurements, thermal device models are more difficult

Manuscript received November 3, 2005; revised March 29, 2006. Recommended for publication by Associate Editor J. A. Ferreira.
T. Brckner is with the Converteam GmbH, Berlin 12277 Germany (e-mail:
thomas.brueckner@gmx.net).
Steffen Bernet is with the Institute of Energy and Automation Technology, Technical University of Berlin, Berlin 10587 Germany (e-mail:
steffen.bernet@tu-berlin.de).
Digital Object Identifier 10.1109/TPEL.2006.886651

are obtained either by

to validate. Their parameters
complex physics-based finite element method (FEM) device
models or by the measurement of via temperature-dependent
electrical semiconductor characteristics in special test circuits
[6][8].
calculation based on these simple electrical
Though the
and thermal models is widely used, nearly no data are available
quantifying its accuracy and its limitations en bloc. Attempts
to experimentally verify the calculated junction temperatures
in a real converter under practical operating conditions are not
known to the authors.
Therefore, the common method for junction temperature calculation is analyzed in this paper. The approach is applied to a
three-level neutral point clamped voltage source converter (NPC
VSC) with insulated gate bipolar transistors (IGBTs) and the
estimated temperatures are compared to junction temperatures
directly measured in this converter. Thus, the difficulties, the
accuracy, and the limitations of the junction temperature estimation method as a whole can be evaluated.
A schematic of the three-level NPC VSC is shown in Fig. 1.
The topology has been chosen as an example, because its thermal
design is more complicated than that of two-level converters.
The NPC VSC displays an inherently uneven loss distribution
amongst its semiconductor switches, dependent on the specific
operating points. The most unbalanced distribution can be
found at the boundaries of the four-quadrant operating area, i.e.,
1.155 and minimum modulation depth
at maximum
0 and power factors
1 (inverter operation) and
1 (rectifier operation). In each of these cases one group of
devices experiences both, significant switching and conduction
losses, while all other devices are substantially less stressed
[9]. The paper provides the experimental proof for the loss and
junction temperature distribution in the NPC VSC. Moreover,
the junction temperature measurements described in this paper
are also used in [10] to validate the loss-balancing principle of
the NPC VSC featuring active NPC switches (ANPC VSC).
In detail, a low-voltage (400 V, 10 kVA) test setup was constructed whose function and design is described in Section II of
the paper. The electrical and thermal modeling of the laboratory
setup is explained in Section III. The difficult task of the junction temperature measurements was performed with the aid of
an infrared camera system on open IGBT modules as described
in Section IV. Following, the results of calculation and measurement are compared and evaluated in Section V. Special attention
is given to the accuracy and the problems associated with both
junction temperature measurement and prediction (Section VI).

0885-8993/$20.00 2006 IEEE

IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 22, NO. 1, JANUARY 2007

TABLE I
RATINGS OF LABORATORY TEST SETUP

Fig. 1. Three-level NPC VSC.

Fig. 2. Schematic of test setup with two phase legs used as

H -bridge converter.

II. LOW-VOLTAGE LABORATORY SETUP

To investigate the junction temperature distribution of the
three-level NPC VSC, a low-voltage test bench was set up. It
is configured as three-phase acdcac converter with 12-pulse
diode-bridge rectifier and three-level inverter. A schematic of
the inverter is shown in Fig. 2. The basic converter data and ratings are summarized in Table I. One phase leg of the inverter
is assembled of specially prepared open IGBT modules to enable the infrared junction temperature measurements. The open
modules are directly taken out of the production process prior to
the encapsulation and the filling with silicon gel. A second phase
leg is equipped with high-frequency Pearson current transducers
to measure the IGBT switching transients.
200 V 690 V) the
In low-voltage IGBT converters (
devices are usually installed on a common heatsink or a higher
number of devices, e.g., a complete three-phase two-level converter, is integrated into one power module. In this situation,
is the
the common base temperature for the calculation of
heatsink or base plate temperature. The devices are thermally
strongly coupled. In contrast, in high-power medium-voltage
converters discrete devices are installed on separate heatsinks.
The devices are thermally coupled by the cooling water or ambient air only. Hence, the differences between the device junction temperatures can be larger than in a low-voltage converter.
An unequal loss distribution amongst the devices also yields an
unequal junction temperature distribution.
Since the three-level NPC VSC is typically applied in
high-power medium-voltage applications and its thermal characteristics are more critical in this situation, several measures

