A Transient Thermal Model for Power Electronics

Systems
Neelakantan Padmanabhan
ZF Passive Safety Systems US Inc
Livonia, USA
neel.padmanabhan@zf.com

Abstract—An equation based reduced order model applica- described in this work is based on the concept of thermal
arXiv:2403.03268v1 [math.NA] 5 Mar 2024

ble to generalized heat equation and thermal simulations of lumped element model, but applies modifications in how the
power electronics systems developed in commercial CFD tools, model parameters are measured and suggests a new equation
is presented in this work. The model considers the physics of
heat transfer between multiple objects in different mediums and that can be used to model the transient thermal behavior of
presents a set of equations that can be applied to a wide range of linear and non-linear systems. The ease of setup, computation
heat transfer scenarios including conduction, natural and forced speed and range of applicability make this approach significant
convection problems. A few case studies including heat transfer in in iterative design improvement process in development of
a power electronic system are simulated in Ansys® Icepak™ and power electronics systems. The model presented in this work is
the temperatures from the simulations are compared with the
temperatures predicted by the models. The models are observed built in steps and presented in the following sections. A model
to be highly accurate when compared with the simulations. The for a single body insulated system is built from the transient
predictive model described in this work reduces large complex form of heat conduction equation. The model is then extended
simulations down to a few parameters which tremendously to two body systems and the appropriate modifications are
improves the computation speed, uses very low physical disk presented. This is then extended to a multi-body system with
space and enables fast evaluation of thermal performance of the
system for any changes in the input parameters. mixed modes of heat transfer. A method to estimate the model
Index Terms—Reduced order models, lumped parameter mod- constants for a general system and evaluation of the model
els, ROM for thermal simulations, Icepak ROM with simulation data are then presented.

