Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/271140988
READS
91
5 AUTHORS, INCLUDING:
Bjrn Lundberg
Jonas Sjblom
3 PUBLICATIONS 5 CITATIONS
SEE PROFILE
SEE PROFILE
Derek Creaser
Chalmers University of Technology
60 PUBLICATIONS 1,071 CITATIONS
SEE PROFILE
Department of Chemical and Biological Engineering, Chemical Reaction Engineering, Chalmers University of Technology, Gteborg SE-412 96, Sweden
Department of Applied Mechanics, Division of Combustion, Chalmers University of Technology, Gteborg SE-412 96, Sweden
Johnson Matthey AB, Vstra Frlunda SE-421 31, Sweden
d
Scania AB, Sdertlje SE-151 87, Sweden
b
c
a r t i c l e
a b s t r a c t
i n f o
Article history:
Received 4 August 2014
Received in revised form 4 December 2014
Accepted 7 January 2015
Available online 19 January 2015
Keywords:
Parameter estimation
D-optimal design
Diesel Oxidation Catalyst
Multivariate Data Analysis
Engine rig experiments
In the current study a parameter estimation method based on data screening by sensitivity analysis
is presented. The method applied Multivariate Data Analysis (MVDA) on a large transient data set to
select different subsets on which parameters estimation was performed. The subset was continuously
updated as the parameter values developed using Principal Component Analysis (PCA) and D-optimal
onion design. The measurement data was taken from a Diesel Oxidation Catalyst (DOC) connected to a
full scale engine rig and both kinetic and mass transport parameters were estimated. The methodology
was compared to a conventional parameter estimation method and it was concluded that the proposed
method achieved a 32% lower residual sum of squares but also that it displayed less tendencies to converge
to a local minima. The computational time was however signicantly longer for the evaluated method.
2015 Elsevier Ltd. All rights reserved.
1. Introduction
Common kinetic mechanisms involve a large number of parameters. These mechanisms and the complex interaction between
different reactions often result in a high correlation between
the kinetic parameters (Bernaerts et al., 2000; Franceschini and
Macchietto, 2008b). This severely complicates the parameter
estimation process and increases the importance of Design of
Experiments (DoE) and the parameter estimation algorithm. DoE
in reaction kinetics has a long tradition reaching back to the pioneering work by Box and Lucas (1959), Box and Hunter (1965),
Draper and Hunter (1966, 1967a,b), Box (1968), Hill et al. (1968),
and Hunter et al. (1969). The focus has however been generally
more directed toward steady-state systems rather than transient
ones (Franceschini and Macchietto, 2008a) even though the need
and benets of transient experiments have been demonstrated
(Goodwin, 1977; Berger et al., 2008). The system studied in the
current paper is a reactor model of a full scale Diesel Oxidation
Catalyst (DOC) connected to a Heavy Duty Diesel engine rig. The
system is transient and the model parameters highly correlated;
DoE is therefore of highest importance.
A wide range of catalyst models with varying level of detail can
be found in literature. Generally, the kinetics are either described
local minimum close to the initial guess and thereby failing to converge to the global minima (Blockley, 2010). The global methods
generally evaluate a larger parameter space which means that they
are more likely to nd the global optima. The main drawback with
the global methods is very high computational cost which means
that they often are not realistic to use for complex systems with
many parameters. To be successful the global algorithms also are
more dependent on robust models since a wide parameter span will
be needed to be evaluated. Examples from literature where global
optimization algorithms have been applied in catalysis modeling
include Glielmo and Santini (1999) who used the Genetic Algorithm
(an evolutionary algorithm) and Ramanathan and Sharma (2011)
who showed that an improved t could be achieved if a global algorithm was followed by a local algorithm to increase the accuracy.
A method of reducing the tendencies of the local methods to converge to a local minimum close to the initial guess may therefore
be the more desirable way of improving parameter optimization
for the current application.
In our previous study (Lundberg et al., 2014), an experimental
plan that included several different DOC catalyst congurations,
was applied to a full scale engine rig system and used for parameter tuning. The best model t was achieved with a model that
included internal transport resistance with the effective diffusivities of the species being tuned in addition to the kinetic parameters.