are used to emulate the medium-voltage converters thermal

behavior in the low-voltage setup. Single IGBT modules are
not available for the low power rating of the test bench. Therefore, each phase leg is assembled of three half-bridge modules
as shown in Fig. 2, viz. T T T T , and T T .
For all experiments described in this paper, the IGBTs T
and T are not active and are permanently turned off by a
gate-emitter short.1 Their anti-parallel diodes D and D act
as NPC diodes. The arrangement of the modules guarantees
2 between
that none of them is stressed with more than
its terminals.
To thermally decouple the half-bridge modules from each
other, every module is installed on its own heatsink. Furthermore, it is desired to achieve a good thermal decoupling of the
two IGBTs in one module. For this purpose bipartite aluminum
blocks are inserted between the modules and the heatsinks.
Fig. 3 shows the mechanical layout. The aluminum blocks
introduce an additional thermal resistance between the junction
and the first point of a low-resistive common coupling, i.e., the
heatsink. The coupling through the module base plate can not
be avoided. The temperature in the middle of each heatsink is
measured by a thermocouple.
To enable a good cooling and highest compactness (necessary for the infrared measurements), the three heatsinks are integrated into one mechanical unit as depicted in Fig. 4. They
are mounted in series and are thermally insulated. HS-23 denotes the heatsink for the module with the IGBTs T and T ,
HS-15 for T and T , etc. The interaction between the three
heatsinks through the cooling air stream (heat convection) and
heat radiation remains.
For the thermal investigations, two phases of the test bench
were operated as four-quadrant -bridge converter with a
single-phase inductive load as indicated in Fig. 2. The load
resistance is the self resistance of the inductor. The phase
under test, i.e., the open phase, is voltage controlled by a
pure sine-triangle modulation with a given modulation depth
and fundamental frequency . The second phase
is used to generate the necessary countervoltage impressing a
and phase
sinusoidal current with the desired amplitude
into phase one. The required modulation depth for
shift
phase two and the phase shift of its fundamental voltage with
1The active NPC switches T
and T
for the ANPC VSC described in [10].

are utilized in the experiments

BRCKNER AND BERNET: ESTIMATION AND MEASUREMENT OF JUNCTION TEMPERATURES

TABLE II
OPERATING POINTS FOR SINGLE-PHASE OPERATION

area. Instead, the four representative operating points given in

Table II were selected for the investigation.
III. ESTIMATION OF JUNCTION TEMPERATURES
In an industrial converter design the thermal dimensioning
of the switches is typically based on the electrical and thermal
data given in the semiconductors datasheet, that are possibly complemented by measured switching and conduction
losses. The switching losses are approximated in terms of the
commutation voltage and current, the conduction losses in
terms of the on-state current. After an electrical simulation
allocating the switching transients and conduction periods to
the specific switches, the accumulated losses per device are
component
fed into a thermal-equivalent network (
model) to determine the junction temperatures. For a stationary
thermal design only the average junction temperatures are
are omitted. The
calculated and the thermal capacitances
junction temperature ripple with the fundamental frequency
does typically not exceed more than a few Kelvin for
5 Hz. An appropriate safety margin for the maximum operating junction temperature taking the ripple into account, e.g.,
5 10 K, is already included in the datasheet value.

Fig. 3. Mechanical layout of one half-bridge module with heatsink.

