You are on page 1of 25

Simulation Modelling Practice and Theory 60 (2016) 1539

Contents lists available at ScienceDirect

Simulation Modelling Practice and Theory


journal homepage: www.elsevier.com/locate/simpat

Mathematical modelling and parameter identification


of a stainless steel annealing furnace
S. Zareba a, A. Wolff b, M. Jelali a,
a
b

Cologne University of Applied Sciences, Lab Control Engineering and Mechatronics, Betzdorfer Str. 2, 50679 Kln, Germany
VDEh-Betriebsforschungsinstitut GmbH, Division Measurement and Automation, Sohnstr. 65, 40237 Dsseldorf, Germany

a r t i c l e

i n f o

Article history:
Received 27 July 2014
Received in revised form 9 September 2015
Accepted 23 September 2015

Keywords:
Continuous annealing
Modelling
Parameter identification
Optimization

a b s t r a c t
A new, comprehensive mathematical model of continuous annealing furnaces is developed,
under consideration of both the radiative and convective heat transfer of the furnace
components. Based on measured normal operating data from an industrial stainless steel
plant, parameter identification is basically carried out using a nonlinear least-squares
optimization algorithm for the whole annealing furnace, to estimate optimal values of
uncertain parameters, such as emissivities. Due to the complexity of the model, a sequential approach for parameter identification is proposed and implemented, i.e. the parameter
set is divided into different subsets, and the parameter estimation is carried out sequentially in several steps and iterations. The performance of the model with the estimated
parameters is then evaluated on a different test data set. It is shown that the obtained
model can predict temperature evolutions along the furnace in good agreement to
measured data, under both steady-state and transient conditions. The presented model
is suitable for controller design and process optimization.
2015 Published by Elsevier B.V.

1. Introduction
1.1. Continuous annealing furnaces
Continuous annealing furnaces are widely used for heat treatment after cold rolling in flat steel production, with the aim
to produce steel strips of high tensile strength and high formability. Commonly used continuous annealing lines are either
vertical or horizontal, depending on the strip-guiding method through the furnace; see Fig. 1. Through the global furnace
structure, the number and length of heating and cooling areas, different annealing treatments can be achieved. With both
types, it is possible to operate at high or low temperature level.
Some furnaces have enclosed heating burners to avoid chemical reaction of the strip with hydrogen or oxygen at high
temperatures. Thereby, the combustion takes place inside a pipe to avoid getting waste gas into the furnace, and the heat
exchange between heat pipe and strip is mainly radiative. Alternatively, the heating power is generated by the electric
method. In both cases, the furnace can be also filled with an expansive inert gas instead of ambient air to reduce the chemical
reactions. In a closed furnace there is no significant air flow inside the furnace room and the heat transfer is mainly radiative.
Furnaces with internal combustion have much greater air flow and, depending on the burner type, a significantly higher air
turbulence. In this case, the convective heat transfer between gas and strip and gas and wall cannot be neglected.
Corresponding author.
E-mail addresses: sebastian.zareba@fh-koeln.de (S. Zareba), andreas.wolff@bfi.de (A. Wolff), mohieddine.jelali@fh-koeln.de (M. Jelali).
http://dx.doi.org/10.1016/j.simpat.2015.09.008
1569-190X/ 2015 Published by Elsevier B.V.

16

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

Fig. 1. Schematic view of a horizontal (upper diagram) and vertical (lower diagram) continuous annealing process.

Usually, the strip temperature within the furnace is not accessible to measurements. Only the strip temperature at the
exit of the furnace is measured by radiation pyrometry, but it is well known that such measurements are often unreliable,
due to radiative interference and inefficiency of installed shieldings [7,33].
1.2. Motivation and control objectives
There is a continuous interest in the process industry to use the fewest possible resources to heat up strip coils. Even
small improvement of control parameters, control structure or process scheduling that allows a reduction in energy consumption can lead to cost savings and indirectly to a reduction of the climate pollution.
The simplified furnace design shown in Fig. 2 is the target plant of this study. It is direct-fired with natural gas for a clean
combustion. The considered furnace has three connected chambers and the strip is running from the first to the third
chamber. Cooled transport rolls are integrated between the chambers for supporting and guiding the strip. The furnace is
divided into several zones, numbered in the direction of the strip movement, and having different lengths. The goal of furnace
is to heat up flat stainless steel as fast as possible to a reference temperature, with a ramp to spike profile, so that there is no
soaking zone.
At normal operation, no additional fresh air is blown in, so that the exhaust gas flow is a result of combustion only. If the
burners power decreases, exhaust gas flow decreases too. The used FLOX burners [39] have no visible flame for a furnace
operating area above 850 C. They are arranged in pairs over the whole length of the furnace with varying number per zone.
The burners in zone 1 are arranged at the end of zone because the first meters are used as a pre-heating area.

Fig. 2. A horizontal annealing furnace.

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

17

Before being guided into the furnace, the flat steel run through a pickling line, where the surface is chemically cleaned and
warmed up by the process to the initial temperature of T in  30  C. Strip tarnishing has a effect on the heat exchange, as it
changes surface properties like radiative emissivity and occurs after reaching a specific strip temperature in the range
T tarn 630700 C [10].
The prediction and control of the strip temperature evolution along continuous annealing lines is of main interest to
ensure the requested mechanical properties of flat steel. The challenge of the temperature control is to manage the transient
operating behavior caused by frequent variations of processing conditions and strip parameters. The direct gas-fired furnace
under investigation is part of a horizontal continuous annealing line for stainless steel. Combined with a downstream cooling
process, the annealing treatment is used to modify the grain size within the stainless steel strip. Because of frequent changes
in the strip parameters, the annealing process must be permanently controlled either in strip velocity or temperature. However, the reduction of strip velocity is an unwanted option, since it reduces process productivity.
Due to the complexity of the control task, high-performance strip-temperature control requires a model-based control
strategy, such as model predictive control, as recently demonstrated by Steinboeck et al. [34] for slab-reheating furnaces.
This relies on an accurate model that describes the behavior of the annealing process during both steady-state and transient
temperature periods [7]. A comprehensive model is required for the realistic simulation and pre-tuning/-commissioning of
the temperature control system as well. Based on the comprehensive model, simplified and fast models can then be derived
and used for online-control and optimization purposes.
1.3. Review of annealing-furnace models
The literature contains some reports on the modelling of annealing furnaces. A detailed model, considering the combustion process and the strip heating by using energy and mass balances inside the industrial furnace, has been presented by
Carvalho et al. [4] for silicon steel strip annealing (a tunnel blast furnace). Marlow [20], McGuiness and Taylor [22], and Prieto et al. [25] approximated gas-fired furnaces using one-dimensional geometry. A similar approach was proposed by Ramamurthy et al. [26] for indirect-fired radiant furnaces. Depree et al. [7] studied furnace modelling using two methods: (1) a 3D
finite element model using COMSOL Multiphysics software, and (2) a simple/rapid 1D/2D finite difference model scheme
using MATLAB software. A comparison between the predictions of both models has been provided. Chen et al. [5] developed
a finite element code ANSYS to analyze temperature and stress distributions in (vertical) continuous annealing lines. In Kang
and Chen [15], 3-D temperature distributions of strip in continuous annealing lines were theoretically evaluated by using
both the techniques of energy balance method and finite element method, in which the view factors, ferriteaustenite phase
transition, and thermal contact conductance between strip and roll were taken into account. A modelling approach for a vertical continuous annealing furnace with only radiative heat exchange is proposed by Wan et al. [37]. A commonly used
approach is to construct data-driven models using different techniques as ARX models, neural networks, or principlecomponent analysis [40,23,24,3,16].
The development and comparison of 2-D and 3-D zone models of large-scale slab reheating furnaces, which are thermodynamically comparable with continuous flat steel furnaces, has been done by Tian et al. [36]. They considered the radiative
heat transfer and validated the models using experimental and simulated plant data. A mathematical model of a slab reheating furnace considering heat conduction within the slab is derived by Steinboeck et al. [33]. The modelling of a direct-fired
continuous annealing furnace, considering the chemical reactions and the flue-gas composition for a steel strip is carried out
in Strommer et al. [35], whereas in this work a stainless steel strip is considered, where the heat treatment is performed at
higher temperatures and the material properties are different.
To the best of authors knowledge, although many mathematical models exist for simulating and controlling of annealing
processes, the topic of identifying model parameters from measurement data to improve thermal models is seldom
addressed in the literature. Yoshitani and Hasegawa [40] proposed to estimate the parameters of a simplified model of a
vertical annealing furnace using an algorithm for recursive parameter estimation. In Chen et al. [5], the emissivity and
equivalent heat-transfer coefficient of the strip in a preheating furnace is estimated based on an energy model of the strip
and roll and the minimization of a least-squares error function. In most cases, however, values for model parameters are, at
best, educated guesses, but even a slight change, especially in the emissivities, may change significantly the results of the
strip temperature in the models [19].
Certainly, parameter identification for a furnace model is a challenging task, and needs a sufficient measurement-data
history for model estimation and validation, as stated by Schurko et al. [30] and Steinboeck et al. [33]. However, the
availability of databases at steel plants today makes it easier to apply powerful identification algorithms for proper tuning/
calibration of the thermal models, to give enhanced predictions required for model-based control.
There is no existing model for a stainless steel continuous strip annealing furnace with the focus on the identification task
of distributed parameters available. Furthermore, no model that considers the used FLOX burners inside a stainless steel
continuous strip annealing furnace can be found.
1.4. Contribution and contents of the work
In this work, a new comprehensive mathematical model for temperature prediction in a stainless steel annealing furnace
is developed, considering all relevant components (strip, gas and insulation) with their mutual interactions, based on