In the current study, the data and catalyst model formulation will be
re-used but with a new approach to parameter tuning will be presented and evaluated. The method is based on a study by Sjoblom
and Creaser (2008) who developed methods for model-based DoE
including transient data. The methodology presented by Sjblom
and Creaser includes Multivariate Data Analysis (MVDA) to reduce
the dimensionality of the largely correlated Jacobian and subsequent selection of a D-optimal sub-set of parameters to estimate
based on the parameter sensitivity of simulation data. The relatively
simple global kinetic model in the current study makes the number
of parameters to estimate less of a concern. Instead the method was
revised to select a set of time points from the large transient data set
with high sensitivity for parameter changes and at the same time
retaining a good representation of the entire data set. To achieve
these properties the D-optimal onion design, introduced by Olsson
et al. (2004), was selected as an algorithm for data point selection. The algorithms combine the D-optimal design, which by itself
tends to only select the most extreme points (Pinto et al., 1990;
Zullo, 1991), with a space lling approach making the selection
more evenly distributed over the entire data set. Parameter estimation was performed based only in the selected time points with
the aim of making use of improved statistical properties to improve
the overall t to measurement data. Since the parameter sensitivity, and thereby the data selection, will depend on the parameter
values, the parameter estimation was performed in an iterative process where the data point selection was updated as new parameter
values were found. This change of data points used for parameter
tuning may also help to avoid local optima. As an evaluation, the
efciency and resulting t to measurement data achieved by the
presented method will be compared to the results of a traditional
method of parameter estimation.
2. Experimental
The most important characteristics of the reactor model and
adjustable parameters are summarized in the current section and
for completeness the dening equations are given in Appendix. A
thorough description of the experimental set-up, measurements
and engine operation design is given in the preceding study
(Lundberg et al., 2014).
145
k1 yCO yO2
(1)
G(yi , Ts )
k2 yC3 H6 yO2
G(yi , Ts )
k3 yNO yO2
G(yi , Ts )
(2)
1
K
Kp
(3)
yNO2
(4)
1/2
yNO yO
2
G(yi , Ts ) = Ts (1 + K4 yCO + K5 yC3 H6 )2 (1 + K6 yCO
yC2
3 H6
0.7
)(1 + K7 yNO
)
(5)
where Kj is the reaction rate coefcient for the inhibition terms
in the denominator G and Kp is the equilibrium constant for NO
oxidation. At thermodynamic equilibrium, Kp will be equal to K and
reaction rate r3 will be equal to zero. Both the rate coefcients kj and
the inhibition terms Kj were described by Arrhenius expressions:
kj = Aj eEA,j /RTs
(6)
146
(7)
147
Table 1
Catalyst conguration used for parameter estimation (ad). Catalyst conguration c consists of two catalysts in series. The Pt-loading is shown in both mass% of washcoat
and g/ft3 total catalyst volume.
Conguration
a
b
c
d
0.30
0.59
0.30
0.59
15
30
15
15
The scores matrix gives a simplied description of the parameter sensitivity of all data points given in the Jacobian. The scores
matrix has fewer columns than the Jacobian but the same number of rows, see Appendix for details. In the third step a D-optimal
onion design is performed on the resulting scores matrix giving a
selection of data points that will have a more equal inuence from
the different parameters. The reduced number of columns in the
scores matrix compared to the Jacobian will reduce the computational time of nding a D-optimal design which would be very high
if the Jacobian would be used directly. The method is described in
detail in Appendix and is schematically summarized in Fig. 2.
Length [cm]
10.2
10.2
2 10.2
10.2
0.110
0.110
0.110
0.055
yi,in
yi,in yi,out
yi,out
yi,in
yi,in
yi,out yi,out
res =
yi,out is the
where wi is a weighting factor (further described below),
simulated outlet mole fraction, yi,out is the measured outlet mole
fraction and yi,in is the measured inlet mole fraction. To compensate
for the fact that the NO conversion is limited by thermodynamic
constraints the residual for NO was instead calculated relative to the
w =
yi,in
wi (8)
Original parameters
Parameter sensivity
calculaons
(at all me points)
Jacobian matrix
PCA
Scores matrix
Parameter tuning
(at selected me points)
No
Improved parameter
values?
Yes
Fig. 2. Summary of the method used for parameter estimation. Improved parameter values means that parameters were found that reduced the residual sum of squares at
the selected time points which will result in a right hand loop.