A. Electrical Modeling and Loss Calculation

Fig. 4. Photograph of heatsink unit.

respect to phase one denote to

(1)

(2)
with
and
Using this method, the stresses of a single phase of a threephase converter can be modeled over nearly the entire spectrum
of operating conditions. The theoretical limits are set by the
1. A further practical limmaximum countervoltage at
itation is invoked by the minimum/maximum pulse width constraint. For these reasons, the temperature distribution can not be
examined exactly at the most critical corners of the operating

The loss data given in datasheets are valid for a standard converter design and can not necessarily be applied to any specific topology. In the case of the low-voltage test bench the
datasheet values are not applicable for a variety of reasons. The
100 A/1200 V-IGBT modules are overrated in terms of voltage
and current carrying capability by a factor of two. Consequently,
no appropriate values are given in the datasheet, e.g., for a com325 V. Due to the drastic overrating
mutation voltage of
a sufficiently large junction temperature rise for a convincing interpretation can not be achieved with the recommended gate re15 . Therefore, to slow down the switching
sistance of
transients and to amplify the IGBT switching losses the gate re56 .
sistance is increased to
Moreover, commutations with different commutation loops
occur in the NPC VSC. For the commutation from to the
0, the outer switch
neutral tap at a positive load current
T is turned off and D takes over the current. Since only
one half-bridge module is involved in the transition (see Fig. 2),
this type of commutation is called short commutation. For the
0, the inner
commutation from the neutral tap to at
switch T is turned off and the diodes D and D take over
the current. The commutation loop involves all three half-bridge
modules of the phase leg. Thus, this commutation is called long
commutation. The higher stray inductance in the larger loop

IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 22, NO. 1, JANUARY 2007

TABLE III
MEASURED SWITCHING LOSSES

Fig. 5. Measured turnoff transient of T (V =2 = 325 V, i

= 40 A,
R
= 15
, time 250 ns/div, voltage 100 V/div, current 20 A/div).

Fig. 7. Measured IGBT and diode forward characteristics and approximations.

diode switching losses can be approximated with good precision by

(3)

Fig. 6. Measured turnoff transient of T

(conditions as Fig. 5).

causes higher turn-off and smaller turn-on losses for T compared to T .

The differences between the short and the long commutation were evaluated by switching loss measurements for both,
the recommended and the increased gate resistance. As an example, the current and voltage waveforms of the aforementioned
turn-off transients of T and T with the gate resistance
15 are shown in Figs. 5 and 6, respectively. From the magnitude of the voltage peak, the stray inductance in the commutation loops can be estimated to 180 nH for the short commutation circuit (turn-off of T ) and 360 nH for the long commutation circuit (turn-off of T ). Measured switching losses for
15 and
56 are given in Table III. The increased gate resistance roughly doubles the switching losses of
the IGBTs, while the diode recovery losses change only slightly.
At the same time, the loss differences between the short and long
commutation are lessened. Nevertheless, they are considered in
the model.
and turn-off losses
It is found that the turn-on
of the IGBTs display an almost linear dependence on the load
of the diodes are small comcurrent. The recovery losses
pared to the IGBT switching losses. Therefore, the IGBT and

where
stands for either
, or
, or
. The voltage
denotes the prospective commutation voltage of the deis the voltage at which the
vices in the NPC VSC and
is a fitting
device loss measurements were performed at.
constant. Since the loss measurements were carried out exactly
2, both
at the level of the commutation voltage
and
cancel from the equation.
Simple approximations for the IGBT and diode on-state parameters (threshold voltage and differential resistance) are given
in the datasheet. A series of on-state measurements shows that
these approximations deliver insufficient results for the rela0 30 A) of interest for the lowtively low currents (
voltage test setup. Therefore, the measured IGBT and diode conduction losses are approximated as
(4)
The measurements and their approximations are shown in Fig. 7.
(4) contains a logarithmic function of the on-state current modeling the behavior of a pn junction and a resistive share for
the drift region of the semiconductor. Though the parameters
, and
represent a physical meaning for the characteristic of a diode (temperature voltage, saturation current, and
ohmic resistance of drift region, respectively), here they are not
more than fitting constants. They were obtained by the means
of a standard curve fitting program.

BRCKNER AND BERNET: ESTIMATION AND MEASUREMENT OF JUNCTION TEMPERATURES

TABLE IV
PARAMETERS OF THERMAL MODEL

Fig. 8. Thermal model of one half-bridge module with heatsink.