18

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

thermodynamic fundamentals. A special characteristic of the investigated furnace are flameless oxidation (FLOX) burners
that are used for a clean combustion, consequently causing a high turbulent environment inside the furnace. Both the
radiative and convective heat transfer of the furnace components are described. Therefore, the performance-related turbulences inside the furnace are considered for convective heat flow and the brightness method is used for radiative heat-flow
computation. Dividing the furnace into several zones, an energy balance for the participating strip, gas and insulation with
opposite directions of motion of gas and strip is realized.
The developed model is suitable for simulation of annealing furnaces, design and pre-tuning of advanced control systems.
A further contribution is to propose and demonstrate a new model identification strategy for estimating and validating the
model parameters using measured normal operating data from a real industrial plant. The parameter identification is basically carried out using a large-scale optimization method, i.e. a nonlinear least-squares optimization algorithm for the whole
heating furnace, and delivers optimal parameter estimates for uncertain parameters, e.g. emissivities. Due to the complexity
of the model, a sequential approach for parameter identification is proposed and implemented, i.e. the parameter set is
divided into different subsets, and the parameter estimation is carried out sequentially in several steps and iterations.
The paper is organized as follows: Section 2 contains the main theoretical derivation of the developed model. In Section 3,
the parameter identification approach is explained and the results achieved are presented. The paper ends with concluding
remarks in Section 4.
2. Mathematical model and computational procedure
For the numerical modelling, the furnace is divided into sub-systems: first in its chambers and then into its zones. This
should be the minimum division of the furnace, since the given control structure manipulates the gass mass within one
zone, resulting in seven controlled zones. To increase the accuracy of the model, each zone is divided again into several equal
segments. The arrangement of several segments is shown on the left-hand side of Fig. 3; one of these segments with the
transport direction inside is detailed on the right-hand side of the figure. The furnace segment is represented as a square
profile with a thickness dw and a heat capacity cpw . The flue gas inside the furnace flows through the segment loosing or
absorbing energy to/from the wall insulation and strip.
2.1. Assumptions
The geometry of a horizontal furnace has a lot of symmetry, so that some simplifications can be made in the modelling
approach:
 The temperature distribution of the gaseous medium and the opposite furnace insulation surfaces within one segment at
the upper and lower side of strip is homogeneous. Therefore, the furnace is divided in the horizontal plane and can be
reduced to a enclosure system with rectangle plates and a participating gas, as shown in Fig. 4. The gaps between the
furnace insulation and the strip are treated as mirror walls.
 The strip is also divided in the horizontal plane considered by a half-size strip thickness.
 All participating surfaces, except the mirror walls, behave as gray bodies with an emissivity 0 <  < 1. The surfaces are
treated as ideal diffuse reflectors.
 It is assumed that the radiation from and to the upper part of the furnace through the gap between the strip and
insulation has the same value. Therefore, mirror walls treated as diffuse reflecting bodies with an emissivity of gap 0
are applied at this place.
 The radiative heat exchange is calculated by the zone method and the usage of view factors. For a similar furnace geometry, Depree [8] already shows on the basis of the view factors that most of the heat exchange occurs inside one segment
and there is only small heat exchange between the segments next to each other. Therefore, radiative heat exchange
between two segments is neglected. To prove that this assumption is valid for this furnace, a calculation of the viewfactors is carried out in Appendix A.
 The heat accumulation of furnace insulation is considered as a stationary flat area with a density qw , heat capacity cpw and
conductivity kw that are homogeneous distributed over the whole thickness dw .

Fig. 3. Arrangement of several segments (left) and one segment with moving direction (right).

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

19

Fig. 4. Considered enthalpy and heat flows.

 To approximate the transient temperature behavior of the insulation, the mass is split into several unequal thick layers
and conductive heat exchange between this masses is considered.
 The furnace is not a perfect isolated system and heat disappear through the furnace insulation and the incoming and
outgoing strip gap of furnace. The heat loss is considered as a radiative heat flow from the furnace outside surface to
the ambiance induced by thermal radiation of an approximated square profile of the same size.
 The combustion of the used FLOX burner happens over a wide space inside the furnace. Therefore, the thermal radiation
of the flue gas as well as the convective heat exchange due to the very high flue-gas turbulences are taken into account.
 All FLOX burners in one zone are manipulated in parallel with a homogeneously distributed combustion. Because more
than one burner per segment is considered, the injection of fuel is assumed to be continuous.
 The flue gas is considered as an incompressible fluid, with a temperature-depending viscosity, density, heat conduction,
heat capacity and emissivity. The flue gas is a mixture of N2 ; H2 O and CO2 .
This furnace is used for continuous annealing of flat stainless steel and some assumptions due to strip properties and heat
exchange are made:
 Flue gas and strip are running inside the furnace in opposite directions and have different velocities.
 An enthalpy balancing of the incoming flue-gas flow, the fuel and the fresh-air injection volumes gives the outgoing
gas-mass flow.
 The strip has a homogeneous temperature distribution along each segment, and conduction from one element to the
other is neglected.
 Only the top and the bottom surface of the strip interact with the environment and are considered for heat exchange. The
lateral strip faces do not participate in the heat exchange process.
 The strip is conveyed through the furnace on water-cooled rolls. Because the roll-surface temperature, as well as the
material properties are unknown, the heat exchange is approximated by a simple heat-conduction method without
considering changing surface or material properties.
2.2. Radiative heat transfer
Non-luminous gaseous radiation is usually the dominant mode of heat transfer from the flame and combustion product in
natural gas-fired furnaces. Gases, such as nitrogen N2 or oxygen O2 , with symmetrical oscillating patterns, do not emit or
absorb in the technical interesting infrared wavelength and take no effect in the thermal radiative heat exchange within the
furnace. Therefore, the most important emitters, in the case of a stoichiometric or fuel-lean combustion, are carbon dioxide

20

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

and water vapor [27]. The furnace under investigation uses high velocity burners, where the combustible fuelair mixture is
preheated and injected into the furnace, where, under special condition, a flameless oxidation occurs [39]. Since this burner
type has no flame front, the generation of nitrogen oxides NOx and carbon monoxide CO is very low. Because of the low
concentration in the flue gas, the influence of CO, NOX and methane CH4 , to the radiative heat exchange, can be neglected.
Nevertheless, for radiative heat exchange, the flue gas must be considered as a participating gas with a specific emissivity
flue and absorptivity aflue .
The determination of the thermal radiation heat exchange inside an industrial furnace is very complex. It is, in contrast to
convection or conduction, a global phenomenon, where the thermal radiation exchange depends on material properties and
temperature of surfaces and gas volume inside the whole enclosure. There are different approaches to solve this problem, e.g.
the brightness method or different methods of energy balancing. The zone method, which was first proposed by Hottel and
Sarofim [13], divides the furnace into a number of zones, assuming a uniform emissivity and temperature distribution of the
surfaces areas and gas volumes.
A multi-surface enclosure, containing a homogeneously distributed participating gas, as shown on the left side of Fig. 5,
with N S 6 surface sections and N G 1 gas volumes, can be defined by applying the zone method. Assuming that the
participating surface areas Ak ; k 1; . . . ; N S and volume V m ; m 1; . . . ; N G , within the enclosure act like diffuse gray or black
radiators, the radiative heat exchange can be derived.
As mentioned in Section 2.1, the symmetry can be used to simplify the radiation exchange problem by dividing the
considered furnace in the horizontal plane. Therefore, the gaps between the strip and furnace insulation are treated as mirror
walls with a emissivity Gap 0. Hence, the three surfaces areas of the furnace insulation A13 , the two mirror walls A4 and
A6 , the strip A5 and the flue gas V G as a participating medium inside this enclosure are considered for the thermal radiation
heat exchange, see also Fig. 5 on the left-hand side.
The view factors uij of the exchange areas Ai are calculated using the derivations of Siegel and Howell [31] for parallel and
perpendicular rectangle surfaces and are merged to a matrix of the form

0
B
/ij B
@

u11

...