I. I NTRODUCTION II. BACKGROUND

® ™
Commercial CFD codes such as Ansys Icepak are indis- The generalized heat conduction equation can be expressed
pensable in development of simulations of electronic thermals as,
in a wide range of industries including automotive, aerospace ∂T ∂ ∂T
and energy. A high fidelity thermal simulation requires sig- ρcp = (k ) + P, (1)
∂t ∂xi ∂xi
nificant computational resources and time. The computational
expense increases as the complexity of the system increases. where, P is the source (input) term, ρ, cp , k represent the
Thus, for product design and development process, reduced or- density, specific heat, and thermal conductivity, respectively, T
der models (ROM) are highly desirable. Development of ROM the temperature, and xi , t the spatial and temporal terms. For
is an active area of research [1]. While Linear Time Invariant constant properties at atmospheric pressures [10], this equation
(LTI) based ROM [2] are available in Ansys® suite of tools, can be expressed as,
they are applicable only to linear systems. Approaches for ∂T P
ROM development for non-linear systems include Response = α∇2 T + , (2)
∂t ρcp
surface methodology (RSM) [3], Proper Orthogonal Decom-
position (POD) [4], Subspace projection [5] and machine where, α = k/ρcp , is the thermal diffusivity and ∇2 is the
learning [6] to name a few. While powerful, these approaches Laplace operator. Depending on the type of heat transfer across
are computationally expensive even for simple systems. For the surface of the solid bodies, the boundary conditions can
complex systems with large number of components, the costs be formulated as insulated walls with no heat flux, conduction
of simulations and subsequent ROM development increase across different materials or convection between a solid sur-
non-linearly [7], [8]. Thus, a simple and computationally face and the surrounding fluid. Integrating this equation with
inexpensive method is advantageous. Thermal lumped element respect to time results in an equation of the form,
method [9] is a simple approach often utilized to model P
the thermal behavior of systems. In this method, the model T (t) = D(t) + t + cint . (3)
ρcp
constants are typically estimated from steady state behavior of
where, D(t) = α ∇2 T dt is the diffusion term and cint is
R
a system. However, in power electronics systems, steady state
assumptions often result in incorrect estimation of tempera- the integration constant. This form of the equation is used as
tures because the inputs are highly transient. The approach a starting point to develop models for the thermal systems.
©2024 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future
media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
Accepted for publication in the IEEE SouthEastCon 24 (March 2024).
 III. M ODELS
A. Model for 1-body insulated system
For one body insulated system, it is observed that the diffu-
sion of temperature across the body is instantaneous. Thus, the
diffusion term (spatial gradient) in Eq. 3 is negligible. Hence,
evolution of temperature of the body can be modeled as,
PB 
T B (t) = T 0 + T L (t) = T 0 + t, (4)
CB
where, T 0 is the initial temperature, C B = ρcp V is the thermal
capacitance of the body, V the volume of the body and P B
the power input applied to the body. This temperature curve
is observed to be linear with the slope P B /C B . Thermal
simulations are created for 1-body insulated system where
wall boundary conditions with zero heat flux are applied at
the cabinet. The temperature from the simulations and the
temperature from the model are presented in Figure 4.
B. Model for 2-body conduction heat transfer
For heat conduction between 2 bodies insulated from the
surrounding, it is observed that there is a spatial gradient in Fig. 1: Temperature of bodies, T L (t) from 2-body simulation
temperature from the heat source to the sink. Further, the heat and deviation terms
transfer in this case can be approximated to 1D heat transfer
since the temperature gradient is observed to be significant
only in the direction of heat flow. The temperatures (at center)
of each body, T1B (t) and T2B (t), can thus be modeled as,
T1B (t) = T 0 + T L (t) + T1Dm (t),
(5)
T2B (t) = T 0 + T L (t) + T2Dm (t),
where, T0 denotes the initial temperature, T L (t) = (PTB /CTB )t
and the term PTB /CTB denotes the ratio of total applied power
in both bodies to the total thermal capacitance of both bodies.
For two body system it is defined as, Fig. 2: Analytical computation of thermal resistance
P2 B
PTB i=1 Pi
= 2 . (6)
CTB B
P
i=1 Ci response of an RC circuit [11]–[13]. The model equation can
be expressed as,
The significance of the term T L (t) is that, it represents a
temperature curve that lies between T1B (t) and T2B (t) and
is always linear with a slope of PTB /CTB . This term can be TiDm (t) = Ai (1 − e−ki t ), (8)
computed a priori using the material properties and the input
where, Ai = |TiBsim (tm ) − T L (tm ))| is the magnitude of
powers. The location xPTB /CTB , within the system at which this
difference in stationary temperature of each body and the
temperature curve is observed is variable and depends on the
linear temperature curve, and tm the time instant after which
applied power, geometry, and material properties. For similarly
the temperature curve has reached a stationary state. The sign
sized bodies with similar material properties, xPTB /CTB is
of Ai is positive if TiBsim > T L and vice-versa. Further, the
typically located close to the midpoint between the two bodies.
term Ai can be modeled as,
However, when there are differences in thermal capacitance of
bodies, this location is closer to body with the larger thermal Ai = PiB (x)RiB , (9)
capacitance. The second term, TiDm (t) is a model for TiD (t)
which is the deviation in temperature of the body measured where, PiB (x) is the applied input power at each body. The
from the simulation TiBsim (t) and the linear temperature curve input power is described as a spatial function because the
T L (t). This is defined as, applied power varies spatially across the system. RiB is the
thermal resistance between the body’s center and the location
TiD (t) = TiBsim (t) − T L (t), (7)
xPTB /CTB . The term ki = 1/(CTB RiB ) represents the time
It is observed that TiD (t) (Fig. 1) has the shape of a character- constant or the rate of change of the temperature. To model
istic exponential curve and can be modeled similar to transient TiDm (t) for a general 2-body system as described in Fig.
2, where the source and sink may vary, the principle of
superposition is applied. It is expressed as,
B(1,0) B(0,1)
T1Dm (t) = (P1B R1 + P2B R1 )[1 − e−k1 t ],
B(1,0) B(0,1)
(10)
T2Dm (t) = (P1B R2 + P2B R2 )[1 − e−k2 t ],