The blocking losses of the IGBTs and diodes as well as the

diode turn-on losses are negligibly small and, thus, not considered in the model.
B. Thermal Modeling
The thermal model for one half-bridge module with heatsink
is depicted in Fig. 8. The purpose of the model is the calculation of the average junction temperatures. Thermal capacitances
are therefore not included. Compared to the usually applied
standard model the introduction of the aluminum blocks (see
Fig. 3) requires some refinements. Due to the increased resistance case-to-heatsink the coupling via the base plate must not
be disregarded. The network is extended by the transition resistance between the two switches on the base-plate level (case-tocase). In continuation of this concept also the heatsink is modeled in two parts which are linked by a heatsink-to-heatsink resistance. According to its actual position the thermocouple to
measure the mean heatsink temperature is placed in the middle
of this link. The two resistances case-to-heatsink include the
transition base plate-to-aluminum block, the resistance of the
block itself, and the transition aluminum block-to-heatsink.
The values of the thermal resistances are given in Table IV.
The case-to-case, heatsink-to-heatsink, and aluminum block resistances are calculated from the specific thermal conductivity
of the material and the geometrical data
(5)
and

(6)
(7)

is the average distance between the silicon dies in the

where
and
are the cross-sectional areas of the copper
module,
,
base plate and the heatsink along the mirror axis, and
are the height, length, and width of the aluminum
and
blocks.
401 W/(m K) and
209 W/(m K) denote

the specific thermal conductivity of copper and aluminum, respectively.

The resistances for junction-to-case (IGBT and diode) are
taken from the datasheet. For the transitions between base plate,
aluminum block, and heatsink, thermal contact grease is applied
and the case-to-heatsink value from the datasheet is used.
The interaction of the three heatsinks via the cooling air is
considered by temperature-controlled temperature sources. The
dominant process is the mutual heating by convection through
the forced air stream. Assume the lowest heatsink HS-23 (see
Fig. 4) to be at a high temperature level. The ambient air passing
through HS-23 is heated and the cooling air temperature for the
heatsinks HS-15 and HS-46 is increased. Therefore, a fraction of
the heatsink temperature of HS-23 must be added to the ambient
temperature for the latter heatsinks
(8)
In equivalence to (8) linear functions are formulated for all
mutual dependencies. The appropriate gains of the temperature
and the thermal resistances heatsink-to-air
sources
were obtained experimentally. The IGBTs of one module at a
time were connected to a dc-current source to produce defined
. The temperatures
of the three
constant losses
heatsinks are measured and their ratios are evaluated
and
(9)
(10)
is the temperature of the directly heated sink
and
is the increased temperature of the considered neighboring heatsink. The results of these measurements
are given in Table IV as well.
As expected the coupling occurs predominantly in the direction of the air flow, but also in the opposite direction some influence can be seen. Beside convection neighboring heatsinks
can affect each other also by radiation in both directions. Moreover, it is interesting to note that the investigation yields difthough all heatsinks are identical. The
ferent values for

IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 22, NO. 1, JANUARY 2007

Fig. 9. (a) Photograph of phase leg with open modules. (b) Infrared picture at operating point M
0.85; pf
1. (c) Infrared picture at operating point M
1; common conditions: ambient temperature #
25 C, V
650 V, I
30 A, f
8 kHz sine-triangle modulation.
0.15; pf

