Highlights

FastCTF: A Robust Solver for Conduction Transfer Function Coefficients and Thermal Response Factors

Khodr Jaber

• Open source solver for CTF coefficients and response factors

• Series expansion estimation of hyperbolic conduction transfer function

• Performance under various time steps and wall weights

arXiv:2103.16312v1 [cs.CE] 30 Mar 2021

• Comparison with ASHRAE Handbook and Frequency-Domain Regression

 FastCTF: A Robust Solver for Conduction Transfer Function Coefficients and
Thermal Response Factors

Khodr Jabera,∗
a Meinhardt Group, Amwal Tower, Doha, Qatar

Abstract

Conduction transfer functions (CTF) are commonly used in the building services to quickly estimate hourly con-
duction heat loads through multilayered walls without resorting to expensive, time-consuming solutions of the heat
equation. It is essential for any software developed for this purpose to be able to simulate walls of varying weight with
a high degree of accuracy. A robust algorithm for computing CTF coefficients and thermal response factors based on
power series expansions of solutions of the governing equations in the complex s-domain is presented and validated.
These series expansions are used to to construct Padé approximants of the system’s transfer functions, which greatly
simplifies the inversion of the solution from the complex domain to the time domain, and allows for an easy recovery
of a time series representation via the Z-transform. The algorithm is also implemented in an open-source C++ code.
Its performance is validated with respect to exact theoretical frequency characteristics and its results are compared
with data generated by previously established methods for computing CTF coefficients / response factors.
Keywords: CTF Coefficients, Thermal Response Factors, Heat Flow Calculations, Transient Heat Conduction

1. Introduction

The need for efficient heat load calculation strategies in the building services has led to the development of a
variety of numerical tools that allow for automated implementations which had previously been performed manually
by hand. Examples of such strategies include the heat balance method [6] and its derivatives - the transfer function
method [7], cooling load temperature difference and the radiant time series method [8]. The heat balance method
directly applies first principles and involves setting up an energy balance over conditioned spaces in terms of surface
and interior temperatures / heat fluxes, resulting in a system of equations whose solution can be cumbersome due to
transient interactions between internal surface conditions and element conduction effects driven by the heat equation
(which is assumed to be 1D for roofs and walls):
∂T k λk ∂2 T k

=


 ∂t ρk C p,k ∂x2 ,


k = 1, ..., Nl

(1)
∂T k


= −λk

qk


∂x
∗ Correspondingauthor. Tel.: +974 3362 8564
Email address: [email protected] (Khodr Jaber)

Preprint submitted to Energy and Buildings March 31, 2021

 
T k (0, t) = T k−1 (Lk , t)




, k = 2, ..., Nl


 (2)
qk (0, t) = qk−1 (Lk , t)




where λk , ρk , C p,k , T k and qk denote thermal conductivity, density, specific heat capacity, temperature distribution
and heat flux through the kth layer of an Nl -layer element. Direct simulation of the governing equations is made
computationally infeasible by the stability constraints imposed on the choice of time step (which would need to be on
the order of an hour to estimate yearly energy profiles in a reasonable amount of time). This constraint has motivated
the development of methods based on time series representations of the conduction process which simplify the internal
surface-conduction relationship and provide the basis for a time marching algorithm with a more feasible time step.
The most popular of these methods are the response factor method and transfer function method, the latter which
continues to be employed in well-known commercial software such as Carrier HAP as it does not require surface /
interior conditions as input and does not place any constraint on outdoor conditions (as opposed to the RTS, which
exploits the steady periodic nature of a day-by-day calculation process). The reader is referred to [13] and [18] for an
exhaustive historical review of transient heat flow calculation methods.
The response factor method is an approach in which heat fluxes are written as a time series in terms of temperature
history:
∞
X ∞
X
qext. (n∆t) = X[k]T ext. [(n − k)∆t] − Y[k]T int. [(n − k)∆t] (3)
k=0 k=0
X∞ X∞
qint. (n∆t) = Y[k]T ext. [(n − k)∆t] − Z[k]T int. [(n − k)∆t] (4)
k=0 k=0

where T ext. , T int. are external and internal temperatures of the element, and the coefficients {X}k , {Y}k , {Z}k are referred
to as external, cross and internal response factors. These coefficients represent the response of a thermal element
when subjected to temperature pulses in discrete time {k∆t}k . Brisken and Reque are credited [13] with establishing
the method [1], which was later improved significantly by Mitalas and Stephenson [2, 3, 4]. It is noted that this time
series could become infeasibly long, motivating Mitalas and Stephenson to develop the transfer function method [7],
which allowed heat fluxes to be represented by a substantially shorter series in both temperature and flux history:
X X X
qnet = bk T o,k−n − dk qk−n − T rc ck (5)
k k k

where T o,k−n is the outdoor temperature at t = (k − n)∆t, T rc is the required room temperature (assumed to b con-
stant) and bk , ck , dk are the transfer function coefficients. To do this, Laplace transforms are usually applied to the
governing conduction equations so that the analysis can be performed in the complex s-domain (although this need
not be the case, as in Davies’ time-domain analysis [10], or the state-space methodology which employs finite dif-
ferences/element analysis to recover the conduction transfer function). Z-transforms are then used to transform the
time-domain solution (which is obtained after applying an inverse Laplace transform) into a discrete-time transfer
function (i.e. the z-transfer function). Direct root finding (DRF) was first developed to obtain the inverted solution by
2
resolving the integral (inverse) transform via the residue theorem. DRF is used in simulation programs such as PREP
(in TRNSYS [20]) and BLAST [21], while the state space method is used in the EnergyPlus software [22]. The RTS
method of Spitler et al. is a special case of the response factor method and assumes that outside temperature can be
put in the form of a 24-hour periodic set, so that the response factors can be written as:
23
X 23
X
qnet (n∆t) = YRTS [k] · T o,k−n − T rc ZRTS [k] (6)
k=0 k=0

where {YRTS }, {ZRTS } are response factors for periodic conditions. It has been shown by Spitler and Fisher [11] that
these response factors can be derived from conduction transfer coefficients using linear algebra. More recently devel-
oped methods include Frequency-Domain Regression (FDR) by Wang and Chen [13, 15] (which uses least squares
regression in the frequency domain to simplify Laplace inversion), Direct Numerical Integration (DNI) by Varela et al
[17] (which provides an alternative numerical Laplace inversion strategy) and Frequency-Domain Spline Interpolation
(FDSI) by Pérez et al. [19] (which employs Fourier analysis to recover the heat flux in the frequency domain).
FastCTF is an open source C++ code capable of computing CTF coefficients and response factors. It implements
a novel algorithm motivated by the FDR method, as presented in [12, 13], in that the transfer functions of the system
in the s-domain (consisting of a complex arrangement of hyperbolic trigonometric functions) are approximated in the
form of rational functions:

βk s
Pm e k
G(s) ≈ Pmk=0 (7)
k=0 eαk s k

This is done by approximating the s-domain solution via high-order Taylor series expansions and using them to
construct rational functions representations of the external, cross and internal flow transfer functions via suitable order
Padé approximations. Due to heavy use of series expansions in the s-plane, the algorithm will be referred to as the
Complex Domain Series Expansion (CDSE) method for the remainder of this paper. Once Padé approximants of the
transfer functions have been determined, the inverse Laplace transform is easily applied to recover the solutions in the
time domain. Finally, Z-transforms are used to derive the z-transfer functions (from which the CTF coefficients are
obtained). An appropriate partial fraction decomposition is performed via the residue method to facilitate the inverse
Laplace transform. Forms of the time-domain solutions can be predicted in advance through knowledge of the poles
and residues, so the z-transfer functions are assembled immediately after these have been calculated.
The CDSE method is a purely mathematical exercise requiring only basic linear algebra for computing Padé
approximants and polynomial roots. Iterative methods are not used at any point, eliminating potential numerical
stability concerns. The conditioning of the relevant linear systems does not require any special care, even when
approximants of up to order 10 are sought (although order 6 is sufficient in practice). Quality control of the results
corresponds to tolerances on errors between the frequency characteristics of the CTF z-transfer functions and their
theoretical counterparts. Analysis of the latter is formalized using an error criterion defined by Chen et al. [16] in
their investigation of a verification and validation strategy for CTF coefficients and response factors.

3
 The s-transfer function estimation procedure will be described first, followed by an illustration of the time-domain
solution corresponding to a ramp input and its corresponding Z-transform. Steps for assembling the z-transfer func-
tions are then shown, and it is demonstrated that the thermal response factors can be accurately recovered from them.
Finally, a set of case studies will be investigated to verify the results of the CDSE formulation with respect to theo-
retical values and in comparison to other results reported in the literature (specifically, those generated by the FDR
method and reported in ASHRAE’s well-known Handbook for HVAC calculations).

2. Modeling of Conduction Through a Multilayered Element

Assuming that the physical properties of all layers in a conductive element are constant, the Laplace transform can
be applied to governing equations (1), (2) to obtain a complex-domain solution of the form:
    
T   A(s) B(s)  T 
 in     out 
  =     (8)
qin C(s) D(s)  qout 
| {z }
M(s)

where the overall transmission matrix M(s) can be written as a product of individual transmission matrices for each
layer of the conductive element:
Nl
Y 
M(s) = Min M j (s) Mout (9)
i=1

  r s  r s  r s 
cosh Lj sinh L j λj
     
αj αj α j 
1   1
Rin  Rout 
Min =   , M j (s) =  r   , Mout (s) = 

(10)
  
r   r s
0 1 1≤ j≤Nl 
 λ j s  s  0 1 
sinh L j cosh L j 
αj αj αj

and Min , Mout correspond to transmission matrices for the inner and outer surfaces films of the element, respectively.
Transfer functions for the heat flux entering / leaving an element can be found by rewriting the system above as a
function of the temperatures:

  D(s)
 1  

 q   B(s) −     
 int.   B(s)  T ext.  G X (s) −GY (s) T ext. 
  =     =     (11)
qext.   1 A(s)   T int.  GY (s) −GZ (s)  T int. 
−
B(s) B(s)
Here, G X (s), GY (s) and GZ (s) are the external, cross and internal transfer functions, respectively. Explicitly, we have:

A(s) 1 D(s)
G X (s) = , GY (s) = , GZ (s) = (12)
B(s) B(s) B(s)

4
3. Numerical Methodology

3.1. Series Expansions Formulation for the s-Transfer Function

The present aim is to construct rational function approximations of the s-transfer functions of the form:

k=0 βφ,k s
Pm e k
Gφ (s) = , φ ∈ {X, Y, Z} (13)
1+ m αφ,k sk
P
k=1 e

In doing so, computing the inverse Laplace transform becomes straightforward after performing partial fraction de-
composition. To obtain this rational function approximation, Taylor series expansions are sought for elements of the
individual transmission matrices. For (M j )1,1 and (M j )2,2 , we have:
 p 2k
[ j] [ j]
 p  X N L s/α
cosh L[ j] s/α[ j] = + O sN+1


(14)
k=0
(2k)!

where L[ j] and α[ j] are the length and thermal diffusivity of the jth wall layer. To get the Taylor series expansions of
the other matrix entries, a Puiseux series (that is, a power series with fractional exponents in the indeterminates) is
√
sought for the hyperbolic sine function in terms of s:
 p 2k+1
[ j] [ j]
 p  X N L s/α
sinh L[ j] s/α[ j] = + O sN+1/2


(15)
k=0
(1 + 2k)!
p
In other words, the Taylor series of sinh (L[ j] y) is found, with y substituted for s/α[ j] . Now, the square-root term can
be divided/multiplied into this series to obtain (M j )1,2 and (M j )2,1 , respectively:
 2k+1  p 2k
[ j] [ j]
 p    p  XN L s/α
sinh L[ j] s/α[ j] s/α[ j] = + O sN+1


(16)
k=0
(1 + 2k)!
 2k+1  p 2(k+1)
[ j]
p   p  XN−1 L s/α[ j]
s/α[ j] sinh L[ j] s/α[ j] = + O sN+1


(17)
k=0
(1 + 2k)!

Multiplication of the individual transmission matrices can now be performed in sequence:

 N N 
X [ j] k X [ j] k

 M11,k s M12,k s 

M j =  k=0 k=0  M j−1 , j = 1, ..., Nl + 1

N N  (18)
X [ j] X
[ j]
 M21,k sk M22,k sk 
|k=0 {zk=0 }
M[ j]

M0 = Mout , M[Nl +1] = Min , MNl = M (19)

 N N 
X X 
 M11,k sk M12,k sk 
 
M =  k=0 N
k=0
N

 (20)
X X
 M21,k sk M22,k sk 
k=0 k=0

5
 To ensure that the resulting approximation of the Laplace-domain solution is of high accuracy, terms of up to
order 20 are included in the Taylor series expansions. The external, cross and internal heat flow transfer functions are
thus represented as ratios of these series entries, however, the numerical conditioning of a polynomial root-finding
problem applied to the denominators of these rational functions at this stage would be poor and the method would
become redundant. On the other hand, the unique structure of the cross flow transfer function can be exploited by
realizing that it is just the multiplicative inverse of a Taylor series. With this series, an appropriate-order s-transfer
function can be estimated via its corresponding [n, m]-order Padé approximant, defined by:
∞ Pn
1 X Pk x k
= T k s ≈ Pmk=0
k
k
(21)
GY (s) k=0 k=0 Qk x

which is computed by moving the left-hand-side into the fraction:

∞ Pn k P∞ k Pm k Pn k
k=0 Pk x k=0 T k s k=0 Qk x − k=0 Pk x
X
k
T k s − Pm k
≈ P m k
=0 (22)
k=0 k=0 Qk x k=0 Qk x

and requiring that the coefficients of the numerator vanish (Q0 = 1 is fixed to ensure uniqueness of the approximant):
    
  Q1   −T 0 

 0
 0 ... 0 −1 0 ... 0     
  Q2   −T 1 
... 0 −1 ... 0    

 T 0 0 0
  .   . 