B(1,0)
where, R1 is the characteristic thermal resistance between
the center point of body B1 (xB 1 ) and the location xPTB /CTB,
B B
when input P1 = 1 is applied to body B1 and input P2 =
0 is applied to body B2 . This term can be determined by
running a parametric simulation as described in Sec. III-D
and determining the difference in the temperature between the
body and T L ,
Bsim (1,0)
B(1,0) T1 (tm ) − T L (tm ) Fig. 3: Estimated vs calculated thermal resistance
R1 = . (11)
P1B
B B B
B (1,0) with PT,j = P1,j + P2,j .
In this equation, T1 sim (tm ) is the temperature of B1 when
nT
input P1B = 1 is applied to B1 and input P2B = 0 is applied X B(1,0) B(0,1)
to B2 . Since a unit input power is applied, the magnitude of T1Dm (t) = B
[(P1,j R1 B
+ P2,j R1 )
j
temperature difference is equivalent to the thermal resistance.
B B(1,0) B B(0,1)
Similarly, −(P1,j−1 R1 + P2,j−1 R1 )]
−k1 (t−t0j−1 )
B(0,1) Bsim (0,1) (1 − e ),
R1 = T1 (tm ) − T L (tm ); P1B = 0, P2B = 1, nT (15)
B(1,0) B(0,1)
X
B(1,0)
R2 =
B
T2 sim
(1,0)
(tm ) − T (tm ); P1B
L
= 1, P2B = 0, T2Dm (t) = B
[(P1,j R2 + B
P2,j R2 )
B(0,1) Bsim (0,1) j
R2 = T2 (tm ) − T L (tm ); P1B
= 1. = 0, P2B B B(1,0) B B(0,1)
(12) −(P1,j−1 R2 + P2,j−1 R2 )]
B(1,0) B
The time constants are expressed as, k1 = 1/(R1 CT ), −k2 (t−t0j−1 )
B(0,1) B
(1 − e ),
and k2 = 1/(R2 CT ). The thermal resistance terms can
also be determined analytically for simple configurations. A where, j is the index of number of changes in the input power,
general form of thermal resistance for conduction is expressed nT is total number of changes (transients) in the input, and
as, t0j is the time instant at which the change in input powers
B B
occurs. The initial values are, P1,0 = P2,0 = 0, t00 = 0.
Li xB
i − xPTB /CT
B
Thermal simulations are created for 2-body insulated system
RiBcalc = = , (13)
K i Ai K i Ai where wall boundary conditions with zero heat flux are applied
at the cabinet. The temperature from the simulations and the
where, the location xPTB /CTB can be approximated close to the temperature from the model are presented in Fig 6.
center of the two bodies. An illustration of this is provided in
Fig. 2. A set of trials are conducted by varying the material C. Model for convection heat transfer
properties, geometry, and sizes of each body. Further the For convection heat transfer, with one solid body, the prob-
sources and sinks of heat for each case are also varied. It is lem essentially has two mediums, the solid and the surrounding
observed that a linear relationship exists between the thermal fluid. When the flow velocity of the fluid is non-zero it is
resistance values determined from optimization and thermal observed that the temperature evolution of the solid body
resistance calculated analytically, (RiB = mi RiBcalc + ci ) as follows an exponential curve. Thus, for one solid body in a
observed in Fig. 3. fluid, the temperature evolution of the solid body is given by,
Further, the changes in thermal resistance terms t