148
res =
(yNO,in yNO,out,lim ) (
(yNO,in yNO,out,lim ) (yNO,out yNO,out,lim )
yNO,out yNO,out,lim )
yNO,in yNO,out,lim
yNO,in yNO,out,lim
(10)
where ni is the number of data points used for residual calculation for species i after points with close to 100% conversion and
points from the rst 120 s of the experimental data sets have been
removed.
3.2. Parameter tuning
After data selection using D-optimal onion design a parameter estimation with the gradient search method is undertaken for
only the selected time points. The gradient search method is very
efcient for linear systems but can also be applied for nonlinear
systems such as catalyst models. For a nonlinear system the residual function is rst linearized for all parameters and then a step
in the parameter space is made in the direction of the steepest
descent (lowest residual). This process is usually repeated until
the change in residual is below a certain tolerance. The gradient
search method of choice in this work is the trust-region-reective
method (Coleman and Li, 1996) which is the standard method in
Matlab, the software used for parameter estimation in this project,
for over-determined nonlinear least square problems (function
call lsqnonlin). In theory the proposed methodology with PCA
and D-optimal design should make a gradient search more efcient compared to if all data points were used, since the correct
direction for parameter adjustment should be better dened by a
well-balanced subset of more sensitive data points.
Since the properties of the sensitivity matrix (Jacobian in all data
points) are dependent on the actual parameter values, the best
selection of time points for parameter tuning will change as the
parameters are tuned. To make the parameter tuning efcient it
wNO =
yNO,out yNO,out
wNO
yNO,in yNO,out,lim
(9)
149
Table 2
Residual sum square of each component and complete data set together with the
summation of residual sum square (rightmost column) for the different methods.
Start
1 step
5 steps
Ref
NO
HC
CO
Sum
18,986
515
558
1238
9736
1161
1193
1173
16,954
730
742
1169
45,675
2405
2494
3580
the residuals in Fig. 4 still show a declining trend even when both
methods have reached residual sum of squares below 2500.
Although the rate of decline appears to be decreasing, no clear
optima appear to have been reached.
4.3. Comparison to measurement data
The parameter values that gave the lowest residual sum of
squares, shown in Fig. 4, shall be regarded as the nal tuning results
of the methods. A compilation of the residual sum of sqares for the
nal results and the starting paramter values are shown for every
gas phase component in Table 2.
From these results it appears that HC was the most difcult component to t to measurement data for the MVDA methods which
may have several different reasons. Firstly HC and CO concentrations are highly correlated in the exhaust of a HDD engine which
means that it can be difcult to generate experimental data, let
alone subsets of data, where the parameters that inuence each
component can be distinguished. Since both components have high
concentrations at similar time points, the larger residual for CO at
the starting parameters may have inuenced the parameter estimation to initially focus more on reducing the CO residual at the
cost of decreased sensitivity for HC concentration. HC also differs from NO and CO since it is modeled as an average of a wide
range of hydrocarbons. The physical properties are more difcult
to model and the HC composition at the catalyst inlet will even
change between different engine operating points. For the reference case, the HC residual was at about the same level as for
the MVDA methods but both NO and CO residuals are noticeably higher. The results in Table 2 also indicate that both MVDA
methods achivieved a very similar nal t, with HC being the
largest residual, but that the reference method differs because the
total residual is more equally distibuted over the three components.
Two examples of the nal results of the tuning of the kinetic and
mass transport parameters for the different parameter estimation
methods are shown in Figs. 5 and 6.
Fig. 5 shows a very similar t for the two MVDA methods for all
components. The reference method shows a similar t for CO and
HC but for NO the transient was poorly captured. The NO production
before the change in engine operating point is likely a result of
NO2 reduction by CO or HC. These reactions are not included in
the kinetic model and can therefore not be captured by any of the
models.
Fig. 6 also shows a very similar t for the two MVDA methods, with the most noticeable difference compared to the reference
method, again mainly being the better t for NO. However the
NO concentrations after the transient did not agree as well with
measurement data as for the case in Fig. 5. The t of HC and CO
is somewhat better for the MVDA methods, but there were still
notable differences compared to measurement data.