 . . .
 ..   .   . 
   
     
  Q  −G 
T m−1 T m−1 ... T m−n 0 ... −1  n   m 

0
    =   (23)
T
 m+1 T m ... T m−n+1 0 0 ... 0    
  P0   0 


T m+2 T m+1 ... T m−n+2 0 0 ... 0   P   0 
     
  1  
 ..
 
  .   . 
 .   .   . 
   .   . 
T m+n T m+n−1 ... Tm 0 0 ... 0 P   0 
    
m

Applying the multiplicative inverse is straightforward, and we now have:

Pn
Qk x k
GY (s) ≈ Pk=0
m k
(24)
k=0 Pk x

It will be shown later that the z-transfer function that is computed from this approximant will be always have an
mth -order numerator and denominator, so the choice of n ≤ m is arbitrary (n = m is fixed in the implementation and in
the remainder of the method description for convenience).
P
The Padé approximants for G X (s) and GZ (s) are derived using the polynomial k Pk by equating their transfer
functions with corresponding transmission matrix entries in the following way:
PN k Pm k
k=0 M11,k x k=0 (QX )k x
G X (s) ≈ PN = P m k
(25)
k=0 Pk x
k
k=0 M12,k x
PN m
M22,k xk k
P
k=0 (QZ )k x
GZ (s) ≈ Pk=0
N
= P m k
(26)
k=0 Pk x
k
k=0 M12,k x

6
After clearing denominators on both sides of the equations, recurrence relations can be established for the (QX )k and
(QZ )k :
N m k−1
1  X X  X 
(QX )k = M11, j x j Pjxj − (QX ) j M12,k− j , k = 0, ..., m (27)
M12,0 j=0 j=0
k
j=0
N m k−1
1  X X  X 
(QZ )k = M22, j x j Pjxj − (QZ ) j M12,k− j , k = 0, ..., m (28)
M12,0 j=0 j=0
k
j=0

3.2. Response Factors and the Conduction Transfer Function

The remainder of the calculation procedure is similar section 4 of [13] albeit with a modified Laplace inversion
formula that accounts for complex poles in the transfer functions (real roots were guaranteed in the FDR method since
the rational function approximation was fitted to the frequency response of the system). The generalized Laplace
inversion formula for a system subjected to a ramp input can be written as:
ΛR ΛI 
βk sk
Pr e K1 K2 X γk X ς1 + jς2 ς1 − jς2 
k=0
= + + + + (29)
s2 (1 + k=1 e
m
αk sk ) s2 s + rk k=1 s + σ1,k + σ2,k j s + σ1,k − σ2,k j
P
s k=1
ΛR ΛI
k=0 βk s
 Pr e k  X X
q(t) = L−1 2 = + + γ −rk t
+ 2e−σ1,k t ς1 cos(σ2,k t) + ς2 sin(σ2,k t)


K 1 t K 2 k e (30)
s (1 + k=1 e αk s k )
Pm
k=1 k=1

where rk are poles of the s-transfer function (with rk = σ1 + jσ2 for complex cases) and γk , ς1/2,k denote residues in
the partial fraction decomposition corresponding to ΛR real roots and ΛI pairs of complex roots (that is, a complex
root and its conjugate are a single element counted by ΛI such that ΛR + 2ΛI = m ), respectively.
This ramp response is used to construct triangular pulse responses that define the response factors. Three ramps
at t − ∆t, t, t + ∆t with slopes of ∆t−1 , −2∆t−1 , ∆t−1 are superimposed, and the response of the system at time t = k∆t
defines the kth response factor Y[k]:

1
θ(t) = q(t − ∆t) − 2q(t) + q(t + ∆t)

(31)
∆t
1
Y[0] = q(∆t) (32)
∆t
1
Y[k] = [q((k − 1)∆t) − 2q(k∆t) + q((k + 1)∆t)], k = 1, 2, ... (33)
∆t

The function θ(t) is the heat flow due to a triangular pulse excitation in the continuous time domain. Applying the
Z-transform to this function recovers the transfer function in the z-domain:

1
GZφ (z) = Z θ {n∆t}∞
n=0 = Z[q {(n − 1)∆t}∞n=0 ] − 2Z[q {n∆t}n=0 ] + Z[q {(n + 1)∆t}n=0 ]
∞
∞




(34)
∆t
1  −1
= z Z[q {n∆t}∞ n=0 ] − 2Z[q {n∆t}n=0 ] + zZ[q {n∆t}n=0 ]
∞
∞



(35)
∆t
(z − 1)2 1
= Z[q {n∆t}∞


n=0 ] (36)
z ∆t

7
 To get Z[q {n∆t}∞


n=0 ], a Z-transform is applied to each sub-element of q(t):

ΛR 
K1 ∆tz γk erk ∆t z

K2 z X
Z q {n∆t}n=0 =
∞

+ +

(37)
(1 − z)2 z − 1 k=1 erk ∆t z − 1
 

ΛI 
2ς1 e2σ1 ∆t z2 + 2eσ1 ∆t (ς2 sin(σ2 ∆t) − ς1 cos(σ2 ∆t)) z
X   
+  2σ ∆t  2  σ ∆t (38)
e 1 z − 2e 1 cos(σ2 ∆t) z + 1

k=1
ΛX
R +ΛI
K1 ∆tz K2 z Nk
= 2
+ + (39)
(1 − z) z−1 k=1
Dk
z M(K1 ∆t − K2 + K2 z) + z(1 − z−1 )2 M
= 2
(40)
(1 − z) M

where:
ΛY
R +ΛI

M= Dk (41)
k=1
ΛX
R +ΛI ΛY
R +ΛI

M= Nk Dj (42)
k=1 j=1
j,i

Now, the coefficients of the z-transfer function are recovered using the following identities:
 M(K ∆t − K + K z) + z(1 − z−1 )2 M  a0 + a1 z−1 + ... + am z−m
1 2 2
GZX (z) = = (43)
M ∆t X 1 + d1 z−1 + .... + dm z−m
 M(K ∆t − K + K z) + z(1 − z−1 )2 M  b0 + b1 z−1 + ... + bm z−m
1 2 2
GZY (z) = = (44)
M ∆t Y 1 + d1 z−1 + .... + dm z−m
 M(K ∆t − K + K z) + z(1 − z−1 )2 M  c0 + c1 z−1 + ... + cm z−m
1 2 2
GZZ (z) = = (45)
M ∆t Z 1 + d1 z−1 + .... + dm z−m