T B−F (t) = T 0 + P1B RB−F (1 − e R

B(1,0) B(1,0) B(0,1) B(0,1) B−F C B
R1 , R2 , R1 , R2 due to changes in geometry T ). (16)
or material can be calculated by scaling the terms L, k, A. The term RB−F = 1/hA denotes the thermal resistance
For two body system with transient powers, P1B (t), P2B (t), between the body and the fluid, h the convection heat transfer
a piece-wise approximation method is applied. The model coefficient, A the cross-sectional area, and CTB = CSB + CFB is
equations are expressed as follows, the total thermal capacitance of the solid body and the fluid.
nT
"
B B
# While determination of convection heat transfer coefficient for
X PT,j PT,j−1 complex geometries typically requires multiple experimental
T L (t) = − (t − t0j−1 ), (14)
j
CTB CTB runs, it can also be calculated from Ansys® Icepak™ for a
given configuration [Report > Full Report Variable: Con- transient inputs in power electronic systems typically
vection Heat Transfer Coefficient]. A set of trial runs were change every few seconds. The model parameters that
performed at different flow velocities, where the RB−F that characterize the temperature curve in the shorter time
best fits the characteristic temperature T B−F (t) for the given range result in higher accuracy. The parametric study can
configuration, were estimated using curve-fitting. Further, the be automated in Ansys® Icepak™ from the Parameters
convection heat transfer coefficient and area calculated from and Optimization > Parametric trial option.
Icepak™ for the given configuration were determined. The • The slope PTB /CTB and the linear temperature curve
measured and estimated values of heat transfer coefficients T L (t) are computed. For complex geometries the volume
are compared in Table I. It is observed that the heat transfer of the body can be determined using a mechanical CAD
coefficient estimated from curve-fit is same as the heat transfer package or by approximating the complex geometry with
coefficient obtained from the tool. This provides a confirma- a number of smaller cuboid geometry whose volume can
tion of the form of the model equation that best captures the be determined easily.
heat transfer in this configuration. For simple geometries and • The deviations TiD (t) = TiBsim (t) − T L (t), in simulation
flow configurations the convection heat transfer coefficient can temperatures from each trial and the linear temperature
also be estimated analytically [14], [15]. are measured.
• The characteristic thermal resistance (RiBchar = Ai ) and
TABLE I: RB−F represents the thermal resistance estimated time constants (ki ) for each body are determined,
from curve-fitting T B−F (t), A the surface area measured
– The parameters, Ai and ki that minimize the sum
from the configuration, and hcalc the convection heat transfer
of square error between the temperature deviation
coefficient calculated by Ansys® Icepak™ .
TiD (t) and the model temperature TiDm (t), SSE =
(TiD − TiDm )2 , are estimated using a non-linear
P
RB−F = hest =
U f low hcalc
1/hest A A(m2 ) 1/ARB−F optimization algorithm. In this work a non-linear
(m/s) (W/m2 K)
(K/W ) (W/m2 K)
0.00098 303.692 0.00015 21.95 21.83 Generalized Reduced Gradient solver in Excel™ was
5 118.9 0.00015 56.07 56.19 used to evaluate the parameters. Estimation was also
50 28.47 0.00015 234.09 235.7 performed with the fminsearch function in Matlab ™
and similar parameters were obtained. The sign of
Ai is negative if TiD − T L < 0.
D. Model for multiple bodies with mixed conduction and • The model deviation terms TiDm (t) are computed and the
convection heat transfer model temperature of the bodies are then computed as,
For a general multi-body system with mixed conduction and TiB (t) = T 0 + T L (t) + TiDm (t).
convection heat transfer modes, the form of model equations
IV. S IMULATION S ETUP
remains consistent with the two-body conduction / convection
models. The equations for the system are expressed as, An overview of the simulation setup used in this study
PNB B to validate the models can be found in this work [16].
PB Pi The simulations include one body, two body and multi-body
T L (t) = TB t = PN i
t,
CT i
B
CiB systems in insulated, pure conduction, natural convection and
TiB (t) = T 0 + T L (t) + TiDm (t), forced convection modes. Detailed mechanical CAD models,
(17)
NB
electrical CAD models and material properties are used for
X Bchar B the components (PCBA, MOSFET, Shunt) in simulation of
TiDm (t) = PiB RiBchar [1 − e−t/Ri CT
].
i
power electronics system. All the simulations are run at
pressure of 1atm and ambient temperature of 20◦ C. For
The term NB is the total number of bodies in the system, natural convection problems, the Boussinesq approximation is
including the fluid, RiBchar the characteristic thermal resis- applied. The grid resolution in every coordinate direction is
tance between the body i and location xPTB /CTB . Further, the set such that the measured statistics become independent of
characterization of thermal resistances are required only for the resolution.
the bodies / sub-systems of interest. The temperature from the
simulations and the temperature from the model are presented V. R ESULTS
in Fig 7 Comparison between the model and simulation for 1-body
1) Estimation of thermal resistance and time constants for with insulated boundaries is presented in Fig. 4, comparison
general multibody system: for convection is presented in Fig. 5, comparison for 2-body
• A parametric study is performed by applying 1W of conduction is presented in Fig. 6 and the comparison for a
input power to each source, while keeping the inputs general multi-body configuration (for power electronics) with
at other bodies at 0. The number of trials is equal to mixed conduction and convection modes is presented in Fig.
the number of sources. The temperature response of 7. For brevity, the results from the finite difference codes for
each body is measured for each trial. Each simulation generalized heat equation are not presented in this paper. The
is run only for a short duration (tm ≈ 20s), since the models are observed to be highly accurate for each case with
 (a) Input power dissipation (b) Material: Silver (metal)