..
.

..

u1NS 1
..
.

C
C:
A

uNS 1 . . . uNS NS
In regard to Beers law, the radiative flux density from a emitting surface is reduced by the transmitting gas, depending on
the gas properties and the path length. The absorptivity of the flue gas inside the furnace is considered by introducing an
additional exchange area

A7

NS
X
Ai ;

i1

situated at the middle of the furnace and modification of the view factor matrix /ij in such a way that

F ij uij ekG SSij

i; j 1; . . . ; NS ;

where kG is the absorptivity coefficient of the flue gas and SSij is the distance between the midpoints of the exchange areas Ai
and Aj . The radiative heat flux from the exchange areas to the absorbing gas volume can be then calculated with the new
view factor

Fig. 5. Multi-zone enclosure with heat exchange surface areas (left) and energy balancing at one exchange surface (right).

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539


NS
X
F ij

F i7 1 

i 1; . . . ; N S :

21

j1

Assuming the flue gas behaves as a gray gas, the emissivity and absorptivity is not equal (Gas aGas ) and absorptivity must
be calculated separately. Therefore, the view factor from the flue gas to the exchange areas is calculated by

F 7i

Ai kG SS7i
e
A7

i 1; . . . ; NS :

As mentioned, there is no radiative heat exchange to other segments considered. Hence, the heat flux from flue gas must be
absorbed either by the exchange surfaces A16 or by the flue gas itself. Since the view factors to the exchange surfaces are
known, the view factor of the flue gas to itself can be expressed as:

F 77 1 

NS
X
F 7j :

j1

For multiple zones, as in this case, it is advantageous to formulate a linear equations system for the N S brightnesses of the N S
exchange areas. Using the Kronecker-symbol, as proposed in Baehr and Stephan [1],


dij

if i j

0;

1; if i j

the brightnesses Hi can be calculated for given temperatures by formulating a linear equation system with
N
S 1
X


dij  1  i F ij Hj i rT 4i ;

j1

where i is the emissivity and T i the temperature of the treated exchange area. The linear equation system, with N S equations, can be solved with common linear algebra.
In general, an energy balance of the incoming and outgoing flux density of an exchange area Ai leads to the radiative heat
flow Q i , as shown in Fig. 5 on the right-hand side. For a multi-surface enclosure, the incoming flux density Ai Ei consist of the
flux densities of the other exchange areas with respect to their view factor and brightness:

Ai Ei Ai

N
S 1
X

F ij Hj :

j1

Then, the radiative heat flow of the exchange area Ai is formulated as

Q_ i Ai

N
S 1
X

F ij Hi  Hj :

10

j1

For further investigations, the radiative heat flows of the exchange surfaces A16 and the gas are merged into the vector
Q_ rad;i Q_ 1 ; . . . ; Q_ NS ; Q_ 7 , where Q_ 7 is the heat flow from or to the gas volume.
2.3. Convective heat transfer
The convective heat transfer is a result of an energetic mass transport of fluid flow and macroscopic particle movement.
Due to the free or forced flow, heat energy can be transported to other places. Assuming a higher flue gas than furnace
insulation temperature, the convective heat flow from fluid to solid is given by Newtons law of cooling:

Q_ K aK AS T F  T S ;

11

where aK represents the convective heat-transfer coefficient between fluid and solid and AS the convective heat-exchange
surface of the solid. The convective heat-transfer coefficient a depends on the fluid properties and the length of the treated
system. The flow velocity has a particularly large effect, because it is responsible for the laminar or turbulent flow behavior.
Due to missing information, the exact computation of aK is associated with high uncertainty [32] and an approximation is
used in practice.
In this work, the gas is treated as an incompressible fluid, but measurement or calculation of gas velocity is not feasible;
hence it must be estimated. The main principle of FLOX burner is injecting fuel under high pressure to cause a very turbulent
flow of fuel and pre-heated air at the burners nozzle. This flammable flow spreads into the furnace, and combustion happens
not at one place, but over a wide space. The calculation of the velocity is therefore very difficult and an approximation is used
here. The effect of burner power on the gas velocity is unknown; hence the parameter must be estimated from measurement
data too.

22

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

The calculation of convective heat-transfer coefficient is largely based on the derivation for a fully developed turbulent
flow inside a pipe, where

Nu kA
;
L

12

with kA as the heat conductivity of air and L as the encircled length of the treated body is. As proposed by Gnielinski [12], the
mean Nusselt Number can be described by

"
 2=3 #
n=8Re  1000Pr
d
p
1
;
Num
l
1 12:7 n=8Pr2=3  1

13

with the resistance coefficient valid for Re P 4000

1
0:79 ln Re  1:64

14

For high temperature difference between fluid and solid it is necessary to consider the changing fluid properties such as the
flue-gas viscosity. Hufschmidt and Burck [14] proposed, to multiply Eq. (13) with a correction factor. Assuming an isothermal
behavior of the flue gas, the Nusselt Number is calculated by


Nu Num

Prflue
Prsolid

0:11
;

15

where Pr flue is the Prandtl Number of the flue gas at the flue-gas temperature and Pr solid the Prandtl number at the solid surface temperature. Assuming that the temperature difference between the insulation surface and flue gas is small, the correction factor is only calculated for the convective heat exchange between the strip and the flue gas with the local strip
temperature. The Reynolds number is computed with the flue-gas velocity between the non-moving surfaces of the furnace
and the moving strip surface separately such that

Rei

vL

16

with

v v 1 v 2 v 3; v 1

_ Gin
m
L;
mG

v 2 u kv ;

17

_ Gin is the
where u is the power state of the burners in one zone and kv is the associated tuning parameter for each zone, m
_ G is the total flue-gas mass inside the segment, m is the kinematic viscosity of flue gas and i the
incoming flue-gas mass, m
considered surface.
The Reynolds number for the non-moving surfaces is equal. In case of the strip surface i 5, the strip velocity is added
to the gas velocity such that

v3

0;

v strip

if i 5
if i 5

18

Computing the heat transfer coefficient ai for each exchange area A16 , the convective heat flow to each element can be
expressed as



Q_ convi ai Ai T gas  T i ;

19

where T i is the surface temperature of the exchange area i 1; . . . ; N S .


2.4. Conductive heat transfer inside the furnace insulation
The insulation-surface temperature inside the furnace changes relatively fast compared to the insulation-core temperature. To improve the accuracy of transient behavior of the model, the insulation mass is divided into several unequal parts
with thermal conduction as internal heat exchange. There are different approaches to divide the insulation layer. Wild et al.
[38] divided the furnace insulation into only two layers, a thin boundary layer and a thick outer layer, with experimental
found thicknesss of the layers. Depree et al. [7] computed a 2D mesh with different spacing to calculate the heat conduction
into the x- and y-axis. The thickness of the layers increase from thin at the furnace inside to thick at the outer side. In this
work, the approach of Marlow [20] is chosen. He proposed a spacing that doubled from the inside layer to the outside surface
and is realized with 4 layers, as shown in Fig. 6. Therefore, only the thin boundary layer inside the furnace has a radiative and
convective heat exchange and heat conduction is the internal heat flow from the inside boundary layer to the furnace outside
surface. The benefit of this method is the simple computation of the layer thicknesses, compared to a discretization of the
means of the finite difference method and an improved numerical stability and accuracy compared to a one-layer method.

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

23

Fig. 6. Radiative, convective and conductive heat flows of insulation.

In real industrial furnaces, the insulation consists of several layers of different material with individual heat conductivity.
Due to the unknown insulation material and structure, the heat conductivity is assumed to be equal over the whole insulation thickness. Assuming an isothermal temperature distribution within one layer, the conductive heat exchange to an insulation layer can be expressed by:

Q_ cond;i

8 kW
T Wi1  T Wi
for i 1
>
>
< di
kW
kW
AW  di T Wi1  T Wi  di1 T Wi  T Wi1 for i 2; 3
>
>
: kW
T Wi1  T Wi
for i 4
di1

20

where kW is the homogeneous distributed thermal heat conductivity of insulation, di are the thicknesss of the insulation
layer and T Wi represent the insulation layer temperatures. In cases where Q_ cond;i is positive, the layer i is heated otherwise,
for negative Q_ cond;i the layer i is cooled.
2.5. Conductive heat transfer between strip and transport roll
The strip is conveyed on water-cooled transport rolls through the furnace. Due to the contact of the hot strip with the
cooler transport roll, the strip cools down. This happens between zone 2 and 3 and again between zone 4 and 5. To consider
this heat loss, a simple heat conduction approximation is realized. In general, the measurement of the contact area between a
strip and a transport roll inside the hot furnace is very difficult. In this case, a constant area is assumed, knowing that due to
changing strip tension or scale between the strip and roll the real contact area changes permanently. In this work, the heat
loss is considered with

Q_ roll Aroll acontact T roll  T S v cor ;

21

with the contact area

Aroll 2pr roll wstrip

c
360

22

where rroll denotes the radius of the transport roll, wstrip is the strip width, c is the angle of contact, acontact is the conductive
heat transfer coefficient between strip and transport roll, T S is the local strip temperature, T roll is the surface temperature of
the transport roll. v cor is a velocity depending correction factor realized by

v cor 1



V in
;
1
V max

23

where V in is the actual strip velocity and V max is the maximum strip velocity. Q_ roll is only considered in the last segment of
zone 2 and 4. In all other segments, there is no transport roll and Q_ roll 0.
2.6. Heat loss considering the flap valve position in the flue-gas funnel
The flap valve in the flue-gas funnel is used to manage the combustion inside the furnace. If it is more closed, the combustion temperature and the heat exchange is marginal higher due to a flue-gas pressure change. In this work, the flue-gas is
treated as an incompressible gas, therefore this effect is considered by using the proposed function

/flap
C flap ;
Q_ flap 
/max

24

where /flap is the actual flap valve position, /max is the maximum flap valve position and C flap an estimated tuning parameter.
Therefore, Q_ flap denotes a heat loss of the flue gas depending from the flap valve position.