The above gures show only two of 16 transients used for
parameter estimation (4 catalyst congurations and 4 operating
point changes for every conguration). The transients were, as previously mentioned, selected to give a wide temperature, ow and
concentration window. The combinations of torque and speed that
150
Fig. 3. Measured outlet mole fraction of NO for an example of a selected data set colored by relative sensitivity. The order of the engine operating points is 1, 7, 2, 8, 3, 4, 5, 6
which were run sequentially for catalyst congurations a, b, c and d (separated by blue vertical lines). The relative sensitivity is an indication of how sensitive the data point
is to changes in parameter value relative to the rest of the data set. A value of 0.9 for example, means that the data point has a higher sensitivity than 90% of the data set. (For
interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)
Fig. 4. Residual sum of squares for every set of parameters for the two evaluated methods. The dashed line is the residual sum of squares for the nal set of parameters from
the reference method. (a) 1 step method (b) 5 steps method (c) 1 step method, magnied (d) 5 steps method, magnied.
Table 3
Residual sum of squares for the experiment displayed in Figs. 5 and 6 for every component.
NO
HC
CO
Figure
1 step
5 steps
Ref
40.9
42.3
85.5
40.4
38.7
43.9
94.9
79.4
68.2
5
5
5
1 step
5 steps
Ref
66.2
76.6
212.8
112.6
111.9
112.2
80.2
101.9
157.2
6
6
6
b)
a)
1600
1200
[ppm]
1000
60
NO in
NO out measured
NO out sim, 1 Step
NO out sim, 5 Steps
NO out sim, Ref
C H in
3 6
50
[ppm]
1400
151
800
600
40
30
20
400
10
200
0
0
200
400
600
0
0
200
400
600
c)
d)
250
500
CO in
CO out measured
CO out sim, 1 Step
CO out sim, 5 Steps
CO out sim, Ref
200
150
450
400
350
[C]
[ppm]
100
300
250
Tin
200
50
150
0
0
200
400
600
100
0
out
measured
Tout simulated
200
400
600
Fig. 5. Measured and simulated outlet concentrations of NO (a), C3 H6 (b), CO (c) and temperature (d) for a change in operating point from 1 to 7 for catalyst conguration c.
result in high HC and CO are few and therefore only two of the
eight engine operating points used in the study generate signicant
HC and CO concentrations at the catalyst outlet. To demonstrate
the performance of the different methods, transients with engine
operating points with high HC and CO have been chosen in both
Figs. 5 and 6. In addition, the high temperature at the start of the
experiments provide an interesting transient change in both CO and
HC conversion for the rst 900 s and the change in engine operating
point provides transients in NO and temperature. To put the t of
b)
a)
1000
800
[ppm]
700
600
NO in
NO out measured
NO out sim, 1 Step
NO out sim, 5 Steps
NO out sim, Ref
50
500
400
300
40
3 6
30
20
200
10
100
0
0
C H in
3 6
60
[ppm]
900
70
200
400
600
0
0
200
400
600
c)
d)
250
450
CO in
CO out measured
CO out sim, 1 Step
CO out sim, 5 Steps
CO out sim, Ref
150
400
350
[C]
[ppm]
200
100
300
250
Tin
200
50
0
0
150
200
400
600
100
0
out
measured
Tout simulated
200
400
600
Fig. 6. Measured and simulated outlet concentrations of NO (a), C3 H6 (b), CO (c) and temperature (d) for a change in operating point from 2 to 8 for catalyst conguration b.
1.00
12.3
10.9
4.85
7.71
1.00
0.706
0.706
0.793
1.00
1.00
0.740
0.730
0.828
0.701
actd a [%]
actb a [%]
3.98
50.6
3.96
1.35
1.93
2.08 10
3.88 105
2.59 105
5.33 104
8.25 104
1.00
1.00
1.00
1.00
0.766
3.10 104
2.28 104
2.68 104
2.87 104
3.05 104
9.65 104
9.67 104
9.81 104
9.74 104
9.69 104
65.5
39.0
0.151
77.7
95.8
4.50 10
3.18 1013
1.77 1014
2.34 1017
2.71 1014
4.00 10
4.70 1017
1.06 1018
1.67 1022
1.25 1019
1.00 10
2.17 1015
2.59 1016
2.78 1020
4.89 1016
3.00 103
4.54 103
2.21 103
3.30 103
1.64 103
7.99 103
5.54 103
1.02 104
8.50 103
8.06 103
7.00 104
7.82 104
8.18 104
7.29 104
7.46 104
Start
1 step
5 steps
Ref
Preceding study
EA,5 [J/mol]
EA,4 [J/mol]
EA,3 [J/mol]
a
acta is activity scaling factor for catalyst conguration a (0.30 wt% Pt), actb is activity scaling factor for catalyst conguration b (0.59 wt% Pt), and actd is activity scaling factor for catalyst conguration d (0.59 wt% Pt with thin
washcoat).
b
Effective diffusivity scaling for small (S) and large (L) molecules.