=0

first heatsink in the air stream is the best cooled; the last one the
worst cooled. This phenomenon is independent of the coupling
described above and can be explained by increasing air turbulences.
estimation based on the
The presented model allows the
(to be manually measured at the air
ambient air temperature
.
inlet) or based on the measured heatsink temperatures
To eliminate the potential error of the heatsink data, the heatsink
temperatures are taken as the basis throughout this work.
IV. MEASUREMENT OF JUNCTION TEMPERATURES
An overview of methods to determine semiconductor operating junction temperatures is given in [11] and [12]. The potential approaches include the indirect measurement via temperature-dependent electrical parameters and the direct measurement with contact or contact-less sensors.
Though the indirect measurement is often used for the validation of thermal device models [7], [8], the approach is not suited
for the application under standard converter operating conditions. Complicated changes in the control would be required to
include measurement cycles into the modulation scheme of a
real converter. Moreover, the tiny parameter variations, e.g., of
the on-state voltage drop in the range of a few millivolts per
Kelvin, can not be resolved in the electromagnetic environment
of an operated pulse-width modulation (PWM) converter. A direct measurement method is necessary that works independently
of the tested converter. The infrared temperature measurement
fulfils this requirement. It provides a strict separation of the converter and the measurement system, a high measurement speed,
and a high precision. Therefore, the authors decided for this
method.
The junction temperature measurements were performed
with an AGEMA Thermovision 900 professional infrared
imaging system [13]. The scanner camera with HgCdTe-detector operates at a frame rate of 15 Hz with a resolution of
136 272 pixels. A photograph of the open phase leg as seen

by the camera and two infrared images at different operating

points of the converter are shown in Fig. 9. The entire phase leg
with six IGBT and six diode chips is recorded in one camera
view. The three IGBT modules and the individual chips can be
clearly recognized in Fig. 9(b) and (c). The chip areas are defined in the processing software and the measured temperatures
are averaged over the individual chip areas.
A number of ambient measurement conditions have to be observed. The sensor and the lens temperature are measured by the
camera and are automatically taken into consideration for the
camera-internal data processing. The ambient air temperature,
the humidity, the distance object-to-camera, and the emissivity
of the measured object must be entered into a protocol. Since
the emissivity of the silicon dies is unknown and for metallic
surfaces is typically dependent on the surface angle, a coating
of the IGBT modules is inevitable. It is realized with a thin layer
0.95.
of black paint with a specified constant emissivity of
Since the camera operates in the spectral range of 8 12 m
(so-called atmospheric window with a transmission coefficient
1), the recombination radiation of
of the ambient air of
the silicon semiconductors at a wave length of
1.1 m
is filtered out. Note, that the recombination radiation does not
depend on the junction temperature but on the charge carrier
density.
Fig. 10 shows a macroscopic infrared image of one IGBT
chip. It can be seen that the bond wires conceal parts of the
chip surface and, thus, their temperature contributes to the measured average chip temperature. To quantify the effect, a series
of macroscopic IGBT and diode measurements for various conditions was carried out. It shows that the temperature of central
bond wires under stress is a few Kelvin higher than that of the
chips. However, the measured average temperature across the
chip surface (including the bond wires) is slightly lower than
the spot temperature in the center of the chip. The reason therefore is the lower temperature at the edges which are also partly
covered by passivation silicon gel. Considering both effects, a

BRCKNER AND BERNET: ESTIMATION AND MEASUREMENT OF JUNCTION TEMPERATURES

a temperature drop across the coating is expected. The radiation

intensity is described by the StefanBoltzmann Law
(11)
5.67 10
W cm K . For a single IGBT chip
with
at 75 C (
1 cm ) the total radiation amounts to 78.8 mW,
which is less than 0.2% of its total dissipated power. Hence, it
can be assumed that the temperature drop across the coating is
negligible small and the measured average temperature at the
surface equals the real average junction temperature.
V. RESULTS

Fig. 10. Infrared image of a single IGBT chip under operation, average temperature 53.2 C, spot temperature 54 C.

Fig. 11. Measured junction temperature ripple versus time.

measurement error for the average temperature of the chip surK remains.
face of
The correct time averaging of the recorded data requires further attention. Although the camera frequency is lower than the
desired output frequency of the converter, the beat frequency
effect can be used to measure the -time characteristic. The
fundamental frequency of the converter is set to
50.25 Hz.
With a camera frequency of 15 Hz the fundamental period (
19.9 ms) is sampled 20 times within 4 s if every third recorded
200 ms). The result is depicted in
camera sample is used (
Fig. 11. The junction temperature ripple with can be clearly
2 K in the examined example.
recognized. It amounts to
As to how much this measured ripple equals the true ripple of the
junction temperature can not be evaluated. Though the silicon
die and the applied coating are very thin ( 220 m 30 m,
respectively), the tiny thermal capacitances of the chip and the
coating damp the ripple at the chip surface.
Moreover, the question arises as to how exact the measured
average temperature at the surface equals the average temperature at the junction. Due to the heat radiation of the chip surface