The response factors can be recovered without explicitly computing the time-domain solutions in (32) and (33)
by computing the Taylor series expansions of the z-transfer functions in terms of z−1 . To see this, the definition of the
z-transfer function is invoked, and we have that the ratio of output (in terms of response factors) to input (in terms of
the unit temperature pulses is given by:
Z[{φ[k]}∞ φ[k]z−k
P∞
GZφ (z) = k=0
= k=0
, φ ∈ {X, Y, Z} (46)
Z[{ϑ[k]}∞
k=0 ] 1
∞
X
= φ[k]z−k (47)
k=0

where ϑ[k] is the unit triangular temperature pulse at time k∆t. Its Z-transform is easily shown to be equivalent to 1:

1 e
k=0 ] =
Z[{ϑ[k]}∞ n=0 ] − 2Z[ f {n∆t}n=0 ] + Z[ f {(n + 1)∆t}n=0 ]
Z[ f {(n − 1)∆t}∞

∞
∞

e e (48)
∆t
(z − 1)2 e
= Z[ f {n∆t}∞


n=0 ] (49)
z ∆t
(z − 1)2 z ∆t
= =1 (50)
z ∆t (z − 1)2
8
where e
f is the unit ramp function. Equating (47) to its corresponding CTF z-transfer function, a recurrence relation
can be established for the response factors as follows:
Pm ∞
aφ,k z−k X
Pk=0
m −k
= φ[k]z−k , {aX } = {a}, {aY } = {b}, {aZ } = {c} (51)
k=0 dk z k=0
k−1

 1 X 
φ[k] = φ[ , k = 0, ..., m

a − j] d

φ,k

 k− j
d0



j=0
=⇒ 

(52)

m−1
1 X

 
φ[k] = φ[k + , k = m + 1, m + 2, ...

− − m j] d


 m− j
d0



j=0

3.3. Implementation Details

Polynomial additions and multiplications are performed numerically during the simplification processes of (40),
(41) and (42) without resorting to complicated analytical expansion formulas such as those based on elementary
symmetric polynomials. The linear algebra required to compute the Padé approximants and polynomial roots is
performed using the LAPACK routines dgesv and zhseqr. The former solves the linear system via LU-decomposition
while the latter exploits the upper Hessenberg structure of the polynomial-root linear system:

− α1 − α2 − α3 ... − αN−1 − αN 

 
 α0 α0 α0 α0 α0  
...

 1 0 0 0 0 

 
 0 1 0 ... 0 0  (53)
 .
 
.

 . 
 
..
 
0 0 0 1 0 

to compute its eigenvalues (whose multiplicative inverses are the roots themselves). The coefficients αk correspond to
αk x k .
PN
a polynomial of the form k=0

4. Case Studies and Validation

To validate the CDSE method and its associated implementation, values of CTF coefficients and response factors
are compared with those generated by the FDR method and reported in the ASHRAE Handbook (1997) using a
variety of test problems. The behavior of the response factors and CTF coefficients can be studied by comparing
the frequency characteristics of their associated transfer functions with exact theoretical values by using the L2 error
criterion defined by Chen et al. [16]:
v
u
t Nf
1 1 X
E= (ψi − ψ̄i )2 (54)
U N f i=1

where ψi , ψ̄i are the magnitudes of the s-domain external/cross/internal flow transfer functions subjected to a sinu-
soidal input s = jωi in the discrete, logarithmically-spaced set of frequencies Ω = {ω1 , ..., ωN f }, with ω1 = 10−8 and
ωN f being chosen according to the class of wall being considered (a very light wall might require ωN f = 10−2 , while
9
numerical behavior for a heavyweight wall would become meaningless beyond ωN f = 10−4 ). Frequency characteris-
tics of the z-transfer functions are recovered by setting z−1 = e− jωi ∆t . This information is also used to generate Bode
plots that help assess the region of convergence of the approximated polynomial transfer functions.
An important condition that must be satisfied by the z-transfer functions is that the ratio of the sums of the
numerator and denominator coefficients should be equal to the U-value:
Pm Pm Pm
k=0 ak k=0 ak ak
Pm = Pm = Pk=0
m =U (55)
d
k=0 k d
k=0 k k=0 dk

This condition is verified in Case Studies I and IV, and can be used in conjunction with the L2 error criterion
defined above to form quality control parameters for results generated by the present algorithm which, in terms of
tolerances, can be summarized as:
Pm
aφ,k
Pk=0
m < 1 , φ ∈ {X, Y, Z} (56)
k=0 dk

Eφ < 2 (57)

4.1. Case Study I: Brick/Cavity Wall

Description Physical Properties

L (mm) λ (W m−1 K−1 ) ρ (kg m−3 ) C p (J kg−1 K−1 ) R (m2 K W−1 )
Outside surface film 0.060
Brickwork 105 0.840 1700 800 0.125
Cavity 0.180
Heavyweight concrete 100 1.630 2300 1000 0.06135
Inside surface film 0.120

Table 1: Physical properties of the brick/cavity wall.

The brick/cavity wall test was considered by Xu et al. [14] when validating their improvement of the FDR method
with respect to the time domain method of Davies [10]. The physical properties of the wall are summarized in Table
1. Results of the CDSE method are tabulated and compared with those of the improved FDR method in Table 2.
Bode plots are used to visually assess the frequency characteristics of the z-transfer functions associated with these
coefficients with respect to the exact theoretical frequency characteristics over a frequency interval [10−8 , 10−3 ]; these
plots are displayed in Figure 1.
There is a clear agreement between CTF coefficients generated by the CDSE and improved FDR methods, with
minor differences on the order of 10−2 and smaller that can be accounted for by round-off errors in the physical
properties of the wall supplied as input. The Bode plots illustrate this strong agreement over the specified frequency
interval in regions where the approximated frequency characteristics agree with the theoretical ones and in regions
where they begin to diverge from exact values. Agreement in both sets of regions indicates that the polynomial s-
transfer functions estimated via Padé approximants are nearly identical to those generated by least squares regression.
10
 P
k 0 1 2 3 4 5
aak 9.547772 -18.534215 10.584111 -1.585679 0.065340 -0.000761 0.076568
abk 9.589008 -18.586934 10.586818 -1.560687 0.049680 -0.001181 0.076703

bak 0.000179 0.013914 0.043449 0.018001 0.001021 0.000005 0.076568

bbk 0.000178 0.013914 0.043475 0.018078 0.001052 0.000006 0.076703

cak 6.953532 -12.228572 5.995942 -0.665431 0.021226 -0.000128 0.076568

cbk 6.959635 -12.226366 5.979019 -0.653456 0.018078 -0.000207 0.076703

dka 1.000000 -1.621649 0.727483 -0.065668 0.001670 -0.000004 0.041833

dkb 1.000000 -1.619844 0.724520 -0.064306 0.001543 -0.000006 0.041906

Table 2: Comparison of CTF coefficients for the brick/cavity wall generated by the CDSE and FDR methods.
a
Complex domain series expansion method.
b
Frequency-domain regression method.