24

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

2.7. Transient temperature calculation of insulation, strip and flue gas


The turbulence inside a non-direct-fired furnace is much lower than inside a direct-fired one. Therefore, it is possible to
neglect the convective heat flow for furnaces with low turbulences inside. In this work, the total heat flow is derived for both,
radiation and convection. The total heat flow between the flue gas and the exchange areas can be computed by adding the
convective heat flow to the radiative heat flow.
2.7.1. Calculation of insulation-temperature change
Applying the first law of thermodynamics, the temperature change of an insulation layer can be described, in regard to
Fig. 6, as


@T Wl
1 _

Q rad Q_ conv Q_ cond;l Q_ loss ;


cpW mWl
@t

l 1; 2; 3; 4;

25

where Q_ rad and Q_ conv only exist for the inside furnace surface l 1 and Q_ loss only exist for the outer layer l 4. Q_ cond;l is the
conductive heat flow from or to this insulation layer. The furnace heat loss is only considered as a convective heat flow from
the outer layer to ambiance and is calculated by



Q_ loss AW aloss T amb  T W4 ;

26

where T amb is the ambient temperature and T W4 is the temperature of the outside insulation layer l 4. aloss is the convective heat-transfer coefficient that is determined using a simplified square profile of the considered furnace zone with natural
convection caused by the higher surface temperature to the ambiance.
2.7.2. Calculation of strip-temperature change
Because of the moving strip, the strip-temperature calculation is an open process, and the incoming and outgoing stripenthalpy flow must be considered. Therefore, considering the notation of Fig. 4, the strip temperature can be obtained from

cpS mS tT S t

@
_ Sin tcpin T Sin t  m
_ Sout tcpout T Sout t:
Q_ rad Q_ conv Q_ roll m
@t

27

For calculation of the strip-temperature change, the specific-heat capacities cpS ; cpin and cpout are considered as external
_ Sin t and
parameters. The temperature-depending determination can be found in detail in Section 2.8.1. The mass flows m
_ Sout t and the strip mass mS t inside the considered segment are computed in a separate strip-transport function.
m
Applying of partial differentiation to Eq. (27) leads to

cpS T S t

@mS t
@T S t
_ Sin tcpin T Sin t  m
_ Sout tcpout T Sout t:
cpS mS t
Q_ rad Q_ conv Q_ roll m
@t
@t

28

Assuming that

@mS t
_ Sin t  m
_ Sout t
m
@t

29

and some transformations yield




@T S t
1
_ Sin tcpin T Sin t  m
_ Sout tcpout T Sout t cpS T S t m
_ Sin t  m
_ Sout t :

Q_ rad Q_ conv Q_ roll m


@t
cpS mS t

30

2.7.3. Calculation of gas-temperature change


Assuming that the gas is an incompressible fluid and the total gas mass also depends on the fuel injection, the gasenthalpy flow can be derived through partial differentiation, according to Fig. 4, to give

cG T G t

@T Gt
@mt
cG mt
H_ Gin H_ Fin  H_ Gout Q_ rad Q_ conv Q_ flap ;
@t
@t

31

where

_ Gin tcGin T Gin t


H_ Gin m

32

is the incoming flue gas enthalpy flow,

_ Gout tcGout T Gout t


H_ Gout m

33

is the outgoing flue gas enthalpy flow and

_ F tHU m
_ air tcair T air
H_ Fin m

34

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

25

is the enthalpy flow of combustion-gas injection. Similar to the strip, the specific heat capacity of the flue gas (cGin ; cGout ) is
_ Gout t the outgoing
_ Gin t is the incoming, m
temperature-dependent and its determination can be found in Section 2.8.1. m
_ air t is the additional fresh_ F t is the fuel-mass flow, HU is the specific heat content of fuel and m
flue-gas mass flow, m
air mass flow. cair is the specific heat capacity of the fresh air at temperature T air . This separation of fuel and fresh-air mass
flow is done because the used FLOX burner allows a separate adjustment of the mass flows in a wide range. With the
substitution

@mt
_ Gout t m
_ F t m
_ air t
_ Gin t  m
m
@t

35

and insertion of Eqs. (32)(34) into Eq. (31), it follows

@T G t
1 
_ F tHU m
_ air tcair T air  m
_ Gout tcGout T Gout t Q_ rad Q_ conv Q_ flap
_ Gin tcGin T Gin t m

m
@t
cG mt


_ Gin t  m
_ Gout t m
_ F t m
_ air t :
cG T G t m

36

2.8. Calculation of various material properties


To improve accuracy of the model, some material properties are calculated as functions of the operating point. Fig. 7
shows the values of the parameters S ; qS and cpS drawn over the temperature.
2.8.1. Calculation of strip-heat capacity
For a specific stainless steel, the calculation of specific heat over an wide temperature range was measured by Rudtsch
et al. [29]. The polynomial interpolation of these values for specific heat capacity leads to the empirical function:

cpx T c1 c2  T c3  T 2 c4  T 3 c5  T 4


J
;
kg K

37

with c1 475:613099; c2 0:322294; c3 5:504965  104 ; c4 6:948614  107 ; c5 3:146146  1010 and T is the
strip temperature. The average specific heat capacity cpS of the strip inside the considered zone is calculated with

cpS cpx T with T

T Sin t T Sout t  273


;
2

38

for the incoming strip with

cpin cpx T with T T Sin t  273;

39

and for the outgoing strip with

cpout cpx T with T T Sout t  273:

40

Fig. 7. Strip property value vs. temperature: emissivity (red), specific heat capacity (black), density (blue). (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)

26

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

2.8.2. Calculation of strip density


As stated in [11], the strip density also depends on temperature and can be calculated with

qS T

q0
1 aT  2933


kg
;
m3

41

where q0 is the density of stainless steel at 20 C, a is the thermal expansion coefficient of stainless steel and T is the strip
temperature.
2.8.3. Calculation of strip emissivity
The computation of the strip emissivity is very complex and depends on different parameters such like the surface properties and the temperature. For approximation, investigations of Liu et al. [17], Maynard [21], Roger [28] are used to formulate an empirical function considering the strip temperature:

S T S

lower


arctan5pT app p2

S

upper

42

shift
with T app Tz
, where T is the strip temperature, T max is the self-defined maximum strip temperature and zshift is the inflecT max

tion point of arctan function. The term 5pT app is derived from measurement data that characterizes the rate of increasing
emissivity depending on strip temperature. Slower is the lower and Supper the upper emissivity bound of the strip.
2.8.4. Calculation of flue-gas properties
British Gas has derived the following properties of the stoichiometric combustion products of natural gas for temperature
up to 4000 K. Following values are taken from Rhine and Tucker [27].
Using the method of least squares, these properties U have been fitted to fifth order polynomials of the form:

U a bz cz2 dz3 ez4 fz5 ;

43

where z T1400
and a, b, c, etc. are the polynomial coefficients, which are given in Table 1.
200
h i


In Table 1, cpx kgJ K is the heat capacity, m Nm2s the dynamic viscosity, k mWK the conductivity and Pr the Prandtl Number of
the flue gas. The specific heat capacity cGin is calculated with Eq. (43) at the flue gas temperature T Gin ; cGout in an analogous
manner at T Gout and the average specific heat capacity cG with T G .
2.8.5. Calculation of gas density
The gas density function depending on temperature is defined for ambient pressure as

qT

p
RT

44

with p = 101325 Pa and R = 287 J/kg K.