(a) Input power dissipation (b) Natural convection

(d) Simulation configuration - 1-

(c) Material: Asbestos (insulator) body (Asbestos)
Fig. 4: Model vs simulation for 1-body insulated system (d) Simulation configuration -
(c) Uf low = 50m/s Forced convection
Fig. 5: Model vs simulation for 1-solid body (Cu-Bronze),
the calculated average error of less than 3%. A small deviation fluid (air) medium natural and forced convection system
in the simulation and model temperature of the PCBA in Fig. 7
is observed. This is seen because the PCBA is a body with high
thermal capacitance and the temperature deviation measured
between the PCBA and the linear temperature curve for the
simulation time range is not purely exponential. The model
files occupy 0.01% of the total physical disk space that detailed
simulation and solution files typically occupy.

TABLE II: Comparison of run time between simulations in

Icepak ™ and model in Matlab ™ (simulations and model code
were run on the same computer). It is to be noted that paramet-
ric simulations are required to obtain the model coefficients
before running the model. The parametric simulations however
are run at a constant (large) ∆t and are required to be run only
(a) Input power dissipation (b) T L (t) computed a priori
once.
Parametric
Simulation Model
Simulation
run time run time
Case run time
in Icepak in Matlab
in Icepak
(s) (s)
(s)
1-Body Insulated 130 0 0.072679
1-Body Convection 137 120 0.22
2-Body Conduction 122 370 0.1790
Power electronics simulation 126 240 0.229
(c) Temperature of body 1 (d) Temperature of body 2