4.2. Case Study II: Comparison with ASHRAE Handbook (1997)

The 1997 ASHRAE Handbook [9] provides a tabulation of CTF coefficients for 42 roof and 41 wall constructions,
converted to SI from Harris and McQuiston’s original study on the categorization of roofs and walls based on thermal
response [5]. For each multilayer element, 7 coefficients for the cross flow z-transfer function numerator and denom-
inator are given. These conductive elements serve as an excellent basis of comparison, as they cover a wide weight
range.
The results of a test run over all roofs and walls are tabulated in Table 3. The L2 error criterion is invoked to
compare the frequency characteristics of the cross flow z-transfer function reported by the Handbook and generated
by the present method with exact theoretical characteristics. A frequency interval of [10−8 , 10−3 ] is specified and
discretized with N f = 100 for the purposes of computing these frequency characteristics, and m = 6 is fixed throughout
the test run.
It is clear that the CDSE method is capable of computing accurate z-transfer functions over the full range of wall
weights, with the largest L2 error among both roofs and walls being 3.43%. The error is also seen to decrease as
wall weight increases (heavier roofs and walls are generally found nearer to the end of the Handbook tables, however,
the order is not strict). For lighter roofs and walls, the L2 errors for ASHRAE’s coefficients and those generated by
the present algorithm are seen to be nearly equal. Beyond the first few roofs and walls, the errors for the CDSE
method remain below 1% consistently. For heavier walls, ASHRAE’s z-transfer functions begin to fail, with notably
high errors being associated with Walls 4, 34, 37 and 38 (whose errors are 57.73%, 64.19%, 21.99% and 24.24%,
respectively). The present solver’s capability of handling all given roofs and walls demonstrates robustness with

11
respect to wall weight.
A test run with m = 10 verified that the linear systems remain well-conditioned, as it was found that the L2 errors
were nearly the same as those computed with m = 6. This also indicates that a default choice of m = 6 will likely be
able to hand roofs and walls encountered in practice.

Roofs Walls
No. ASHRAE CDSE No. ASHRAE CDSE No. ASHRAE CDSE No. ASHRAE CDSE
1 2.87% 2.87% 22 0.07% 0.07% 1 3.43% 3.43% 22 0.14% 0.09%
2 2.24% 2.24% 23 3.96% 0.06% 2 1.22% 1.22% 23 0.66% 0.07%
3 0.98% 0.98% 24 0.12% 0.05% 3 1.05% 1.05% 24 0.13% 0.06%
4 0.59% 0.59% 25 0.23% 0.04% 4 57.73% 1.19% 25 0.64% 0.05%
5 0.57% 0.57% 26 0.20% 0.09% 5 0.49% 0.49% 26 0.10% 0.04%
6 0.60% 0.60% 27 0.06% 0.06% 6 0.43% 0.43% 27 0.65% 0.03%
7 0.49% 0.49% 28 0.20% 0.05% 7 0.40% 0.40% 28 0.35% 0.03%
8 0.41% 0.41% 29 0.14% 0.04% 8 0.25% 0.25% 29 1.06% 0.02%
9 0.35% 0.34% 30 1.20% 0.04% 9 0.26% 0.26% 30 2.62% 0.05%
10 0.23% 0.23% 31 0.22% 0.03% 10 0.21% 0.21% 31 4.59% 0.05%
11 0.19% 0.19% 32 0.70% 0.03% 11 0.18% 0.18% 32 2.44% 0.04%
12 0.18% 0.18% 33 1.31% 0.02% 12 0.17% 0.16% 33 0.51% 0.03%
13 0.19% 0.19% 34 0.36% 0.02% 13 0.14% 0.14% 34 64.19% 0.02%
14 0.15% 0.15% 35 1.43% 0.04% 14 0.17% 0.10% 35 12.30% 0.27%
15 0.15% 0.10% 36 0.60% 0.03% 15 0.23% 0.15% 36 0.75% 0.02%
16 0.08% 0.08% 37 0.69% 0.03% 16 0.12% 0.11% 37 21.99% 0.03%
17 0.20% 0.19% 38 4.70% 0.02% 17 0.85% 0.10% 38 24.24% 0.03%
18 0.37% 0.11% 39 1.16% 0.02% 18 0.28% 0.08% 39 9.73% 0.02%
19 0.13% 0.13% 40 0.39% 0.02% 19 0.25% 0.07% 40 10.71% 0.02%
20 0.12% 0.11% 41 3.09% 0.01% 20 0.37% 0.06% 41 8.83% 0.01%
21 0.28% 0.09% 42 6.58% 0.01% 21 0.15% 0.04% - - -

Table 3: A comparison of the L2 error norms associated with the CTF coefficients reported by ASHRAE and generated by the CDSE method (with
m = 6) for a set of 42 roof and 41 wall constructions.

4.3. Case Study III: A Heavyweight Wall

The physical properties of a heavyweight wall reported to be commonly used in China [16] are displayed in Table
5. The response factors for cross heat flow computed by the CDSE method with m = 6 are compared with the set
generated by the FDR method and reported by Chen et al. by visual inspection of the coefficient values (the first
72 of which are listed in Table 6) and the Bode plots of their corresponding z-transfer functions over a frequency

12
range [10−8 , 10−3 ] displayed in Figure 2 (curves of the CTF z-transfer function are included for reference). 144-term
response factor sequences are used to generate the Bode plots. Additionally, the L2 error for the z-transfer functions
constructed by both sets of factors are tabulated in Table 4 under varying truncation limits. The error terms are
evaluated using an interval [10−9 , 10−3 ] and N f = 50.

k Eka Ekb
72 9.125% 7.065%
96 3.740% 2.905%
120 1.535 1.195%
144 0.632% 0.492%

Table 4: Comparison of L2 errors for response factor z-transfer functions of varying truncation limits for the CDSE and FDR methods.
a
Complex domain series expansion method.
b
Frequency-domain regression method.