The NPC VSC is investigated at the four operating points

specified in Table II. The measured temperatures (averaged over
the chip areas and the fundamental period) are compared to the
calculated temperatures in Fig. 12. It is found that the estimated
values meet the measured values with tolerances of
1
5 K. In the average, the estimated temperatures of the
IGBTs and diodes are 0.6 and 1.7 K higher than the measured
IGBT and diode temperatures, respectively. This tendency indicates a small systematic error. The maximum relative deviation
of the estimated to the measured values with respect to the ambient temperature amounts to 16.7%.
The comparison of Fig. 12(a)(d) also shows that in every
operating point other groups of devices are stressed. The measurements prove the fundamental behavior of the NPC VSC
as analyzed in [9] but also reveal some characteristics specific
to the investigated low-voltage test bench. For inverter operation at large modulation depth [Fig. 12(a)] the outer switches
T and T are the most stressed devices, whereas at small
modulation depth [Fig. 12(c)] the NPC diodes D and D
experience the highest temperatures. This result matches with
the general theoretical analysis. However, both operating points
for rectifier operation display a temperature distribution where
the inner switches T /T and diodes D /D reach by far
the highest junction temperatures. At large modulation depth
[Fig. 12(b)] the outer inverse diodes D and D experience
both conduction and recovery losses, while the inner switches
are only stressed with significant switching and the inner diodes
with significant conduction losses. Nevertheless, the inner devices face the highest temperatures because they are joint in
one half-bridge module and their entire power losses have to
be carried off by only one heatsink. The situation for rectifier
operation at a small modulation depth [Fig. 12(d)], once again,
matches with the general theoretical analysis. Here, the inner
switches are the most stressed devices.
Further discussion and comparison of the NPC VSCs loss
and junction temperature distribution is given in [9] and [10].
VI. EVALUATION OF ERRORS
The good match of the measurement results with the estimated temperatures must not lead to rash conclusions about the
calculation. A detailed analysis of errors is
accuracy of the
given in Table V. The included estimates are based on semiconductor datasheets, the discussion with manufacturers, and own
calculations.

10

IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 22, NO. 1, JANUARY 2007

Fig. 12. Comparison of estimated and measured junction temperatures at four operating points. (a) Case A : M = 0.85; pf = 1. (b) Case B : M = 0.70; pf =
: M = 0.15; pf = 1. (d) Case D : M = 0.15; pf =
1. (# = 25 C, V = 650 V, I = 30 A, f = 8 kHz, sine-triangle modulation).

01. (c) Case C

The summation of all partial errors yields a total maximum

deviation of the estimated junction temperature from its true
value of 8.25 9.29 K. This equals almost 30% of the measured junction temperature with respect to the ambient air temperature. In this worst case scenario all measurement and modeling tolerances affect the total error in the same direction. This
is possible, but rather unlikely. If the individual errors of the
loss approximations, thermal resistances, etc. are assumed to be
independent some errors can compensate each other. If the errors are further assumed to be normal distributed, according to
the Gaussian law of error propagation the maximum total error
stays within 11% with a probability of 68%. This number corresponds to the practically achieved results in Section V.
The electrical and the thermal model contribute about equal
shares to the total error. The electrical model is based on measured conduction and switching losses. Three major sources of
errors can be identified, viz.
the measurements of the losses themselves;
the missing reflection of the temperature dependence;
the approximation of the losses.
Though the loss measurements were performed with the
highest possible precision, the temperature dependence could
not be quantified. All measurements were carried out in the
25 45 C, whereas the operating junction temrange of
peratures in the converter vary from
25 70 C. Thus, an
error of 10% for the effect of on the losses must be assumed.
Also the linear switching loss approximations do not enable a
precision higher then 10%. If both, the loss approximations