VI. M ODEL LIMITATIONS AND CORRECTION FOR STEADY

STATE SYSTEMS
The approach presented in this work, models the deviation
T D (t) in temperature between a body and the linear tem- (e) Simulation configuration
perature curve. This approach is however applicable only to
transient systems where the inputs vary rapidly. For steady Fig. 6: Model vs simulation for 2-body conduction. Material
state systems where a body’s temperature T B (t) follows an 1: Ag, Material 2: FR4
exponential curve, temperature deviation cannot be modeled
in the same manner as described in this work. For such
 ROM-Transient-thermal-model-for-Power-Electronics-
R EFERENCES
[1] Alfio Quarteroni and Gianluigi Rozza, Reduced Order Methods for
Modeling and Computational Reduction. Springer, 2014.
[2] S. Asgari, X. Hu, M. Tsuk, and S. Kaushik, “Application of POD plus
LTI ROM to Battery Thermal Modeling: SISO Case,” SAE International
Journal of Commercial Vehicles, vol. 7, no. 1, pp. 278–285, Apr. 2014,
doi: https://doi.org/10.4271/2014-01-1843.
[3] L. Raeisian, H. Niazmand, E. Ebrahimnia-Bajestan, and P. Werle, “Ther-
(a) Input power dissipation (b) Temperature of MOSFET mal management of a distribution transformer: An optimization study of
the cooling system using CFD and response surface methodology,” In-
ternational Journal of Electrical Power & Energy Systems, vol. 104, pp.
443–455, Jan. 2019, doi: https://doi.org/10.1016/j.ijepes.2018.07.043.
[4] N. Padmanabhan, ”Reduced order model of a convection-diffusion
equation using Proper Orthogonal Decomposition,” arXiv preprint
arXiv:2303.07176. 2023.
[5] Jaroslav Šindler, A. Suleng, T. Jelstad Olsen, and Pavel Bárta, “Krylov
Model Order Reduction of a Thermal Subsea Model,” International
Journal of Mechanical and Mechatronics Engineering, vol. 7, no. 5, pp.
842–849, May 2013.
[6] S. Fresca and A. Manzoni, “POD-DL-ROM: Enhancing deep learning-
(d) Temperature of multilayer based reduced order models for nonlinear parametrized PDEs by proper
(c) Temperature of shunt PCBA orthogonal decomposition,” vol. 388, pp. 114181–114181, Jan. 2022,
doi: https://doi.org/10.1016/j.cma.2021.114181.
[7] E. Parish, C. R. Wentland, and Karthik Duraisamy, “The
Adjoint Petrov–Galerkin method for non-linear model
reduction,” Computer Methods in Applied Mechanics and
Engineering, vol. 365, pp. 112991–112991, Jun. 2020, doi:
https://doi.org/10.1016/j.cma.2020.112991.
[8] N. Padmanabhan and R. S. Miller, “Assessment of subgrid
scale mixing models used in LES at high pressures,” Journal
of Turbulence, vol. 19, no. 8, pp. 683–715, Jul. 2018, doi:
https://doi.org/10.1080/14685248.2018.1498590.
[9] A. S. Bahman, K. Ma, and F. Blaabjerg, “A Lumped Thermal
Model Including Thermal Coupling and Thermal Boundary Con-
ditions for High-Power IGBT Modules,” IEEE Transactions on
Power Electronics, vol. 33, no. 3, pp. 2518–2530, Mar. 2018, doi:
https://doi.org/10.1109/tpel.2017.2694548.
[10] “Appendix A: Physical Properties,” Thermoelectrics:
Design and Materials, pp. 323–352, Sep. 2016, doi:
(e) Simulation configuration https://doi.org/10.1002/9781118848944.app1.
[11] [1]E. Brennan, “RC Circuit Formula Derivation Using Calculus,” Owl-
Fig. 7: Model vs simulation for a general multi-body power cation. https://owlcation.com/stem/RC-Circuit-Time-Constant-Analysis
electronics system - PCBA with MOSFET and Shunt in a [12] Thermal Resistance Theory and Practice, SMD Packages, Infineon
Technologies, 2000.
natural convection environment [13] Silva, “Modeling the Transient Response of Thermal Circuits,” Ap-
plied sciences, vol. 12, no. 24, pp. 12555–12555, Dec. 2022, doi:
https://doi.org/10.3390/app122412555.
aPsystem, T B can be directly modeled as T B (t) = T 0 + [14] F. P. Incropera, Fundamentals of heat and mass transfer. Hoboken, Nj:
NB Wiley, 2013.
i PiB Richar [1 − e−ki t ]. [15] V. Ranmode, M. Singh, J. Bhattacharya, ”Analytical formulation of
effective heat transfer coefficient and extension of lumped capacitance
VII. C ONCLUSIONS method to simplify the analysis of packed bed storage systems,” Solar
The approach described in this work can accurately model Energy, 183,2019, 10.1016/j.solener.2019.03.066.
[16] N. Padmanabhan, A. Gebregergis and S. Veigas, ”An integrated approach
the temperatures from simulations of highly transient power to developing thermal models for automotive electric drives,” 2020 IEEE
electronics systems in various heat transfer modes. Future Energy Conversion Congress and Exposition (ECCE), Detroit, MI, USA,
work based on this approach includes evaluation of appli- 2020, pp. 3186-3192, doi: 10.1109/ECCE44975.2020.9236332.
[17] N. Padmanabhan, ”https://github.com/neelp-87/ROM-Transient-thermal-
cability of the model to experimental test setups of power model-for-Power-Electronics-”, 2024.
electronics systems with mixed heat transfer modes, and
including the effects of radiation. This includes development
of analytical models for thermal resistance and capacitance of
various sub-components, measurement of thermal parameters
from test setups and validation of model temperatures against
temperatures from real systems.
A PPENDIX
A code of the model in Matlab ™ for various configurations
is presented in the following link: https://github.com/neelp-87/