The z-transfer function curves are seen to agree well with the theoretical frequency characteristics over [10−8 , 10−3.2 ].
Beyond this region, all phase curves begin to diverge from the theoretical curve (on the other hand, magnitude curves
continue to vanish). The CDSE and FDR response factor sequences are very similar (with errors on the order of 0.001
due to round-off errors in the wall parameters) except for the first four terms, the variance of which is due to sensi-
tivity in the computation of the CTF z-transfer function. Since the first term obtained from the recurrence relation
between the CTF and response factor z-transfer functions will always be Y[0] = b0 /d0 = b0 , round-off errors in the
wall parameters will be more apparent in the initial few sequence terms. The L2 error terms reveal that the variation
in the coefficient values does not impact accuracy of the overall sequence, with errors for the CDSE method turning
out to be slightly smaller than those of the FDR method (which, for confirmation, were found to be nearly identical to
those reported in [16]).

Description Physical Properties

L (mm) λ (W m−1 K−1 ) ρ (kg m−3 ) C p (J kg−1 K−1 ) R (m2 K W−1 )
Outside surface film 0.0538
Common brick 370 0.814 1800 879
Foam concrete 100 0.209 600 837
Wood wool board 25 0.163 400 2093
Stucco 20 0.814 1600 837
Inside surface film 0.1147

Table 5: Physical properties of the heavyweight wall.

13
4.4. Case Study IV: Varying the Time Step

Although hourly analysis (i.e. with a time step of ∆t = 3600 s) is standard for heat load calculations in the building
services, other contexts may involve analyses in which smaller time steps are required (such as control simulations
of air conditioning systems). Wang et al. [15] investigated the performance of the FDR method as the time step was
shortened, noting that previously established methods such as Direct Root Finding and State Space required a larger
number of coefficients to sufficiently simulate heat conduction when ∆t was made small enough, leading to unstable
behaviour in the resulting z-transfer functions. CTF coefficients for Wall Group 2 of the ASHRAE Handbook (whose
physical properties are summarized in Table 7) were reported for time steps ∆t = 3600, 1800, 1200, 900, 600, 300
and 60 s. CTF coefficients computed using the CDSE method with a choice of m = 5 are tabulated along side
those generated by the FDR method in Table 8, with results for time steps ∆t = 1200 and 900 s omitted for brevity.
The frequency characteristics of corresponding z-transfer functions are represented by Bode plots in Figure 3 in
comparison with theoretical values.
A visual inspection of the Bode plots reveals that the CDSE method remains accurate when the time step is
reduced. In fact, the frequency characteristics of the CTF z-transfer functions are more inline with theoretical values
when the time step is smaller. The L2 errors for time steps 3600, 1800, 600, 300 and 60 s were found to be 1.174%,
0.305%, 0.0345%, 8.622e-3% and 3.457e-4%, respectively. For time steps smaller than 60 s, the error begins to
increase rapidly. For example, the error for ∆t = 10s was found to be 1.819%. There is a general agreement between
the CTF coefficients computed using the FDR and CDSE methods, with the most significant variations appearing
strictly in the bk coefficients. U-values associated with both sets of coefficients are seen to match the true U-value of
the Wall Group.

5. Conclusion

An algorithm has been presented for computing response factors and CTF coefficients based on series expan-
sions of the solutions of the governing equations in the s-domain. This approach was motivated by the recently
developed frequency-domain regression approach in that external, cross and internal heat flow transfer functions are
approximated with rational functions that simplify the Laplace inversion process. However, the rational functions are
estimated using Padé approximants rather than solving a weighted least squares problem. The method is mathemat-
ically straightforward, and its open source C++ implementation only requires two FORTRAN routines to perform the
necessary linear algebra.
A set of case studies were chosen to study the performance and results of the CDSE method, which in many
ways can be directly compared with those of the FDR method due to the fact that the two methods generate ’optimal’
polynomial approximations of the systems’ transfer functions. The behaviour of the z-transfer functions was found
to be nearly identical in both methods, regardless of the choice of time step ∆t, however, some differences in the
coefficient values (especially in the bk ) was detected. Nonetheless, comparisons of z-transfer function frequency

14
characteristics with theoretical values and with characteristics of other results reported in ASHRAE’s Handbook and
in the literature demonstrate a high degree of accuracy and robustness with respect to the potential range of wall
weights encountered in practice.

6. Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit
sectors.

7. Acknowledgements

I would like to thank Kassem Jaber for assisting me with technical details regarding the dynamic system analysis
and in assessing the overall composition of this paper.

References

[1] Brisken WR, Reque SG. Heat load calculations by thermal response. ASHVE Transactions 1956;62:391–424.
[2] D. G. Stephenson and G. P. Mitalas, ”Cooling load calculations by thermal response factor method,” ASHRAE Transactions, vol. 73, (1), pp.
1-7, 1967.
[3] G. P. Mitalas and D. G. Stephenson, Room Thermal Response Factors. United States: American Society of Heating, Refrigerating, and Air
Conditioning Engineers,New York, 1967.
[4] G. P. Mitalas, ”Calculation of transient heat flow through walls and roofs,” ASHRAE Transactions, vol. 74, (2), pp. 182-88, 1969.
[5] S. M. Harris and F.C McQuiston, ”A Study to Categorize Walls and Roofs on the Basis of Thermal Response.” , ProQuest Dissertations
Publishing, 1988.
[6] Pedersen, C.O., D.E. Fisher, R.J. Liesen. 1997. Development of a Heat Balance Procedure for Calculating Cooling Loads, ASHRAE Trans-
actions, Vol. 103, Pt. 2, pp. 459-468.
[7] Stephenson DG, Mitalas GP. Calculation of heat conduction transfer functions for multilayer slabs. ASHRAE Transaction 1971;77(2):117–26.
[8] J. D. Spitler, D. E. Fisher and C. O. Pedersen, ”The Radiant Time Series Cooling Load Calculation Procedure,” ASHRAE Transactions, vol.
103, pp. 503, 1997.
[9] American Society of Heating, Refrigerating and Air-Conditioning Engineers, ASHRAE Handbook: 1997 Fundamentals. (Inch-Pound ed.)
Atlanta GA: ASHRAE, 1997.
[10] M. G. Davies, ”Wall transient heat flow using time-domain analysis,” Building and Environment, vol. 32, (5), pp. 427-446, 1997.
[11] J. D. Spitler and D. E. Fisher, ”On The Relationship between the Radiant Time Series and Transfer Function Methods for Design Cooling
Load Calculations,” HVAC&R Research, vol. 5, (2), pp. 123-136, 1999.
[12] Y. Chen and S. Wang, ”Frequency-domain regression method for estimating CTF models of building multilayer constructions,” Applied
Mathematical Modelling, vol. 25, (7), pp. 579-592, 2001.
[13] S. Wang and Y. Chen, ”Transient heat flow calculation for multilayer constructions using a frequency-domain regression method,” Building
and Environment, vol. 38, (1), pp. 45-61, 2003.
[14] X. Xu, S. Wang and Y. Chen, ”An improvement to frequency-domain regression method for calculating conduction transfer functions of
building walls,” Applied Thermal Engineering, vol. 28, (7), pp. 661-667, 2008.
[15] J. Wang et al, ”Short time step heat flow calculation of building constructions based on frequency-domain regression method,” International
Journal of Thermal Sciences, vol. 48, (12), pp. 2355-2364, 2009.