2.9. Combustion dynamics
To consider the combustion dynamics of changing power state and fresh air, a discrete first-order LTI-system is used and
realized with

yk T

1
dyn

Dt

uk  yk1 yk1 ;

45

_F m
_F m
_ air T is the output vector at time step k, yk1 is the solution at the time step before, uk m
_ air T is the
where yk m
input vector, T dyn is the time constant and Dt the sample time.
Table 1
Polynomial coefficients of the properties of flue gas.

cpx

m
k
Pr

1401.65961
50.249626
0.99972
0.704524

36.60067
7.731996
0.012950
0.006326

4.43026
0.12346
0.000153
0.000334

0.316066
0.008790
0.000016
0.000137

0.046715
0.001328
0.0
0.000014

0.0099
0.000167
0.0
0.000003

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

27

2.10. State-space formulation


The equations derived above can be transformed to a nonlinear state-space representation that is suitable for further
analysis, observer or controller design. The subsystem of one segment, according to Fig. 4, can be expressed as

_
xt
f x; H; t; u;

46

where f is the corresponding nonlinear function. H denotes the vector of model parameters to be estimated,


_ Gin ; m
_ F; m
_ air ; m
_ Sin ; m
_ Sout ; /flap ; V in ; T Sin ; T Gin T the vector of manipulated (input)
x T 1 ; T 2 ; . . . ; T 7 T the state vector, and u m
variables. To realize the model structure of a whole furnace, Eq. (46) can be extended by merging several segments to
one system. In this case, some input variables (such as T Sin ) can be taken from the state vector/matrix, because they are
results of the previous segment.
2.11. Numerical solution and stability analysis
To compute the temperatures of the strip, the gas and the insulation of one segment, the finite-difference method is
used. A stability analysis based on Tian et al. [36] was done, and a computation with an explicit integration method like
fourth-order RungeKutta is not feasible because of stability constrains. The limitations result from furnace configuration
and cannot be changed. The temperature dependence of some property parameters is a second reason why they are implicitly computed. Furthermore, the dynamic behavior of the furnace is equal to a stiff system where the usage of implicit integration methods is necessary. To avoid stability problems the more computationally-intensive implicit integration method is
chosen. Contrary to the commonly used explicit (fourth-order) RungeKutta method, implicit Heun of second order is suitable for stiff systems. The system follows the second law of thermodynamics and stability of system is given [33]. To proof
stability of the linearized system, a stability analysis of the used differential equations has been performed by computing the
Jacobian matrix and checking the Euclidean norm, which confirms the usability for a wide range of plant operating points.
2.12. Computational procedure
The computational model considers heat-transfer fundamentals with opposite motion directions of gas and strip.
Therefore, the necessary parameters are computed separately in every simulation step. Fig. 8 shows the flow chart of the
computational realization, which can be divided into four main parts: the pre-processing and the initialization procedure,
the transport modules, the heat flow calculation and the energy balance. The pre-processing and the initialization procedure
can be described in detail:
 With the furnace geometry data, the simulation model is customized to the given case study. Defining the number, the
width, the height and the length of the zones as well as the burner arrangement and insulation thickness, the model
calculates all necessary surfaces.
 The measurement data used here are a record over a whole month, including maintenance and down-time periods. For
the identification procedure, a feasible time frame of about 150 min is selected and outliers are replaced by specific
values. Extending the data from measurement time base T meas 10 s to simulation time base T sim 0:5 s the data are
used as input for the model.
 The initial conditions inside the furnace, such as the gas T G and insulation temperature T Wl , are set up from measurement data. To reduce the computational time, the pre-allocation of the initial strip width, thickness and temperature for
each segment is carried out.
 Because the initial states are only estimates, a steady-state calculation is carried out and the results are set as initial
condition of the simulation.
The transport module can be described in detail:
 The first module of the main simulation loop calculates the gas-mass transport of each segment, considering the incoming
gas-mass flow and power state. The exit gas-mass flow and total gas mass within the segment are saved as the result of
calculation.
 Secondly, the strip transport module calculates the incoming and outgoing strip properties and the necessary average
strip mass mS inside the segment. The strip-mass flow is calculated by multiplication of strip thickness, width, and density with the strip velocity.
The heat flow calculation can be described in detail:
 The computation of the flue-gas properties and dynamics in a first step, and the corresponding heat-transfer coefficients
in a second step leads to the convective heat flow of each exchange area.

28

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

Fig. 8. Flow chart of the annealing-furnace model.

 Calculating the view factor in the first step, the absorption of the flue gas is considered in a second step. In the third and
fourth step, the emissivities and brightnesses of each exchange area are computed to get in the last step the radiative heat
flow from and to each exchange area.
 Conductive heat flow within the furnace insulation is calculated by Fouriers law. Additionally, the furnace heat loss to the
ambiance is considered in this module.
The energy balancing is carried out in the last step, where the temperature changes of the strip, the gas and the insulation,
considering the enthalpy flows as described in Eqs. (25), (30) and (36) are computed.

3. Model-parameter identification
When using the developed model to predict the strip and furnace temperatures in steady-state or dynamic conditions and
further for controller design or real-time optimization, all model parameters must be known. However, the product

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

29

characteristics and operating conditions of annealing processes frequently change, and a set of values for the model parameters, determined from standard physics-based relationships, cannot ensure sufficient accuracy and utility of the mathematical model for online and real-time applications in a particular furnace. Therefore, model parameters should be adjusted such
that the predicted temperatures are in agreement with the actual measured steel temperatures.
3.1. Basic identification method
The measurement of strip temperature inside a direct-fired furnace is very difficult and often not possible. In this case
study, only gas-temperature measurement data at each zone and strip-temperature measurement data at the exit of the furnace are available. The measurement positions of gas temperature are in the strip running direction at the end of each zone.
The approach of the simulation model is an isotherm behavior within one segment. Therefore, the simulation data of the last
segment of a zone are taken for the identification procedure. Fig. 9 shows the distributed gas-temperature measurement
positions TG17 and the strip-temperature measurement point TS at the end of furnace. The characteristic of the basic gas
control loop is unknown so that the furnace and the basic gas controllers have to be merged to one system for identification.
This is due to the used FLOX burners that have to be heated up before the operational point can be changed. The unknown
logic manages this pre-heating process and is not separately realized.
The derived mathematical model is used to estimate the temperature dynamics, especially of the strip, inside the furnace.
To estimate the uncertain model parameters, such as the emissivities, parameter identification was carried out. For identification, the inputs and initial conditions of plant are taken from measurement data um and the parameters are chosen such
that the model output closely follows the measured plant output. Identifying the parameters of the plant the incoming strip
parameters, such as the thickness or the width, and the reference values of the gas controllers serve as the input variables of
system. For given initial conditions x0 , inputs um and outputs ym , the parameter vector H has to be estimated from a set K to
minimize the least-squares error ek ym k  yk, where yk is the model output at the time step k 1; 2; . . . ; N, with N is
the number of simulation steps.
Taking similar notations as in Ljung [18], the identification problem can be formulated as

^ arg minVH
H

47

H2K

subject to Hmin 6 H 6 Hmax ;


where

VH

1
jjFHjj22 :
2

48

To compute the error at one furnace zone el H; l 1; 2; . . . ; 8, the temperature differences between simulated and measured
data over the data length N are constructed:

^1
el H y1  y

^2
y2  y

^3
y3  y



^N  ;
yN  y

49

^k is the value of calculated and yk the value of measured temperature at time step k. The error vectors of each data set
where y
l are merged into the vector

3
eTG1 H
6 e H 7
7
6 TG2
7
6
7
6
.
.
FH 6
7:
.
7
6
7
6
4 eTG7 H 5
2

eTS H

Fig. 9. Inputs and outputs of model identification (TG: gas temperature; TS: strip temperature; GC: gas control loop).

50

30

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

There are many existing approaches to solve Eq. (47), where some of them are built into software packages like MATLAB or
MAPLE. For smooth error functions VH, gradient-based methods can be successfully employed to determine an estimate
of H. Since the model is nonlinear, a nonlinear least-squares optimization algorithm with the iterative trust-region reflective
method is used in this work, to update H in the direction of the steepest descent of error function, i.e.

