Copyright 2002 IFAC

15th Triennial World Congress, Barcelona, Spain

ESTIMATION OF PROCESS VARIABLES IN A GLASS

MELTING FURNACE

Leo Huisman

Department of Chemical Engineering, Eindhoven University of

Technology, 5600 MB Eindhoven, l.huisman@tue.nl

Abstract: In this paper the application of observers for glass melting furnaces is discussed.
In glass melting furnaces only few variables, of those that are important for the glass quality,
can be measured. Here it is demonstrated how low complexity mathematical furnace models
can be used to estimate the other variables that cannot be measured.

Keywords: Luenberger observer, Kalman filter, glass melting furnace

1. INTRODUCTION duced by a so called fining agent, that dissociates at a

higher temperature than the melting temperature of the
In glass melting furnaces glass is molten at high tem- glass to form a fining gas. For the melting, fining and
peratures and the result is a glass melt that contains a homogenizing processes not only the temperature is
high amount of very small bubbles (containing mainly important but also the velocities in the melt. In figure 1
CO2 and N2 ) and also some solid material (mainly it is shown that temperatures in the glass melt are not
SiO2 ) that has to be dissolved in the melt. Further- measured. To determine the temperatures in important
more, the glass melt may not yet be homogeneous just glass melt zones but cannot be reached through online
after melting. A sketch of a glass melting furnace is measurements we use an estimator which is designed
shown in figure 1. For the melting of the remaining based on mathematical models that are available for
solid particles it is important that the particles have the glass melting furnaces. Here the design of such
a sufficiently long residence time in the furnace and estimator for glass melting furnaces is discussed. First
sufficiently high temperatures along their trajectories. the derivation of a relatively simple simulation model
The small bubbles are released from the melt by letting is presented. This model is used in the Kalman ob-
server that will be discussed next.
TT TT pO TT TT
LT 2
pCO Burner port
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xx
xxx
Refiner
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xx
xxx xxxx
xxxx Feeder
xx
xxx
xx
xxx xxxx
xxxx
xx
xxx
xx
xxx xxxx
xxxx PT LT
pO xx
xxx
xx
xxx
xx
xxx
xxxx
xxxx
xxxx
xx
xxx
2 xxx
xxx
xx
xx
xx
xxx
xxxxx
xx
xxx
xxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxx
xxxx
xxxx
xxxx
xxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxx
xxx
xxx
2. MODELS FOR THE GLASS MELTING
xxxxx
xx
xxx xxxx xxx
xxxxx
xx
xxx Combustion space xxxx xxx
xxx
xxx
xxx
xxx
xxx
Batch xxxx
xxxx
xxxx
xxxx
xxxx
xxxx
xxx
xxx
xxx
xxx
xxx
xxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxx
FURNACE
xxx xxxx xxxxxxxxxxxxxxxxxxxxxx
xxx
xxx
xxx
xxx
Glass melt xxxx
xxxx xxxxxxxxxxxxxxxxxxxxxxxxxxx
xxx
xxx
xxx
xxx
xxx xxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxx xxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx T T T T
Throat Gob
Glass melting tanks are often modeled with CFD-
TT TT TT TT TT
codes (computational fluid dynamics (Patankar, 1980),
(Post, 1988), (Ungan and Viskanta, 1987), (Viskanta,
Fig. 1. Sketch of a glass melting furnace with its measurements.
1994)) for design purposes or process operation im-
T T is temperature transmitter, PT is pressure transmitter, LT
provement. Typical properties of these models are
is level transmitter pO is partial oxygen pressure and pCO
2 the high level of detail and high computational effort
is partial carbon monoxide pressure. The sensor with dotted
needed to do calculate a solution. If a model is used
lines is a sensor for oxygen partial vapor pressure in the melt
in a controller, it must allow fast enough calculations
and is in a test phase.
and predict the dynamic behavior in the important
them grow and rise to the surface. This process is frequency region well enough. Fast enough would typ-
called the fining process. The bubble growth is in- ically mean a hundred times real time and well enough
would mean that the controller will make the process properties that are not influenced by in or outflows.
meet any performance criteria one might have. Know- Connections are mass flows (indicated by a solid line)
ing this, one can verify that a CFD-code with a high or energy flows (indicated by dotted lines), that could
number of grid points will in general be to complex. be work or heat flows. Table 1 gives the indices of the
Much effort is being put in developing reduced models systems in figure 1. In the next section a derivation of
that are derived from a glass tank model (GTM), a the model is given.
CFD-code written especially for glass melting fur-
naces by the research institute TNO. Until these mod-
2.1 Ideally stirred regions
Cr1 F1 F2 Cr2
For the ideally stirred region under the batch blanket
s; 1 s; 2 an energy equation for a constant pressure zone leads
BB to:
T3 T2 T1
dT