15
[16] Y. Chen, J. Zhou and J. D. Spitler, ”Verification for transient heat conduction calculation of multilayer building constructions,” Energy and
Buildings, vol. 38, (4), pp. 340-348, 2006.
[17] F. Varela et al, ”A direct numerical integration (DNI) method to obtain wall thermal response factors,” Energy and Buildings, vol. 81, pp.
363-370, 2014.
[18] H. Wang and Z. Zhai, ”Advances in building simulation and computational techniques: A review between 1987 and 2014,” Energy and
Buildings, vol. 128, pp. 319-335, 2016.
[19] J. Sanza Pérez et al, ”A new method for calculating conduction response factors for multilayer constructions based on frequency–Domain
spline interpolation (FDSI) and asymptotic analysis,” Energy and Buildings, vol. 148, pp. 280-297, 2017.
[20] Klein SA, Beckman WA, Mitchell JW, et al. TRNSYS—a transient system simulation program. Solar Energy Laboratory, University of
Wisconsin–Madison, Madison, USA, July 1994.
[21] BLAST: Building loads analysis and system thermodynamic. http://www.bso.uiuc.edu/blastmain.htm.
[22] Strand R, Winkelmann F, Buhl F, Huang J, Liesen R, Pedersen C, Fisher D, Taylor R, Crawley D, Lawrie L. Enhancing and extending the
capabilities of the building heat balance simulation technique for use in energyplus. Proceedings of Building Simulation ’99, vol. II, Kyoto,
Japan, September 1999. p. 653–7.

16
 12.0 0.7
Theoretical Theoretical
11.0 CDSE CDSE
−1

FDR 0.6 FDR

10.0
K

Argument, radian
9.0 0.5
−2

8.0
Magnitude, W m

0.4
7.0

6.0
0.3
5.0

4.0 0.2

3.0
0.1
2.0

1.0 0.0
10−8 10−7 10−6 10−5 10−4 10−3 10−8 10−7 10−6 10−5 10−4 10−3
Frequency, radian s
−1 Frequency, radian s
−1

(a)

2.0 3.0

1.8
−1

2.0
K

1.6
Argument, radian
−2

1.4
1.0
Magnitude, W m

1.2

1.0 0.0

0.8
-1.0
0.6

0.4
Theoretical -2.0 Theoretical
0.2 CDSE CDSE
FDR FDR
0.0 -3.0
10−8 10−7 10−6 10−5 10−4 10−3 10−8 10−7 10−6 10−5 10−4 10−3
Frequency, radian s
−1 Frequency, radian s
−1

(b)

8.0 0.7
Theoretical Theoretical
CDSE CDSE
−1

7.0 FDR 0.6 FDR

K

Argument, radian

6.0 0.5
−2
Magnitude, W m

5.0 0.4

4.0 0.3

3.0 0.2

2.0 0.1

1.0 0.0
10−8 10−7 10−6 10−5 10−4 10−3 10−8 10−7 10−6 10−5 10−4 10−3
Frequency, radian s
−1 Frequency, radian s
−1

(c)

Figure 1: External (a), cross (b) and internal (c) flow Bode diagrams for the brick/cavity wall of Case Study I.

17
k Yka Ykb k Yka Ykb
0 -0.0000263957 0.0000196345 36 0.0121633014 0.0121727560
1 0.0000675681 0.0000108668 37 0.0117379521 0.0117479420
2 -0.0000468852 0.0000040855 38 0.0113253164 0.0113356370
3 0.0000004003 0.0000231081 39 0.0109254660 0.0109359310
4 0.0001465373 0.0001405920 40 0.0105383625 0.0105488070
5 0.0006073107 0.0006138580 41 0.0101638835 0.0101741660
6 0.0016778107 0.0016942230 42 0.0098018425 0.0098118420
7 0.0033991044 0.0034144010 43 0.0094520055 0.0094616240
8 0.0055920679 0.0056038820 44 0.0091141032 0.0091232590
9 0.0079971773 0.0080073320 45 0.0087878415 0.0087964740
10 0.0103774754 0.0103869760 46 0.0084729095 0.0084809710
11 0.0125625475 0.0125711980 47 0.0081689851 0.0081764420
12 0.0144529693 0.0144606400 48 0.0078757406 0.0078825730
13 0.0160069035 0.0160140340 49 0.0075928461 0.0075990410
14 0.0172224324 0.0172296410 50 0.0073199722 0.0073255280
15 0.0181219045 0.0181294380 51 0.0070567927 0.0070617130
16 0.0187402231 0.0187477510 52 0.0068029861 0.0068072800
17 0.0191169751 0.0191237540 53 0.0065582369 0.0065619190
18 0.0192915961 0.0192967990 54 0.0063222362 0.0063253250
19 0.0193006888 0.0193036980 55 0.0060946828 0.0060971980
20 0.0191767574 0.0191773220 56 0.0058752838 0.0058772480
21 0.0189478135 0.0189460690 57 0.0056637540 0.0056651920
22 0.0186374833 0.0186338750 58 0.0054598172 0.0054607550
23 0.0182653747 0.0182605400 59 0.0052632054 0.0052636670
24 0.0178475567 0.0178422040 60 0.0050736593 0.0050736720
25 0.0173970623 0.0173918750 61 0.0048909278 0.0048905160
26 0.0169243691 0.0169199390 62 0.0047147683 0.0047139590
27 0.0164378325 0.0164346180 63 0.0045449464 0.0045437640
28 0.0159440641 0.0159423810 64 0.0043812354 0.0043797050
29 0.0154482543 0.0154482800 65 0.0042234167 0.0042215610
30 0.0149544418 0.0149562290 66 0.0040712792 0.0040691220
31 0.0144657381 0.0144692410 67 0.0039246191 0.0039221830
32 0.0139845107 0.0139896060 68 0.0037832398 0.0037805450
33 0.0135125326 0.0135190440 69 0.0036469516 0.0036440200
34 0.0130511042 0.0130588220 70 0.0035155716 0.0035124220
35 0.0126011508 0.0126098510 71 0.0033889233 0.0033855740

Table 6: The first 71 response factors for the heavyweight wall generated by the CDSE and FDR methods.
a Complex domain series expansion method.
b Frequency-domain regression method.