[]
1.00
1.49
1.34
1.62
2.11
fDscl ,L
[]
b
fDscl ,S
8.00 10
7.33 104
7.85 104
7.90 104
7.67 104
4.79 10
4.85 105
2.99 106
4.84 1010
4.71 107
EA,2 [J/mol]
EA,1 [J/mol]
A7 []
A6 []
A4 []
A5 []
14
17
20
A2 [mol, K/m2 , s]
A3 [mol, K/m2 , s]
Table 4
Results of kinetic parameter tuning.
10.0 104
9.51 104
9.52 104
9.86 104
9.80 104
A1 [mol, K/m2 , s]
152
the kinetics. The result of the parameter tuning with MVDA shows
the importance of having a well formulated model and also displays
how an efcient parameter tuning method can tune the parameters
to less realistic values, and still achieve a good t, even if the model
formulation is decient.
In general the parameter values from the preceding study are
closer to the values of the MVDA methods than the reference case.
The parameters from the preceding study also gave an overall better t with a residual sum of squares of 3094 compared to 3580
for the reference case. There are two main differences between the
parameter estimation method of the preceding study and the reference case. Firstly the residuals were calculated differently, see
Section 3.1, and secondly a ten times higher parameter weighting
for kinetic parameters (w in Eqs. (A15) and (A17)) was used in the
current study to make the Jacobian less sensitive to numerical noise.
It appears that this increased parameter weighting has resulted in
estimated parameter values further away from the original ones
with an overall poorer t. This could be an effect of some parameters, such as A1 A3 , having a large inuence on the t at an early
stage of the parameter estimation which makes them change fast
and thereby reach values where the sensitivity for the other parameters are low. In other words the risk of nding a local minimum
may be increased by the increased parameter weighting. One of the
aims of the MVDA method was to make the inuence of parameters
more equal and it appears that this has prevented the parameters
from reaching extreme values for both the 1 step and the 5 steps
method even though the parameter weighting was high.
4.5. Final remarks
By continuously changing data points on which the parameter
estimation is performed the risk of nding a local minimum appears
to have been reduced at the cost of longer computational time.
As mentioned in the introduction, these characteristics are typically associated with global parameter estimation methods such
as simulated annealing. The difference with the method at hand is,
however, that since a gradient search method is applied, the parameters will evolve without evaluating a large number of parameter
sets (as is characteristic of a global search method) which will put
less demand on the model stability. This will also reduce computational cost which already is high for such a complex system of
equations and large data set.
The method of parameter estimation should be useful for complex dynamic models where model stability limits what parameter
values are possible and where the number of data points is large.
The model is not likely to be suitable if the diversity of the data
points is low, for example if the model should be tuned to only a
few stationary data points. With low diversity in the data it is not
likely to achieve a notable improvement in parameter sensitivity
by changing the data point selection. The computational effort is
then better spent on tuning the model to the few available data
points directly. For models where the stability is highly dependent
on starting conditions, such as detailed kinetic models, the large
number of short simulation sections where starting conditions are
needed may make the method less suitable. For example it may be
very difcult to nd stable starting conditions if the rst data point
is in the middle of a sharp transient such as a catalyst ignition point.
5. Conclusions
153
0=
Vk,0
dci,k,0
dt
0=
dci,k,n
Vk,n
dt
(A2)
for n1
where rj,k,n is the reaction rate in mol/m2 Pt/s and Am,k,n is the catalyst active surface area in channel segment k and wall layer n,
Vk,n is the volume of channel segment k and layer n and is the
washcoat porosity. Index k indicates the segment number (axial
discretization) where k = 1 represents the rst segment in the direction of the ow and k = K represents the last. The index n indicates
the layer number (radial discretization) where n = 1 represents the
rst washcoat layer closest to the gas bulk, n = N represents the last
layer and n = 0 represents the gas bulk. The index i indicates the
species and index q represent catalyst conguration.