c pVUB UB = c pVT 3jUB TT 3 TUB
m pull m pull dt
UB B11 B12 B13 NT

e f f AUBjBB1 TUB TBB1
B21 B22 B23 wall AUBjE

TUB TE + Se;UB
wall
Boosting This equation can briefly be written as:
E C


TUB = fUB TUB ; TT 3
Fig. 2. Physical topology of a glass tank with a single
D

Se;UB
recirculation loop. + fUB TUB ; TBB1 ; TE
(1) +
c pVUB
els become available testing of controllers can be done Chemical reactions are assumed to take place outside
on simple test models derived from first principles or region (that is in the batch blanket) and the glass is
identified by using input-output data from the GTM treated as a single component having average proper-
or a real furnace. Identification of input-output data ties that depend on the composition.
((Haber et al., 1981), (Ljung, 1987), (Wertz and De- The region near the throat can be modeled similarly
meuse, 1987), (Wertz et al., 1992)) will lead to better with:
models for a particular furnace in a certain operating dT

range but it can only be applied in that range and for c pVNT NT = c pVB13jNT TB13 TNT
dt

that furnace. Here a simple model is proposed that + c pVB23jNT TB23 TNT
may be useful for:
wall ANT jE

 Gaining insight in the dynamics of the chemical
wall
TNT TE
and physical processes involved in the melting  
of glass and their effects on the overall time e f f ;NT ACSjNT TNT ;S TNT + Se;NT
dependent behavior of glass melting furnaces.
 Having a flexible model available that can easily This can be written as:


C
be extended or simplified and used for the design TNT = f NT TNT ; TB13 ; TB23
and testing of state observer or controllers and   Se;NT
D
very fast simulations. + fNT TNT ; TNT ;S ; TE + (2)
c pVNT
The model derived here is based on a container glass
furnace for making green bottles. This furnace has in
2.2 Heat transfer in laminar flow regions
the glass melt a single recirculation loop and some
smaller loops. For the laminar flow regions in the tank we can use
Table 1 : Indices used in the model description a two dimensional model, with convective transport
in the 1 -direction and conductive transport in the 2 -
UB Under batch BB Batch blanket
direction. This results in a two dimensional form of
B Bottom FS Fuel supply
the energy equation:
NT Near throat CS Combustion space
T Top E Environment  
(c p T ) (c p T ) qwall ;k
t
1
= ke f f T v1 Se
1 + +
In figure 2 a physical topology is given of the glass 2 2
8k
melt, and the combustion space. It consists of sys-
which can be discretized (Patankar, 1980):
tems (capacities) and connections. A circle indicates
dTi j ke f f ke f f
a lumped system, that is a region with uniform prop- dt = c p 2 (Ti; j +1 Ti; j ) (T Ti; j 1)
c p 22 i; j
hv i (T
2
erties, a rounded rectangle indicates a distributed sys-
Se;i j q
tem, that is described by a partial differential equa- 1
1
i+1; j Ti 1; j ) + c p + wall k
c pVi j
;

tion, and an arc indicates a reservoir, a region with 8k

Here, a central difference interpolation scheme for VT
the diffusive terms and an upwind scheme for the with R = a dimensionless backflow ratio and a
V pull
convective terms are applied. For the top and the constant depending on the situation (one plate ore two
bottom region the following equations can be found: plates). For the friction parts we can derive:

Se;B
T B = f D (T B ; T T ; TE ) + f
C
T B ; TUB + c (3) 3V 2 L 12VT2 LT 2
B B p E f ;T 1 = T 3T 1 and E f ;T 2 = (6)
  WT HT WT HT3

Se;T
T T = fD T T ; T B ; T T ;s + f CT T T ; TNT + c
T p 12VB2 LB
(4) E f ;B = (7)
WB HB3
Figure 2 shows that for the bottom and top region 6
and 3 grid points (lumps) are chosen, and so T B 2 R6
and the buoyancy terms can be written:
and T T 2 R3 .