Hs1 Hs  grVjHs ;

51

where rVjHs is the gradient of V evaluated at Hs in the iteration step s; g is the learning rate. The iteration continues until
rV 0 or some other stopping criterion is satisfied. A detailed discussion of steepest descent and other gradient-based optimization methods is given in Bishop [2] and Dennis and Schnabel [6]. The optimization routine is realized as shown in
Fig. 10. Starting with an initial set of parameters Hs , the model is run once to compute the first error vector V as defined
in Eq. (48). Depending on the stopping criterion, a new parameter vector Hs1 as in Eq. (51) described is calculated and
the simulation is runs again until a specified threshold of the criterion is achieved.
3.2. Parameter set to be estimated
In this case study, a set of 16 parameters (Hi ), as specified in Table 2, is identified:
 Insulation density and conductivity. Due to model simplification, the wall isolation is treated as a single mass inside one
segment and no heat conduction between the segments is assumed. To match the dynamic temperature characteristic,
the insulation density qwall and the conductivity kwall is estimated.
 Heating value. The heating value HU of the used natural gas is varying over the year and measurement data are not
available. Based on the energy content of domestic gas, the resulting heating value is estimated.
 Emissivities. The approximation of the emissivity is complex and often not feasible due to unknown parameters, such as
chemical or surface properties. Surface emissivities are very difficult to measure in practice, and vary greatly due to surface conditions that may also change with age [8]. In this work, the emissivity is considered as an unknown parameter
and separately estimated for the wall W . The strip emissivity is approximated by three parameters Slower ; Supper and
zshift , as described in Eq. (42).
 Absorptivity. The flue-gas behaves as a gray gas, wherefore the absorptivity in the flameless case G is estimated.
 Burners flow velocity. As already explained in Section 2.3, a tuning factor kv (Eq. (17)) is introduced to take into account
the uncertainty in the calculation of gas velocity, indicating the turbulence inside the furnace, due to the burners nozzle
configuration. Depending on the injection angle, the gas flow velocity can vary near the strip surface.
 Combustion dynamics. The combustion of the flammable mixture is a dynamic process, and gas temperature does not
raise or decrease instantly due to changes of input mixture. Therefore, the heat exchange from the fuel to the gas is
approximated with a first-order (LTI) system with the time constant T dyn .

Fig. 10. Flow chart of basic identification method.

31

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539


Table 2
Model parameters (Hi ) to be estimated, and their bounds specified for the parameter-estimation process.
No.

Name

Description

Unit

Lower bound

Upper bound

1
2
3
4
5
6
7
8
9
1016

qW

Insulation density
Insulation emissivity
Insulation conductivity
Gas absorptivity
Heating value of fuel
Strip emissivity (lower bound)
Strip emissivity (upper bound)
Strip emissivity (point of inflection)
Time constant of combustion dynamics
Tuning knob of burners flow velocity in each zone

kg/m3

W/m K

MJ/kg

K
s
m/s

200
0
0.001
0
30
0
0
300
0.1
0

3000
1
1
1
60
1
1
1500
300
100

W
kW

aG
HU

S;lower
S;upper

zshift
T dyn
kv1-v7

To avoid running into a local minimum, the model parameters are set into physical constraints, such as the maximum air
velocity or variance of similar parameters. The bounds of each parameter are given in Table 2.
3.3. Sequential identification approach
Due to the complexity of the model, a sequential approach for parameter identification is proposed and implemented, i.e.
the parameter estimation is carried out sequentially in several steps and iterations. To get information about the effect of
each parameter, a structural identifiability analysis to determine the parameter subset Hstruc , that is identifiable for the given
mathematical model structure is carried out. To ensure that the parameter subset Hstruc is identifiable from the given
measurement data, in a second step a practical identifiability analysis based on a sensitivity analysis is performed. The result
of this step is the parameter subset Hprac , where Hprac 2 Hstruc .
The parameters are then divided into the three effect groups: Hstatic contains the mainly static effecting parameters
(W ; kW ; aG ; HU ; S;lower ), Hdynamic contains the mainly dynamic effecting parameters (qW ; T dyn ) and Hdistribution contains the
parameters that mainly effect the heat exchange distribution over the furnace length (S;upper ; zshift ; kv1-v7 ).
The sequential identification approach (Fig. 11) starts with the estimation of the parameters for Hstatic and is carried out
until an optimal fitting of measurement and simulation data for this effect group is reached. The parameters are stored in a
global file and used for the next iterations for the other effect groups. As second step, the parameters for Hdistribution are optimized until the optimal fitting for this group is reached. In the next step, the groups Hstatic and Hdistribution are alternately
identified until both groups shows no benefit in the error vector. In the third step, the identification routine is extended
by the last effecting group, Hdynamic , and the estimation of the three groups is carried out alternately until no benefit of

Fig. 11. Flow chart of the sequential identification approach.

32

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

the error vector in all groups is reached. With this sequential identification approach, optimal parameter estimates for all
groups could be found.
The advantage of the proposed sequential identification algorithm is the estimation of only identifiable parameter subsets, whereas not identifiable parameters are not considered. In this case, no computational time is lost on not identifiable
parameters. For instance, the dynamic parameter shows their highest sensitivity to the model output at the region of optimal
parameters for the static and distributional group.
The sequential identification procedure is presented for a reduced parameter set, where the heat content HU characterizes
Hstatic , the inflection point zshift characterizes Hdistribution , and qW characterizes Hdynamic . The time series in Figs. 12 and 13
shows the evolution of strip and gas temperature exemplary for the identification approach with the mentioned parameters.
The figures show the results of the parameter identification: (a) at the beginning of the whole routine, (b) first optimal
estimates for the Hstatic group (after the second square of Fig. 11), (c) after finding the optimal parameters for the Hstatic
and the Hdistribution parameter group (after the reaching of the first stopping criterion), and (d) for the optimal parameters
found in all groups.

3.4. Identification and verification


The validation of the model is carried out in two steps, wherefore two different measurement data sets for validation
process are used. The first (identification) data set is for identification process only, the second for verification of the derived
model with the estimated parameters from step one. In both, the treated measurement data shows some changes in strip
width and thickness and a wide dynamic velocity profile; see Figs. 14 and 16. Due to confidentiality reasons, the data are
normalized for all shown process variables. To catch the slow temperature dynamic of wall insulation a long duration of
simulation was chosen. The estimated strip parameters have a direct effect on the strip and gas-temperature characteristics,
so that they are observable.

Fig. 12. Evolution of the strip temperature using the sequential identification approach.

33

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

3.4.1. Identification process


For parameter identification, an investigation time frame of 150 min is taken. The incoming strip parameter are shown in
Fig. 14. As described in Section 3.1, the estimation of parameters is carried out by the comparison of real measurement data
and simulation data, shown in Fig. 15. Due to similar characteristic of the temperature trend inside the furnace for each zone,

Fig. 13. Evolution of the gas temperature in Zone 7 using the sequential identification approach.

[v i /v nor m] [t i /t nor m] [w i /w nor m] [ i / nor m]

Strip width(red), thickness(blue), velocity(black) and ap posion [green]


1.1
1
0.9
0.8
0.7
0.6
0

10

20

30

40

50

60

70
80
Time [min]

90

100 110 120

130 140

150

Fig. 14. Time plot of strip width (red), thickness (blue), velocity (black) and flap position (green) of identification data. (For interpretation of the references to
color in this figure legend, the reader is referred to the web version of this article.)

34

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

the results are shown for selected zones only. Fig. 15a shows the time trend of the gas temperature at the zone 1, where the
strip is inserted into the furnace, Fig. 15b shows the gas temperature of zone 3 at the middle of the furnace, Fig. 15c shows
the time trend of the gas temperature in zone 7 at the end of the furnace and Fig. 15d shows the exit strip temperature.
Considering that the mathematical model is a simplified representation of the real plant, the model output with the estimated parameters is in good accordance to the measurement data. Only in the first zone, a small offset between measured
and simulated values can be observed. The reason can be found in the segment spacing and the assumption of an isothermal
behavior of the flue gas within one zone. Because the first zone is mainly used as pre-heating area, where the burner are
arranged at the end of the zone, as suggested in Fig. 9, the average turbulence in zone 1 is lower then in the other zones.
Thus, the circulation of the flue gas is lower and the temperature at the top and the bottom of the furnace zone is different.
Taking into account that the thermocouple at the real plant is situated under the ceiling, the measured temperature is higher
than the average simulated one. Table 3 shows the estimated parameters found in the identification process.

Fig. 15. Measurement data (red) simulated data (black): (a) gas temperature in zone 1, (b) gas temperature in zone 3, (c) gas temperature in zone 7, and (d)
exit strip temperature. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 3
Estimated model parameters (Hi ).
No.

Name

1
2
3
4
5
6
7
8
9

qW

10
11
12
13
14
15
16

Description

Unit

Value
3

zshift
T dyn

Insulation density
Insulation emissivity
Insulation conductivity
Gas absorptivity
Heating value of fuel
Strip emissivity (lower bound)
Strip emissivity (upper bound)
Strip emissivity (point of inflection)
Time constant of combustion dynamics

kg/m

W/m K

MJ/kg

K
s

1000
0.89
0.016
1.37
55.4
0.015995
0.292523
375.25
1.5

kv1
kv2
kv3
kv4
kv5
kv6
kv7

Tuning
Tuning
Tuning
Tuning
Tuning
Tuning
Tuning

m/s
m/s
m/s
m/s
m/s
m/s
m/s

5.565
5.565
5.565
5.565
17.40
0.0612
26.896

W

aloss
aG
HU

S;lower
S;upper

knob
knob
knob
knob
knob
knob
knob

of
of
of
of
of
of
of

burners
burners
burners
burners
burners
burners
burners

flow
flow
flow
flow
flow
flow
flow

velocity
velocity
velocity
velocity
velocity
velocity
velocity

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

35

3.4.2. Verification process


The verification of the derived model with the estimated parameters is done with similar, but different measurement
data. The incoming strip parameters are shown in Fig. 16.
The comparison of real measurement and simulation data confirms the accuracy of developed model. Figs. 17 and 18
present the temperature trends of the continuous furnace for a duration of about 150 min. In this period, 8 different strip
coils are annealed through the furnace. The control objective is to reach and hold the strip temperature at a specified
reference value. Disturbances like changes in velocity or strip properties have a direct effect on the exit strip temperature.