are improved (error 5%) and the temperature dependence is

included (error 0%), the total maximum error of the estimated
junction temperature can be reduced to 18 22%.
Whereas the electrical model can be improved with acceptable
efforts, there are no simple means to verify the parameters of
the thermal model. The thermal resistance for junction-to-case
given in the datasheet includes a safety margin for the aging
of the modules due to thermal cycling. Moreover, the whole
structure of the model bears the essential weakness of a radically
simplified geometry. E.g., the coupling of IGBT and diode
value
through the substrate layer is not considered. The
for the transition from the module base plate to the heatsink is
not more than a rough estimate. Here, the heat flow is affected
by a number of aspects as the flatness of the base plate as well
as the thickness and migration of the thermal contact grease.
Since in the special situation of the test setup there are two
such transitions (from and to the inserted aluminum block)
this estimate clearly becomes the weak point of the thermal
model. Finally, the heat convection from the heatsink to the
ambient air is strongly dependent on the heatsink temperature,
the ambient temperature, air turbulences, etc. The datasheet
values of the heatsink do not describe the dependence on these
parameters in detail. Therefore, the error of the heatsink data is
(accuracy
eliminated as the measured heatsink temperature
0.5 k) is taken as a basis for the
calculation.
If the aluminum block and one of the transitions with thermal
contact grease are removed from the calculation, the total
maximum error reduces to 20 24%. Considering a standard

BRCKNER AND BERNET: ESTIMATION AND MEASUREMENT OF JUNCTION TEMPERATURES

TABLE V
ERROR ANALYSIS FOR JUNCTION TEMPERATURE ESTIMATION (EXAMPLE
CALCULATION FOR T AT OPERATING CONDITIONS OF CASE A )

11

VII. CONCLUSION
The authors investigate the junction temperature distribution
calof a three-level NPC VSC with IGBTs by the standard
culation method and by direct infrared measurements. Both apdistribution of
proaches are described in detail. The unequal
the NPC VSC is experimentally proven.
However, close inspection yields that the analytical or simulative junction temperature estimation with simple electrical
component) models based on datasheet values
and thermal (
30%. If
cannot deliver results with tolerances less than 25
the datasheet values are complemented by precise loss measurements in terms of current, voltage, and junction temperature, an accuracy higher than 20% can be achieved. Since
estimathere is currently no simple alternative this common
tion method will be continuously used for the thermal design
of power converters. Due to its simplicity the discussed models
are also useful for the real-time calculation of junction temperatures [3], [4], [9]. Nevertheless, the users should be aware of
the tolerances and limitations.

ACKNOWLEDGMENT
The authors wish to thank Dr. U. Scheuermann and Dr. U.
Nicolai, SEMIKRON, for granting the open IGBT modules and
for valuable advice and discussion, and Dr. R. Schacht, Fraunhofer IZM, for providing the infrared camera system.

mechanical design, i.e., without aluminum block, together

with an enhanced electrical model (see above) an accuracy of
14 18% with respect to the ambient temperature can be
achieved. However, this precision can only be reached with
loss measurements for every individual device. A survey of
semiconductor datasheets shows that the typical manufacturing
tolerances yield a variation of the conduction and switching
25%. Thus, a
estimation based
losses in the range of 20
on datasheet values can not deliver an accuracy higher than
about 25 30%.
As a final point, the accuracy of the junction temperature measurement is assessed. It is mainly determined by
the specified tolerance of the infrared camera ( 1 K);
the effect of the bond wires ( 1 K);
the definition of the chip areas ( 1 K).
The effect of the bond wires has been discussed in Section IV.
Further, it is estimated that the definition of the chip areas across
which the temperatures are averaged causes another error of
1 K. Deviations from the assumed emissivity of the
chip surface can be disregarded due to the applied coating. Also
the temperature difference of the chip surface to the junction can
be neglected (see Section IV). Considering the three above mentioned factors a total measurement accuracy for absolute junc3 K can be concluded. The reltion temperatures of
ative precision for the comparison of the junction temperatures
at different operating points against each other is significantly
higher since the definition of the chip areas remains unchanged
0.08 K.
and the specified resolution of the camera is