hvz i gz Tb T Vb =
8b
1

gV V pull

2 +R WL TNT TUB
2.3 Model equations for the recirculation flow Using these equations we can rewrite the equation for
kinetic energy and we obtain a nonlinear differential
Multiplication of the Navier-Stokes equations with the equation for R:
velocity vector and integration over the entire system

volume gives us an expression for the total kinetic R = fRB + fRF = fR (8)
energy change in the system (Bird et al., 1960): with

3 ! gV  1


v pm fRB = W L Vpull 2 + R TNT TUB
Ktot = m m m 1
2 hvi + m +

+w
pl L L 2
8m Z
fRF = 15 2W HT 3 V pull
2
R2 12 W HB3 V pull
2
(R + 1)
1 T T
 B B

E f ;i + v
g (T T ) d 1 1
T2 2 + T2 1 2 VT 2 B2 VB 2
8i |V {z } fR = 60 W15 2H2 V pull R + 60W 2 H 2 Vpull (R + 1)
T T B B
8b hvz igz (Tb T )Vb
This equation describes the acceleration of the recir-
Now if one makes the assumption that the kinetic culating glass, due to a change in the temperature dif-
energy of the recirculating flow is only influenced by ferences in the melt. The acceleration is counteracted
the size of the buoyancy forces in the melt, then the by the friction forces. Combination of equations (1),
following form of this equation can be written: (2), (3), (4) and (8) gives:

2 3 2 3
E f i + hvz i gz
C D
TUB fUB + fUB + Se;UB
Krc = Tb T Vb (5)
;
6 f + f CB + Se;B 7
8i 8b 6 T 7 6
6 B 7 6 CB
D
7
6 T 7 D 7
The friction terms can be derived from the laminar 6 NT 7 = 6 6
fNT + fNT + Se;NT 7
7
4 T Top 5 4 f D + f C + S 5
flow patterns that
R
are assumed by calculating the in- Top Top
e;Top
R
Vi ( : v) d. In this paper it is as-
tegral E f ;i = B F
fR + fR = fR
sumed that the friction losses occur in the top and These equations can briefly be written as:
bottom regions indicated by T and B in the physical
topology given in figure 2 and that the flow can be x (t ) = f (x (t ) ; u (t ) ; )
modeled as one dimensional laminar flow between If we define x = x xss then we can write a linear
horizontal plates. approximation near the steady state:
x (t ) = Ax (t ) + Bu (t )
iVi

6 i 2 H 2
1 2
v1 = where
W Hi3  
f  f 
A= B=
Two cases can be considered: x x=xss ;u=uss u x=xss ;u=uss
 One dimensional flow (in 1-direction) over one
A step response of the model that was derived here
plate: i = 3
 One dimensional flow between two plates: i = 6 was compared with simulation results of a detailed
CFD-model. Some results are shown in figure 3. Ma-
With these velocity profiles the accumulation term can jor differences occur in the steady state behavior.
be written as:
Z Z Z
2 v d + dt 2 v d + dt 2 v d
d 1 2 d 1 2 d 1 2
Krc = dt 3. LINEAR OBSERVERS
VT 1 VT 2 VB
"  
1 2 1 2 1 Suppose that in the neighborhood of a certain operat-
60 T 2 + 15 T 1 2 VT Vpull
2 B2 VBV pull
2
=
WT2 HT2
R + 60W 2 H 2 (R + 1) dR
dt ing point the behavior of the glass melting furnace can
B B
be described by a linear state space model of the form:

1
6 T
P = APA T CPC
APC T + Rv T
CPA
T
5 + Gw Qw Gw (9)

4 where Rv and Qw are the covariance matrices of the

disturbance vectors v and w respectively. The solution
T [K]

P is the steady state value of the covariance matrix P

3

2 of x:
 
lim P (t ) = lim E x (t ) x (t )T
1 t ! t !

0
The Kalman gain matrix L is then given by:

1
1
0 2 4 6 8 10 12
T CPC
L = APC T + Rv
t[s] x 10
4

The observer equation in the discrete time case is

Fig. 3. Step response results of a detailed computa- given by:
tional fluid dynamics model (CFD) and the sim-
ple simulation model (smooth curves) x (t jt + 1) = A x (t jt (t )
1) + Bu
x = Ax + Bu + w + L (y (k) Cx (t jt 1))
y = Cx + v The Luenberger configuration can be put in the stan-
where xus 2 is the state vector, w 2
Rn v2 Rn , Rp dard form in figure 5 where P can be partitioned as:
are disturbance vectors, u 2 R m is the input vector 
P11 P12