[v i /v nor m] [t i /t nor m] [w i /w nor m] [ i / nor m]

Strip width(red), thickness(blue), velocity(black) and ap posion [green]


1.1
1
0.9
0.8
0.7
0.6
0

10

20

30

40

50

60

70 80 90 100 110 120 130 140 150


Time [min]

Fig. 16. Time plot of strip width (red), thickness (blue), velocity (black) and flap position (green) of validation data. (For interpretation of the references to
color in this figure legend, the reader is referred to the web version of this article.)

Fig. 17. Gas temperature: (a) zone 1, (b) zone 3, (c) zone 5, and (d) zone 7: measurement data (red) simulated data (black). (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of this article.)

36

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539


Zone 7 strip temperature: meas (red), simu (black)
0.92

Temperature [-]

0.91
0.9
0.89
0.88
0.87
0.86
0.85
0

10

20

30

40

50

60

70

80

90

100

110

120

130

140

150

Time [min]
Fig. 18. Exit strip temperature: measurement data (red), simulated data (black). (For interpretation of the references to color in this figure legend, the reader
is referred to the web version of this article.)

Table 4
Comparison of zone-wise prediction quality measures.
Data pos.

Zone
Zone
Zone
Zone
Zone
Zone
Zone
Strip

1
2
3
4
5
6
7

Identification data

Verification data

MAE [K]

MSE [K2]

RMSE [K]

MAX [K]

MAE [K]

MSE [K2]

RMSE [K]

MAX [K]

5.7218
2.8519
3.9679
3.1518
3.1136
5.2033
2.2113
4.8090

42.738
15.460
24.018
16.294
16.466
38.188
7.819
38.111

6.5375
3.9319
4.9009
4.0366
4.0578
6.1796
2.7963
6.1734

15.428
23.565
17.176
14.356
16.501
15.080
9.224
35.272

16.002
7.4780
3.6460
7.1179
3.0341
5.8463
2.9594
4.5445

269.93
67.215
20.885
61.509
13.740
48.216
12.939
33.927

16.430
8.1985
4.5700
7.8428
3.7068
6.9438
3.5971
5.8247

26.089
15.905
16.694
23.634
14.916
20.075
14.529
31.694

As seen in Fig. 17, the gas-temperature estimation agrees very well with measurement data for zone 3, 5 and 7. The
temperature trend of the zones 1 have a small offset caused by different assumed incoming strip temperature. Because of
missing measurement data, the unknown incoming strip temperature is assumed to be constant.
Fig. 18 shows the strip temperature at the exit of the furnace. The comparison shows a good transient behavior. As
mentioned for the gas-temperature trend of the zone 1, the strip-temperature trend has a low offset. The reason could be
a lower incoming strip temperature into the furnace as estimated for identification data.
To quantify the model-prediction quality, some measures have been computed for each data position: the mean absolute
error (MAE), the mean squared error (MSE), the root mean squared error (RMSE), and the maximum difference (MAX)
between the measurement data and the simulated trends. The results are given in Table 4 for both, identification data
and verification data.
On the left side of Table 4 the quality measures for the identification data are shown. The model fit to the real process is
good in all zones. The gas-temperature estimation is within the tolerance band of measurement sensor. It has to be
mentioned that there is no dynamic measurement behavior realized and the comparison is carried out for the estimated
gas temperature and measured gas temperature. To protect the gas temperature sensor, it is coated in a thin ceramic pipe
and heat transfer from the coat surface to the sensor is a result of heat conduction. Due to that model simplification, the
variance of the estimated gas temperature is greater than measured.
The strip-temperature estimation shows a very good fit to the measurement data and only for the time range of 45 min a
difference to the measured value appears. Therefore, it is possible to use the model for strip temperature prediction.
As expected, the quality measures for verification process are different to those for the identification process. The
gas-temperature offset of zone 1 is reflected in the high value of MAE, although the dynamic behavior can be approximated
very good. Similar to the identification data, the gas-temperature estimation is in good accordance with the measurement
data.
Generally, the prediction of the strip and gas temperature is very good. Compared with other simulation studies,
i.e. Strommer et al. [35], the prediction quality is even higher. This could be reached by using a sequential identification
approach. As a result, the parameter estimation for the furnace properties and heat-exchange variables leads to a applicable
simulation model. Especially the dynamic behavior of disturbances, such as a strip-property changes, can be predicted with a
very good accordance to the measurement data.

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

37

4. Conclusions
In this paper, a mathematical model for a horizontal continuous annealing furnace was developed by heat-transfer fundamentals. For the calculation of strip, gas and wall temperatures, both convective and radiative heat transfer are considered.
To apply the finite difference method, the furnace is split into several segments, and computation is solved by an implicit
integration method. The strip and flue gas are treated as 1-D models, so that there is no heat conduction within the material.
Conductivity between several layers of the furnace insulation was considered to take the transient temperature behavior
into account.
Because of the direct-fired furnace, the gas transport is calculated and is considered for convective heat flow. Due to the
segment-wise computation, the model is completely variable in length, height and width and can be easily transferred to
other applications.
Tuning the parameters of the model is crucial for its prediction quality and reflecting the heat characteristics of a
particular furnace. Model-parameter identification is carried out by a nonlinear constrained least-squares algorithm, embedded in a sequential identification scheme, to estimate optimal values of uncertain parameters, such as emissivities. It is
shown that the model predictions agree well with measurement data and is suitable for accurate prediction of strip, gas
and wall temperature along the annealing furnace, under both steady-state and transient conditions. It still remains to
improve the model prediction at low strip velocity by recording the incoming strip temperature.
Since the parameter-identification approach applied in this paper is an iterative method for the whole annealing furnace,
it is computationally intensive and still not yet suited for real-time (online) use. This would give the possibility to detect
changes of important process parameters and initiate a suitable counteraction. A possible improvement direction is to
reduce the number of iterations that are used for the sequential distributed parameter-estimation technique that is
proposed. The usage of parallel computing techniques should also lead to a shorter optimization duration.
The parameter identification for the considered annealing plant was a very difficult task, and revealed that model
parameters of such complex models cannot be straightforwardly estimated from any given data set of measured data. This
may be due to identifiability problems. Therefore, work is going on to develop a new parameter-estimation strategy for such
complex physically-based models, performing structural and practical identifiability analysis in a first stage prior to the
parameter-estimation process itself.
Appendix A
In this work, it is assumed that there is no significant heat flow from one segment to the other. In the following section, it
is shown that this assumption is justified. The following calculation is based on the work of Ehlert and Smith [9]. Fig. A.19
shows the considered surface location to each other. For surfaces that lies parallel to each other, the view factor can be computed by

F 12
with

2 X
2 X
2 X
2
X


1
1ijkl G xi ; yj ; gk ; nl ;
x2  x1 y2  y1 l1 k1 j1 i1


 1
h
i1=2
yg
y  g x  n2 z2
tan1
1=2
C
B
xn2 z2 
B

C
C
h
i1=2
1 B
C:
B
2
xn
1
2
G

z
tan

x

n

y

C
B
1=2
2p B
C
yg2 z2 
A
@
h
i
2
2
z2
2
 2 ln x  n y  g z

A:1

A:2

For surfaces that lies in a perpendicular plane to each other, the view factor can be determined by

F 12
with

2 X
2 X
2 X
2 h
X

i
1
1ijkl G xi ; yj ; gk ; nl ;
x2  x1 y2  y1 l1 k1 j1 i1

0
 2

2 1=2
tan1 K 
1 @ y  g x n
h

G


2p 1 x2 n2 ln1 K 2  y  g2 ln 1
4

A:3

1
1
K2

i A;

A:4

where


1=2
K  y  g= x2 n2
:

A:5

38

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

Fig. A.19. Arrangement of considered surfaces to each other, parallel (left), perpendicular plane (right) Ehlert and Smith [9].

Fig. A.20. Considered surfaces.