REFERENCES
[1] SEMKIRON International, SEMISEL simulation tool, Nuremberg,
Germany, 2006 [Online]. Available: http://www.semikron.com
[2] Infineon Technologies, DIPOSIM dimensioning program, Warstein,
Germany, 2006 [Online]. Available: http://www.eupec.com
[3] D. A. Murdock, J. E. Ramos, J. J. Connors, and R. D. Lorenz, Active thermal control of power electronics modules, in Proc. IEEE-IAS
Annu. Meeting, Salt Lake City, UT, 2003, pp. 15111515.
[4] R. Krmmer et al., On-line calculation of the chip temperature of
power modules in voltage source converters using the microcontroller,
in Proc. EPE Conf., Lausanne, Switzerland, 1999, [CD ROM].
[5] U. Nicolai et al., SEMIKRON Application Manual Power Modules.
Ilmenau, Germany: Verlag ISLE, 2000.
[6] C.-S. Yun, P. Malberti, M. Ciappa, and W. Fichtner, Thermal component model for electrothermal analysis of IGBT module systems,
IEEE Trans. Adv. Packag., vol. 24, no. 3, pp. 401406, Aug. 2001.
[7] F. Profumo, A. Tenconi, S. Fracelli, and B. Passerini, Implementation
and validation of a new thermal model for analysis design and characterization of multichip power electronics devices, IEEE Trans. Ind.
Appl., vol. 35, no. 3, pp. 663669, May/Jun. 1999.
[8] J. Reichl, D. Berning, A. Hefner, and J.-S. Lai, Six-pack IGBT dynamic electro-thermal model: Parameter extraction and validation, in
Proc. IEEE-APEC, Anaheim, CA, 2004, pp. 246251.
[9] T. Brckner, S. Bernet, and H. Gldner, The active NPC converter and
its loss-balancing control, IEEE Trans. Ind. Electron., vol. 52, no. 3,
pp. 855868, Jun. 2005.
[10] T. Brckner, S. Bernet, and P. K. Steimer, The active NPC converter
for medium-voltage applications, in Proc. IEEE-IAS Annu. Meeting,
Hong Kong, China, 2005, pp. 8491.
[11] R. Lappe and F. Fischer, Leistungselektronik-Metechnik. Berlin,
Germany: Verlag Technik, 1993.
[12] J. Lohan, Validation and application of different experimental techniques to measure electronic component operating junction temperature, IEEE Trans. Comp. Packag. Technol., vol. 22, no. 2, pp. 252258,
Jun. 1999.
[13] Thermovision 900 Users Manual, AGEMA Infrared Systems AB,
1994.

12

Thomas Brckner (S99M06) received the

Dipl.-Ing. and Dr.-Ing. degrees in electrical engineering from the Dresden University of Technology,
Dresden, Germany, in 1999 and 2005, respectively.
In 1996 and 1997, he was a Guest Student at the
Virginia Power Electronics Center, Virginia Polytechnic Institute and State University, Blacksburg.
From 1998 to 2005, he worked on several projects
for ABB in Germany and Switzerland as well
the Technical Universities in Dresden and Berlin,
Germany. In 2002, he was a Visiting Researcher at
Monash University, Clayton, Australia. Since 2006, he has been with CONVERTEAM (former Alstom Power Conversion), Berlin. His research interests
include topologies, devices, and controls for high-power conversion.

IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 22, NO. 1, JANUARY 2007

Steffen Bernet (M97) received the M.S. degree

from Dresden University of Technology, Dresden,
Germany, in 1990 and the Ph.D. degree from Ilmenau
University of Technology, Ilmenau, Germany, in
1995, both in electrical engineering.
During 1995 and 1996, he was a Postdoctoral Researcher in the Electrical and Computer Engineering
Department, University of Wisconsin, Madison. In
1996, he joined ABB Corporate Research, Heidelberg, Germany, where he led the Electrical Drive
Systems Group. From 1999 to 2000, he was responsible for ABB research worldwide in the areas Power Electronics Systems,
Drives, and Electric Machines. In 2001, he joined the Berlin University of
Technology, Berlin, Germany, as a Professor of power electronics.