(manipulated variables), y 2 R p is the output vector P =
P21 P22
(measured variables) and = dtd in continuous time
and x (t ) = x (t + 1) in discrete time. This state space If we take
model is used in an observer to estimate the state  1=2  1 
variables. R v 0 v
w =
0 Qw1=2 w

and w as white noise with  unit intensity then the

3.1 Luenberger observer Kalman filter minimizes Fl (P; F )2 = jjP11 + P12 (I
P22 F ) 1 P21 jj2 . So the Kalman filter is a H 2 -filter.
A type of observer that is often used is a Luenberger Minimization of the 2-norm means minimizing all
observer. The structure of such observer type is shown
in figure 4. A Kalman filter ((Gelb, 1974), (Jazwinski,
P
D z e
w
x x u
P
u B 1 C y y

L F
- z
D
x x Fig. 5. Filter problem in standard form: Find a filter
B 1
C y
F such that the norm of the closed loop transfer
from w to e is minimized
A

singular values of the transfer matrix Fl (P; F ) in the

Fig. 4. Luenberger observer
entire frequency range. By filtering the white noise
1970), (Muske and Rawlings, 1993)) also has this disturbances (coloring) emphasis can be put on im-
structure and the gain matrix L is then determined portant frequency ranges. Let these filters have the
such that the variance of the estimation error x = x x following state space representation:
is minimized when v and w are taken as normally
xv = Av xv + Bvw 1
distributed white noise, that is:
  v = Cv xv + Dvw 1
min tr E x (t ) x (t )T xw = Aw xw + Bw w 2
L;x(0)
w = Cw xw + Dw w 2
The solution is then found in the discrete time case
through the solution of the following Discrete time with xv 2 Rnv and xw 2 Rnw Then the extended plant is
Algebraic Ricatti Equation (DARE): described by:
 0 1
0 1 0 1xv 0.045

xv Av 0 0 B v 0 0 B x
B wC
C 0.04

B xw C B 0 A w 0 0 B w 0C B x C
B C B CB C
0.035

@ x A = @ 0 Cw A 0 Dw B AB C
B w 1 C
0.03

y Cv 0 C Dv 0 D @ w 2 A 0.025

x [K]
u 0.02

0.015
The filter that is designed for this extended system
has dimension n v + nw + n. The filter can be extended 0.01

in the Luenberger setting by adopting the extended 0.005

model also in the observer and then search for a 0

gain matrix L 2 R( nv +nw +n) p ((Muske and Rawlings,

Temperature under batch (no extension)
Temperature under batch (output disturbance integrator)
0.005
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

1993)). First we consider the case where we only have t [s] x 10

5

integrating coloring filters for the output disturbance Fig. 6. The estimation error x = x x of a variable that
v: is not measured: the temperature in the region
0 1
0 1 0 1 xv under the batch blanket (disturbance: first a step
xv Av 0 B v 0 0 B B x C
C and then also a ramp).
@ x A = @ 0 A 0 Q1=2 B A B w C
w B 1C filter states xv that were added to the observer. If we
y Cv C Dv 0 D @ w 2 A
consider the case we only have an input disturbance
u
filter we get:
where h i 0 1
Rv1;=12 0 0 1 0 xv 1
Av = I B v =
xw Aw 0 0 Bw 0 B B x C
C
h i @ x A = @ Cw A 0 Dw B A B w

B 1C
C
Cv = I Dv = 0 Rv1;=22 y 0 C Rv1=2 0 D @ w 2 A
If we consider the evolution of the filter states, x E , we u
get: The evolution in discrete time of the observer state
     xE (t ) is then given by:
xv (t + 1) I L1 L1C xv (t )
=     
x (t + 1) L2 A L2C x (t ) xw (t + 1) Aw L1C xw (t )
    =
0 L x (t + 1) Cw A L2C x (t )
+ u (t ) + 1 (Cx (t ) + v (t ))  
B L2 0
+ u (t ) + LCx (t ) + LRv1=2w 1 (t )
Now we assume the following: B

 Steady state (t ! ), which means xE (t + 1) = Now if we make again some assumptions:

xE (t )  We only have a constant offset on the state vari-
 u (t ) = 0 ables: w 2 (t ) = w 2
 The system is assymptotically stable: limt ! x (t ) =  u (t ) = 0 and w 1 (t ) = 0
0  Steady state: xE (t + 1) = xE (t )
 The disturbance v (t ) = v remains constant (off-
set) With these assumptions we get:
  