Table A.5
Results of the view-factor calculation.
Name

Value of the view factor

F 14
F 13
F 24
F 23

0.0049
0.1811
0.0223
0.5734

Eqs. (A.1)(A.5) are used to compute the view factors from segment one to the segment 2. First, the view factor F 14 of the
furnace insulation surface at the side of segment 1, colored in red, to the strip surface inside the segment 2, colored in blue,
is computed. Secondly, the view factor F 24 from the furnace insulation surface at the top of segment 1, colored in green, to
the strip inside segment 2 (blue) is calculated. This view factors are compared with the view factors F 13 and F 23 , which are
the view factors from the surfaces inside the segment 1 to the strip. Fig. A.20 illustrates the comparison. The considered
segment is 1.25 m high, has a width of 2.5 m and a length of 10 m. The strip width is 2 m and there is no participating
gaseous medium considered, the results of the computations are shown in Table A.5.
The comparison of the view factors shows that the radiative heat flow from segment 1 to 2 is significantly lower than
inside segment 1. For the considered case, FF 14
0:0272 and FF 24
0:0388. The summarized heat flow inside the segment 1
13
23
to the strip is F 11 F 23 2F 13 0:9355, whereas the heat flow from segment 1 to the strip in segment 2 is
F 12 F 24 2F 14 0:0321, which is ca. 3.4% of the heat flow inside the segment 1. As assumed, the radiative heat flow from
one segment to the other can be neglected.
References
[1] H.D. Baehr, K. Stephan, Heat and Mass Transfer, second ed., Springer-Verlag, 2006.
[2] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press, 1995.

S. Zareba et al. / Simulation Modelling Practice and Theory 60 (2016) 1539

39

[3] C. Capdevila, C. Garcia-Mateo, F.G. Caballero, C. Garca de Andrs, Neural network analysis of the influence of processing on strength and ductility of
automotive low carbon sheet steels, Comput. Mater. Sci. 38 (2006) 192201.
[4] S.R. Carvalho, T.H. Ong, G. Guimaraes, A mathematical and computational model of furnaces for continuous steel strip processing, J. Mater. Process.
Technol. 178 (2006) 379387.
[5] T.C. Chen, C.H. Ho, J.C. Lin, L.-W. Wu, 3-D temperature and stress distributions of strip in preheating furnace of continuous annealing line, Appl. Therm.
Eng. 30 (2010) 10471057.
[6] J. Dennis, R. Schnabel, Numerical Methods for Unconstrained Optimization and Nonlinear Equations, Society for Industrial and Applied Mathematics,
1983.
[7] N. Depree, J. Sneyd, S. Taylor, M.P. Taylor, J.J.J. Chen, S. Wang, M. OConnor, Development and validation of models for annealing furnace control from
heat transfer fundamentals, Comput. Chem. Eng. 34 (2010) 18491853.
[8] N.B. Depree, Mathematical Modelling of an Annealing Furnace, Ph.D. thesis, University of Auckland, 2009.
[9] J.R. Ehlert, T. Smith, View factors for perpendicular and parallel, rectangular plates, J. Thermophys. Heat Transfer 7 (1993) 173174.
[10] L. Gardner, K. Ng, Temperature development in structural stainless steel sections exposed to fire, Fire Saf. J. 41 (2006) 185203.
[11] Gesellschaft, VDI, VDI-Wrmeatlas, VDI-Buch, Springer, 2013.
[12] V. Gnielinski, New equations for heat and mass transfer in turbulent pipe and channel flow, Int. Chem. Eng 16 (1976) 359368.
[13] H.C. Hottel, A.F. Sarofim, Radiative Transfer, McGraw-Hill, 1967.
[14] W. Hufschmidt, E. Burck, Der einfluss temperaturabhngiger stoffwerte auf den wrmebergang bei turbulenter strmung von flssigkeiten in rohren
bei hohen wmestromdichten und prandtlzahlen, Int. J. Heat Mass Transf. 11 (1968) 10411048.
[15] Z.W. Kang, T.C. Chen, Three-dimensional temperature distributions of strip in continuous annealing line, Appl. Therm. Eng. 58 (2013) 241251.
[16] S. Li, Q. Chen, G.B. Huang, Dynamic temperature modeling of continuous annealing furnace using GGAP-RBF neural network, Neurocomputing 69
(2006) 523536.
[17] Y.F. Liu, Z.L. Hu, D.H. Shi, K. Yu, Experimental investigation of emissivity of steel, Int. J. Thermophys. 34 (2013) 496506.
[18] L. Ljung, System Identification, John Wiley & Sons, Inc, 2001.
[19] Marlow, D., Mathematical Modelling of the Heat Treatment in the Continuous Processing of Steel Strip, Ph.D. thesis, University of Wollongong, 1995.
[20] D.O. Marlow, Modelling direct-fired annealing furnaces for transient operations, Appl. Math. Model. 20 (1996) 3440.
[21] K.R. Maynard, Total Hemisperical Emissivity of Very High Temperature Reactor (VHTR) Candidate Material: Hastelloy X, Haynes 230, and Alloy 617, Ph.
D. thesis, University of Missouri, USA, 2011.
[22] M. McGuiness, S. Taylor, Strip Temperature in a Metal Coating Line Annealing Furnace, Report series/Department of Mathematics, School of
Mathematical and Information Sciences, University of Auckland, 2004.
[23] J.B. Ordieres Mer, A. Gonzlez Marcos, J.A. Gonzlez, V. Lobato Rubio, Estimation of mechanical properties of steel strip in hot dip galvanising lines,
Ironmak. Steelmak. 31 (2004) 4350.
[24] A. Pernia-Espinoza, M. Castejon-Limas, A. Gonzalez-Marcos, V. Lobato-Rubio, Steel annealing furnace robust neural network model, Ironmak. Steelmak.
32 (2005) 418426.
[25] M.M. Prieto, F.J. Fernndez, J.L. Rendueles, Development of stepwise thermal model for annealing line heating furnace, Ironmak. Steelmak. 32 (2005)
165170.
[26] H. Ramamurthy, S. Ramadhyani, R. Viskanta, A thermal system model for a radiant-tube continuous reheating furnace, J. Mater. Eng. Perform. 4 (1995)
519531.
[27] J.M. Rhine, R.J. Tucker, Modelling of Gas-Fired Furnaces and Boilers and Other Industrial Heating Processes, British Gas, 1991.
[28] C.R. Roger, High Temperature Emissivity of Stainless Steel AISI 304, University of New Orleans, 1978.
[29] S. Rudtsch, H. Ebert, F. Hemberger, G. Barth, R. Brandt, U. Gro, W. Hohenauer, K. Jaenicke-Roessler, E. Kaschnitz, E. Pfaff, W. Pnecker, G. Pottlacher,
M. Rhode, B. Wilthan, Intercomparison of thermophysical property measurements on an austenitic stainless steel, Int. J. Thermophys. 26 (2005) 855
867.
[30] R.J. Schurko, C. Weinstein, M.K. Hanne, D.J. Pellecchia, Computer control of reheat furnaces: a comparison of strategies and applications, Int. J. Heat
Mass Transfer 64 (1987) 3742.
[31] R. Siegel, J.R. Howell, Thermal Radiation Heat Transfer, fourth ed., CRC Press, 2001.
[32] I. Sokolova, Einfluss der Transportrollen auf die Guterwrmung in Rollenfen, Ph.D. Thesis, Technische Universitt Freiberg, Germany, 2004.
[33] A. Steinboeck, D. Wild, T. Kiefer, A. Kugi, A mathematical model of a slab reheating furnace with radiative heat transfer and non-participating gaseous
media, Int. J. Heat Mass Transf. 53 (2010) 59335946.
[34] A. Steinboeck, D. Wild, A. Kugi, Nonlinear model predictive control of a continuous slab reheating furnace, Control Eng. Pract. 21 (2013) 495508.
[35] S. Strommer, M. Niederer, A. Steinboeck, A. Kugi, A mathematical model of a direct-fired continuous strip annealing furnace, Int. J. Heat Mass Transf. 69
(2014) 375389.
[36] Y.C. Tian, D.C. Levy, T. Gu, Implementing a process model for real-time applications, in: Proc. Asia-Pacific Conference on Complex Systems, Central
Queensland University, Cairns, Australia, 2004, pp. 287296.
[37] F. Wan, Y.Q. Wang, S.R. Qin, Modeling of strip heating process in vertical continuous annealing furnace, J. Iron Steel Res. Int. 19 (2012) 2936.
[38] D. Wild, T. Meurer, A. Kugi, Modelling and experimental model validation for a pusher-type reheating furnace, Math. Comput. Model. Dyn. Syst. 15
(2009) 209232.
[39] J.G. Wnning, FLOX Flameless Combustion, Thermprocess Symposium, 2003.
[40] N. Yoshitani, A. Hasegawa, Model-based control of strip temperature for the heating furnace in continuous annealing, IEEE Trans. Control Syst. Technol.
6 (1998) 146156.

You might also like