18
 0.8 4.0

0.7 3.0
Magnitude, W m−2 K−1

Argument, radian
0.6 2.0

0.5 1.0

0.4 0.0

0.3 -1.0

0.2 -2.0
Theoretical Theoretical
FDR FDR
0.1 CDSE -3.0 CDSE
CDSE (CTF) CDSE (CTF)
0.0 -4.0
10−8 10−7 10−6 10−5 10−4 10−3 10−8 10−7 10−6 10−5 10−4 10−3
Frequency, radian s−1 Frequency, radian s−1

Figure 2: Cross flow Bode diagrams for the heavyweight wall of Case Study III.

Description Physical Properties

L (mm) λ (W m−1 K−1 ) ρ (kg m−3 ) C p (J kg−1 K−1 ) R (m2 K W−1 )
Outside surface film 0.060
Stucco 25 0.692 1858 840 0.036
Insulation 125 0.043 91 840 2.907
Plaster or gypsum 20 0.727 1602 840 0.028
Inside surface film 0.120

Table 7: Physical properties of ASHRAE Wall Group 2.

0.3 4.0
−1

0.3 3.0
K

Argument, radian

0.2 2.0
−2
Magnitude, W m

0.2 1.0

0.2 0.0

Theoretical Theoretical
0.1 ∆t = 3600 -1.0 ∆t = 3600
∆t = 1800 ∆t = 1800
0.1 ∆t = 600 -2.0 ∆t = 600
∆t = 300 ∆t = 300
∆t = 60 ∆t = 60
0.0 -3.0
10−8 10−7 10−6 10−5 10−4 10−3 10−8 10−7 10−6 10−5 10−4 10−3
Frequency, radian s
−1 Frequency, radian s
−1

(a)

Figure 3: Cross flow Bode diagrams for Wall Group 2 of Case Study IV.

19
k 0 1 2 3 4 5 U
Time step ∆t = 3600 s
bak 9.238270E-04 3.134899E-02 5.424187E-02 1.188745E-02 2.738740E-04 2.703650E-07 0.317398
bbk 9.356678E-04 3.165971E-02 5.460031E-02 1.193587E-02 2.768229E-04 3.041151E-07 0.317479
cak 4.964126E+00 -7.538352E+00 3.020772E+00 -3.499972E-01 2.131200E-03 -3.968320E-06 0.317398
cbk 4.950058E+00 -7.483352E+00 2.971137E+00 -3.401356E-01 1.706114E-03 -5.092724E-06 0.317479
dka 1.000000E+00 -9.408329E-01 2.774543E-01 -2.585314E-02 1.226296E-04 -2.377552E-08 * -
dkb 1.000000E+00 -9.355651E-01 2.738394E-01 -2.526493E-02 1.095122E-04 -3.532504E-08 -
Time step ∆t = 1800 s
bak 3.523998E-06 1.834509E-03 1.144690E-02 9.920927E-03 1.424604E-03 2.253814E-05 0.317398
bbk 4.727864E-06 1.853722E-03 1.158368E-02 1.000414E-02 1.438610E-03 2.348765E-05 0.317479
cak 6.104093E+00 -1.284061E+01 8.882510E+00 -2.266845E+00 1.476614E-01 -2.155427E-03 0.317398
cbk 6.080903E+00 -1.275364E+01 8.788538E+00 -2.234344E+00 1.460655E-01 -2.616704E-03 0.317479
dka 1.000000E+00 -1.730097E+00 1.026203E+00 -2.322156E-01 1.393706E-02 -1.541931E-04 -
dkb 1.000000E+00 -1.725393E+00 1.020708E+00 -2.306784E-01 1.400800E-02 -1.879496E-04 -
Time step ∆t = 600 s
bak 1.888654E-05 -9.879201E-05 2.288415E-04 -3.73172E-05 4.788933E-04 1.967507E-04 0.317398
bbk 1.445008E-05 -7.594148E-05 1.830175E-04 7.711732E-06 4.688070E-04 1.867713E-04 0.317536
cak 7.154002E+00 -2.330286E+01 2.899612E+01 -1.704157E+01 4.666951E+00 -4.718573E-01 0.317398
cbk 7.120910E+00 -2.327280E+01 2.911951E+01 -1.727687E+01 4.812082E+00 -5.020521E-01 0.315547
dka 1.000000E+00 -3.099307E+00 3.687662E+00 -2.082222E+00 5.499704E-01 -5.362348E-02 -
dkb 1.000000E+00 -3.110953E+00 3.723589E+00 -2.123080E+00 5.701970E-01 -5.728142E-02 -
Time step ∆t = 300 s
bak -8.604667E-06 8.803074E-05 -3.226204E-04 5.804443E-04 -5.418111E-04 2.512651E-04 0.317398
bbk -1.127836E-05 9.493795E-05 -3.165406E-04 5.426042E-04 -4.961109E-04 2.324727E-04 0.317479
cak 7.479679E+00 -2.937890E+01 4.551999E+01 -3.472326E+01 1.301909E+01 -1.916569E+00 0.317398
cbk 7.444271E+00 -2.934634E+01 4.568333E+01 -3.506161E+01 1.325239E+01 -1.972001E+00 0.317479
dka 1.000000E+00 -3.840745E+00 5.826007E+00 -4.355829E+00 1.602282E+00 -2.315674E-01 -
dkb 1.000000E+00 -3.855263E+00 5.876049E+00 -4.420216E+00 1.638910E+00 -2.393354E-01 -
Time step ∆t = 60 s
bak -1.550119E-04 8.435387E-04 -1.841489E-03 2.016482E-03 -1.107979E-03 2.444852E-04 0.317398
bbk -1.426411E-04 7.742094E-04 -1.686415E-03 1.843344E-03 -1.011445E-03 2.229736E-04 0.317478
cak 7.769836E+00 -3.683664E+01 6.981223E+01 -6.611053E+01 3.128182E+01 -5.916703E+00 0.317369
cbk 7.733101E+00 -3.670617E+01 6.965236E+01 -6.604659E+01 3.129529E+01 -5.927998E+00 0.317486
dka 1.000000E+00 -4.721551E+00 8.911928E+00 -8.405527E+00 3.961488E+00 -0.746338E-01 -
dkb 1.000000E+00 -4.727177E+00 8.933721E+00 -8.437176E+00 3.981911E+00 -7.512792E-01 -

Table 8: Comparison of CTF coefficients for Wall Group 2 generated by the CDSE and FDR method with various choices of time step ∆t.
a
Complex domain series expansion method.
b
Frequency-domain regression method.

20