We can then write: Aw I L1C xw (t )
0= + LCx (t )
     Cw A L2C I x (t )
L1 L1C xv (t ) L1
+ v = 0
L2 A L2C I x (t ) L2 If we choose A w = I and C = I then we would expect
the error x = x x to converge to zero. Results are
And calculate the steady state value for the extended
shown in figure 7 and 8. After applying a step dis-
state xE :
turbance the error converges to zero for this type of
    1 
xv (t ) L1 L1C L1 Kalman filter.
= v
x (t ) L2 A L2C I L2 Finally, we can add both an input and an output filter.
If we use the M ATLAB routine dare to calculate Then the states of the observer x E satisfy:
the Kalman gain matrix L and then construct x E 0 1 0 10 1
xv Av L1Cv 0 L1C xv
where the output disturbance is a constant offset v = @ xw A = @ L2C A @ xw A

T L2Cv Aw
1 1 1 1 we get for xE : x L Cv Cw A L3C x
  0 1 3 0 1
xE  xv
0
0 L1
+ @ 0 A u + @ L2 A (Cx + Cv v)
B L3
and the error x converges to zero. This is indeed
observed and shown in figure 6. The corrections for In this case offsets lead to steady state estimation
the output offset disturbance are applied to the extra errors. Results are shown in figure 9. If both input
 14
x 10
3
simulation times. The simple simulation model ap-
Temperature under batch (no extension)
Temperature under batch (input disturbance integrator) proximates the dynamics of the detailed model quite
12
well. The obtained simulation model can be linearized
10 and used in a Kalman filter to estimate state variables
8
that are not measurable but important for the glass
melting process. Effects of disturbances than white
x [K]

6
noise (e.g. offsets because of sensor aging) can be
4 dealt with in the Kalman filter by extending the as-
sumed process model with coloring filters.
2

2
5. REFERENCES
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t [s] x 10
5

Bird, R.B., W.E. Stewart and E.N. Lightfoot (1960).

Fig. 7. Response to an input disturbance (first a step Transport Phenomena. Wiley.
and then also a ramp). In this case C = I Gelb, Arthur (1974). Applied optimal estimation. The
M.I.T. press.
0.025
Haber, R., J. Hetthessy, L. Keviczky, I. Vajk, A. Feher,
0.02
N. Czeiner, Z. Csaszar and A. Turi (1981). Iden-
tification and adaptive control of a glass melting
0.015 furnace. Automatica 17(1), 175185.
Jazwinski, Andrew H. (1970). Stochastic processes
x [K]

0.01
and filtering theory. Academic press.
Ljung, L. (1987). System Identification: Theory for the
0.005
User. Prentice Hall.
0
Muske, Kenneth R. and James B. Rawlings (1993).
Temperature under batch (no extension)
Model predictive control with linear models.
Temperature under batch (input disturbance integrator)
0.005
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
AIChE Journal 39(2), 262287.
t [s] x 10
5

Patankar, S.V. (1980). Numerical Heat Transfer and

Fluid Flow. Hemisphere.
Fig. 8. Response to an input disturbance. In this case
Post, Lourens (1988). Modeling of Flow and Com-
C =[0 I 0].
bustion in a Glass Melting Furnace. PhD thesis.
and output disturbances are present as discussed above Technische Universiteit Delft.
then no improvements if compared with the ordinary Ungan, A. and R. Viskanta (1987). Three dimensional
Kalman filter are observed, up to now. In glass melting numerical modeling of circulation and heat trans-
furnaces low frequency input disturbances (like load fer in a glass melting tank: Part 1. mathematical
changes and changes in cullet properties) are most formulation. Glastechnische Berichte 60(3), 71
likely to occur. 78.
0.12
(no extension)
Viskanta, R. (1994). Review of three-dimensional
(input disturbance integrator)
(output disturbance integrator) mathematical modeling of glass melting. J. Non-
0.1 (both input and output filter)
Cryst. solids 177, 347367.
0.08 Wertz, V. and P. Demeuse (1987). Application of
clarke-gawthrop type controllers for the bot-
x [K]

0.06
tem temperature of a glass furnace. Automatica
0.04
23(2), 215220.
Wertz, V., M. Gevers and J.F. Simon (1992). Adap-
0.02
tive control of the temperature of a glass furnace.
0
IFAC Adaptive systems in control and signal pro-
cessing pp. 311316.
0.02
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t [s] x 10
5

Fig. 9. Response to input and output disturbances. In

this case C =[0 I 0].

4. CONCLUSIONS

A relatively simple simulation model was derived.

Simulation results obtained with this model were com-
pared with a detailed simulation model that takes large