The mass transfer coefcients i,k,n are given by:
i,k,0 =
Ak
(1/kc,i,k ) + (0.5x1 /Deff,i,k )
(A3)
i,k,n =
Deff,i,k Ak
0.5xn + 0.5xn+1
(A4)
i,k,N = 0
(A5)
where Ak is the radial mass and heat transfer area in channel segment k. For simplicity, Ak was assumed to be constant for all wall
layers. Deff,i,k is the effective pore diffusion coefcient of component
154
i in channel segment k, kc,i,k is the lm transfer coefcient of component i in channel segment k and xn is the thickness of washcoat
layer n.
The gas energy balance is given by:
0=
Vk,0 cp,g
dTg,k
dt
Deff,i,k =
= Ftot cp,g (Tg,k1 Tg,k ) hk Ak (Tg,k Ts,k )
(A6)
where cp,g is the heat capacity for the gas, hk is the heat transfer
coefcient in channel segment k, Tg,k and Ts,k are the temperatures
in channel segment k in the gas bulk and of the catalyst respectively. Note that the solid temperature, Ts,k , was not discretized
radially since the high solid conductivity and short conduction distance will result in very small radial temperature gradients. The
main temperature transport resistance is in other words assumed
to be in the lm. The solid energy balance is given by:
dTs,k
dt
cp,s ms,k +
sheat
K
n
(A7)
where ms,k is the mass of catalyst material and substrate in channel segment k, cp,s is the mass weighted average of washcoat and
substrate heat capacity, As is the channel cross sectional area of the
substrate and washcoat, Hj is the heat of reaction j. sheat is a heat
sink originating from solid material not included in the catalyst
(e.g. catalyst canning and insulation). The heat loss to the environment is modeled by the lumped heat transfer parameter tot and
the effective environmental temperature T . K is the total number
of channel segments.
The solid axial heat ux was calculated as:
Ts,k Ts,k1
0.5zk + 0.5zk1
(A8)
for k = 2,. . ., K 1, K
qk = 0
(A9)
hk =
ShDi,k
d
Nug
d
Di,k = Dref,i
Ts,k
Tref
dp
3
8RTs,k
Mi
(A14)
(A15)
where w is a weight factor for every parameter and p is the parameter that is changed by the gradient search method. The superscript
o in the equations above indicates that this is the original values
taken from literature or previous successful parameter tuning. The
reaction rate coefcients were centered around a reference temperature to decrease the parameter correlation (Box et al., 1973).
o
kj,ref
= Aoj e
(E o /RTref )
(A16)
A,j
(A17)
When Eq. (A18) is inserted into the expression for the reaction
rate coefcient in Eq. (6), the following expression for the tuned
reaction rate coefcient is obtained
kj = kj,ref e[(EA,j /R)(1/TS 1/Tref )]
(A19)
The mass transport was tuned by adjusting the effective diffusivities for the species taking part in the reactions.
(A10)
fDscl ,i = fDo
(A11)
DKi,k =
for k = 1, k = K + 1where s is the heat conductivity for the solid material (same conductivity assumed for the washcoat and the cordierite
support) and zk is the length of channel segment k (the total heat
transfer distance is half the distance of each segment).
The mass and heat transfer is described with the lm model and
the mass and heat transfer coefcients are given by:
kc,i,k =
(A13)
o
EA,j = EA,j
+ WEa,j pEa,j
tot
(Ts,k T )
K
fD
(1/Di,k ) + (1/DKi,k )
Both the pre-exponential factors, Aj , and the activation energies, EA,j , can be tuned. The scaling of EA,j was straightforward and
performed according to:
qk = s
1.75
(A12)
scl ,i
Deff,i,k =
+ wDscl ,i pDscl ,i
fD
fD ,i
(1/Di,k ) + (1/DK,i,k ) scl
scl ,i
(A20)
(A21)
is typically around 1.
(A22)
(A23)
o
sheat = sheat
+ Wsheat psheat
(A24)
(A25)
f
155
j=1
j = [2, 3, . . ., nlayers , nlayers + 1]
(A26)
k
was
calculated
to
fulll
npoints
where
(1:nlayers ) m npoints (1:nlayers + 1)
6. A vector denoted int was created to divide the data set into
intervals:
int (j) = 0
npoints (1 : j 1)
int (j) =
j=1
j = [2, 3, . . ., nlayers , nlayers + 1]
j=1
j = [2, 3, . . ., nlayers ]
j=1
j = [2, 3, . . ., nlayers ]
where the rst input to D-optimal design was the rows in the
selected interval and the second input was the number of points
to be selected.
The method results in a selection of 1% of the total number of
data points on which PCA was applied. The exponentially increasing
number of points in every onion layer resulted in the majority of
the points being selected at high sensitivity since an equal number
of points (number of components in the PCA +1) was selected from
every layer except the rst. An example of the sectioning of the
data set into layers and the relative parameter sensitivity of the
data points in the layers is shown in Fig. A1.
In the rst onion layer, where the parameter sensitivity is highest, the selected rows in the centered and scaled scores matrix
was expanded by additional columns containing the same columns
again but squared. The reason for this was to capture the nonlinear
behavior in the parameter sensitivity.
156
Fig. A1. D-optimal onion layer thickness colored by relative sensitivity. The relative sensitivity is an indication of how sensitive the data point is to changes in parameter
value relative to the rest of the data set. The color for one bar is set to the average relative sensitivity of the data points in that onion layer.
References
Ansell GP, Bennett PS, Cox JP, Frost JC, Gray PG, Jones AM, et al. The development
of a model capable of predicting diesel lean NOx catalyst performance under
transient conditions. Appl Catal B: Environ 1996;10:183201.
Berger RJ, Kapteijn F, Moulijn JA, Marin GB, De Wilde J, Olea M, et al. Dynamic
methods for catalytic kinetics. Appl Catal A: Gen 2008;342:328.
Bernaerts K, Versyck KJ, Van Impe JF. On the design of optimal dynamic experiments
for parameter estimation of a Ratkowsky-type growth kinetics at suboptimal
temperatures. Int J Food Microbiol 2000;54:2738.
Blockley R. Encyclopedia of aerospace engineering. Chichester, West Sussex,
UK/Hoboken, NJ: John Wiley & Sons; 2010. p. 522938.
Box MJ. The occurrence of replications in optimal designs of experiments to
estimate parameters in non-linear models. J R Stat Soc Ser B (Methodol)
1968;30:290302.
Box GEP, Hunter WG. The experimental study of physical mechanisms. Technometrics 1965;7:2342.
Box GEP, Lucas HL. Design of experiments in non-linear situations. Biometrika
1959;46:7790.
Box GEP, Hunter WG, MacGregor JF, Erjavec J. Some problems associated with the
analysis of multiresponse data. Technometrics 1973;15:3351.
Chen DKS, Bissett EJ, Oh SH, Van Ostrom DL. A three-dimensional model for the analysis of transient thermal and conversion characteristics of monolithic catalytic
converters. SAE, international technical paper number: 880282; 1988.
Coleman TF, Li YY. An interior trust region approach for nonlinear minimization
subject to bounds. Siam J Optim 1996;6:41845.
Crocoll M, Kureti S, Weisweiler W. Mean eld modeling of NO oxidation over
Pt/Al2 O3 catalyst under oxygen-rich conditions. J Catal 2005;229:4809.
Draper NR, Hunter WG. Design of experiments for parameter estimation in multiresponse situations. Biometrika 1966;53:52533.
Draper NR, Hunter WG. The use of prior distributions in the design of experiments for
parameter estimation in non-linear situations. Biometrika 1967a;54:14753.
Draper NR, Hunter WG. The use of prior distributions in the design of experiments for
parameter estimation in non-linear situations: multiresponse case. Biometrika
1967b;54:6625.
Dubien C, Schweich D, Mabilon G, Martin B, Prigent M. Three-way catalytic converter modelling: fast- and slow-oxidizing hydrocarbons, inhibiting species, and
steam-reforming reaction. Chem Eng Sci 1998;53:47181.
AA. The microkinetics of
Dumesic JA, Rudd DF, Aparicio LM, Rekoske JE, Trevino
heterogeneous catalysis. Washington, DC: American Chemical Society; 1993.
Ericson C, Westerberg B, Odenbrand I. A state-space simplied SCR catalyst model
for real time applications. SAE number: 2008-01-0616; 2008.
Franceschini G, Macchietto S. Model-based design of experiments for parameter
precision: state of the art. Chem Eng Sci 2008a;63:484672.
Franceschini G, Macchietto S. Novel anticorrelation criteria for model-based experiment design: theory and formulations. AIChE J 2008b;54:100924.
Froment GF, De Wilde J, Bischoff KB. Chemical reactor analysis and design. Hoboken,
NJ: Wiley; 2011.
Glielmo L, Santini S. A two-time-scale innite-adsorption model of three way
catalytic converters during the warm-up phase. J Dyn Syst Meas Control
1999;123:6270.
Goodwin GC. Dynamic system identication: experiment design and data analysis/Graham C. Goodwin and Robert L. Payne. New York: Academic Press; 1977.
Guojiang W, Song T. CFD simulation of the effect of upstream ow distribution
on the light-off performance of a catalytic converter. Energy Convers Manag
2005;46:201031.
Gthenke A, Chatterjee D, Weibel M, Krutzsch B, Koc P, Marek M, et al. Current status of modeling lean exhaust gas aftertreatment catalysts. In: Marin GB, editor.
Advances in chemical engineering. Academic Press; 2007. p. 103283.
Hill WJ, Hunter WG, Wichern DW. A joint design criterion for the dual
problem of model discrimination and parameter estimation. Technometrics
1968;10:14560.
Hunter WG, Hill WJ, Henson TL. Designing experiments for precise estimation of all
or some of the constants in a mechanistic model. Can J Chem Eng 1969;47:7680.
Kandylas IP, Koltsakis GC. NO2 -assisted regeneration of diesel particulate lters: a
modeling study. Ind Eng Chem Res 2002;41:211523.
Koltsakis GC, Stamatelos AM. Modeling dynamic phenomena in 3-way catalytic
converters. Chem Eng Sci 1999;54:456778.
Koltsakis GC, Konstantinidis PA, Stamatelos AM. Development and application range
of mathematical models for 3-way catalytic converters. Appl Catal B: Environ
1997;12:16191.
Lafossas F, Matsuda Y, Mohammadi A, Morishima A, Inoue M, Kalogirou M, et al.
Calibration and validation of a diesel oxidation catalyst model: from synthetic
gas testing to driving cycle applications; 2011.
Lundberg B, Sjoblom J, Johansson , Westerberg B, Creaser D. Parameter estimation of a DOC from engine rig experiments with a discretized catalyst washcoat
model. SAE Int J Engines 2014;7:1093112.
Martens H, Naes T. Multivariate calibration. John Wiley & Sons; 1989.
Massman WJ. A review of the molecular diffusivities of H2 O, CO2 , CH4 , CO, O3 ,
SO2 , NH3 , N2 O, NO, and NO2 in air, O2 and N2 near STP. Atmos Environ
1998;32:111127.
Mladenov N, Koop J, Tischer S, Deutschmann O. Modeling of transport and
chemistry in channel ows of automotive catalytic converters. Chem Eng Sci
2010;65:81226.
Murzin D, Salmi T. Mass transfer and catalytic reactions. Catalytic kinetics. Amsterdam: Elsevier Science; 2005. p. 341418.
157
Stamatelos AM, Koltsakis GC, Kandylas IP, Pontikakis GN. Computer aided engineering in diesel exhaust aftertreatment systems design. Proc Inst Mech Engin Part
D: J Automob Eng 1999;213:54560.
epnek J, Koc P, Marek M, Kubcek M. Catalyst simulations based on coupling of
St
3D CFD tool with effective 1D channel models. Catal Today 2012;188:8793.
Tronconi E, Forzatti P. Adequacy of lumped parameter models for SCR reactors with
monolith structure. AIChE J 1992;38:20110.
Tsinoglou DN, Koltsakis GC. Effect of perturbations in the exhaust gas composition
on three-way catalyst light off. Chem Eng Sci 2003;58:17992.
Voltz SE, Morgan CR, Liederman D, Jacob SM. Kinetic study of carbon monoxide and
propylene oxidation on platinum catalysts. Prod R&D 1973;12:294301.
Wang TJ, Baek SW, Lee JH. Kinetic parameter estimation of a diesel oxidation catalyst
under actual vehicle operating conditions. Ind Eng Chem Res 2008;47:252837.
Watling TC, Ahmadinejad M, Tutuianu M, Johansson , Paterson MAJ. Development
and validation of a PtPd diesel oxidation catalyst model. SAE technical paper
number: 2012-01-1286; 2012.
Zullo LC. (PhD thesis) Computer aided design of experiments: an engineering
approach (PhD thesis). University of London; 1991.