Proceedings of the 18th International Conference on System Theory, SaE3.

4
Control and Computing, Sinaia, Romania, October 17-19, 2014

Temperature control of reheating furnace based on distributed

model predictive control*

X.M. Nguyen, P. Rodriguez-Ayerbe, F. Lawayeb, D. Dumur, A. Mouchette

(air/gas ratio, oxygen rate…). In addition, the distribution of

Abstract— In steel industry, reheating furnaces are the biggest temperature is not uniform inside the furnace which makes the
energy consumers after the blast furnace. The purpose of a slab
reheating furnace is to heat steel slabs to a suitable temperature
control task more difficult. The slab target temperature is
profile (around 1250°C) before entering to hot rolling phase. The achieved by adapting zone temperature set point. Then,
reheating furnace is divided in zones, each zone being controlled temperature of each zone is regulated by adjusting the power
independently. In this paper, a model considering the coupling (air/gas flow rate) injected to it. Consequently, in the common
effect between zones is proposed and identified. This model is design, the control of a reheating furnace is divided into 2 main
afterwards used in a distributed model predictive control strategy levels as shown in Figure 2. The level 2 determines the
(DMPC). The experimental results in industrial plant show an temperature set point for each zone of furnace accordingly to
energy saving of about 3% while maintaining the heating quality the scheduling of products, their desired discharged
of products. In addition to energy saving, industrial online temperature and the instantaneous thermal state of furnace. At
measurements showed more stable furnace performance,
especially in transient (oxygen level, pressure, global air/gas
this level, a heat transfer model is used to estimate slab
supply). temperature profile which is not measurable. Level 1 of the
control system calculates the required power injected to each
I. INTRODUCTION zone to reach the set point given by level 2.

In hot rolling steel plants, the first step is the reheating

furnace; it consists in heating steel slabs (0.24m x 1.5m x10m)
from ambient temperature to around 1250°C at 400tons/hour
per furnace. To achieve this task, the furnace dimensions are
quite important (60m x 12m x 6m). The energy consumption of
such furnace is equivalent to the consumption of 100000
houses (1toe†/house, 1toe=42GJ, 1.4GJ/ton, and 3Mtons of
production per year). Thus, energy saving is a crucial challenge
in furnace control.
Figure 1. Walking-beam reheating furnace
A walking-beam reheating furnace is usually divided into
several control zones: recuperation zone, preheating zone, one In current furnaces, the level 1 of the control system is
or more heating zones, and soaking zone as shown in Figure 1. based on conventional PID controllers which do not take into
Recuperation zone has no burners in order to collect heat from account furnace dynamics and coupling. The strong
exhaust fumes. Slabs are primarily heated in preheating and interactions between zones of furnace are due to the flow of
heating zones where burners are installed. The temperature exhaust fumes, and the roof/walls radiation. Therefore, PID
uniformity of slabs is achieved in soaking zone. controllers can show strong oscillations and temperature
The reheating furnace is a complex system with strong overshoot, especially in case of transient, which leads to poor
thermal coupling between zones, considerable thermal inertia control performances and energy losses.
of walls and roof, high sensitivity to combustion parameters The study explores the potential of model predictive
control (MPC) techniques to improve the temperature control
of level 1 , and particularly whether or not the energy
consumption can be reduced by exploiting these advanced
*
Research supported by ArcelorMittal Global Research & Development control techniques:
Center. MPC techniques are well-known for their capacity of
X. M. Nguyen is PhD candidate with the Control Department, Supélec handling industrial constraints, dealing with multivariable
Systems Sciences E3S, Gif-sur-Yvette, France (e-mail:
[email protected]). processes, and their ability to transform a control problem to
P. Rodriguez and D. Dumur are with the Control Department, Supélec an online optimization under constraints [2]. Distributed MPC
Systems Sciences E3S, Gif-sur-Yvette, France (e-mail: (DMPC) is an adapted version of MPC dealing with large-scale
[email protected]; [email protected]). systems which consist of strongly coupled sub-systems. In
F. Lawayeb and A. Mouchette are with the Measurement and Control
Department, ArcelorMittal Research, Maizières, France (e-mail: DMPC techniques, local MPC controllers cooperate,
[email protected]; [email protected]).
†
toe : ton of oil equivalent : is a unit of energy, the amount of energy
released by burning one tonne of crude oil

978-1-4799-4601-3/14/$31.00 ©2014 IEEE 726

Authorized licensed use limited to: ArcelorMittal Global R&D. Downloaded on August 24,2022 at 11:53:09 UTC from IEEE Xplore. Restrictions apply.
communicate their actions in order to handle the coupling inferior sub-zone is obtained by applying a variable ratio
effect of sub-systems and achieve together the optimal sup/inf. The system to be considered consists of five sub-
operation of the overall system [2]. In case of reheating systems as illustrated in Figure 3.
furnace, a dynamic model that considers interaction between In each zone of the furnace, temperature variation at
zones is used to predict the evolution of temperature in a given thermocouple location is modeled by transfer functions
prediction horizon. At each sample time, a criteria that takes considering coupling effects between zones. At level 1 of the
into account temperature tracking performance and energy control system, inputs to each zone are power signal and
saving, is minimized to obtain the optimal control sequence. temperature signal of neighbor zones, the output to be
controlled is zone temperature.
The coupling effect between zones is mainly due to the
flow of exhaust fumes in the opposite direction of slab
movement. In most cases, the coupling is limited to first or
second neighbor. The interactions of furnace zones can be
modeled as output coupling, as shown in Figure 3, or they can
also be modeled by conventional input coupling, as discussed
in [4]. Both methods were tested, but industrial feedback
showed that output coupling method gives more satisfactory
Figure 2. Existing and new structure of control system
results. Therefore, in the following, an output coupling model
is described and used to identify the furnace around a nominal
operation point. The transfer functions of furnace zones are
In [5] and [1], a controller based on MPC is applied to formed as in (1).
reheating furnace (MPC level 1 in [5] and level 2 in [1]). The
furnace is modeled as a multiple-input, multiple-output y1 = H 11u1 + H 12 y 2
(MIMO) system based on energy and mass balances.
Simulation results give significant improvement in terms of y 2 = Η 21 y1 + Η 22 u 22
energy consumption. In [4], authors presented a DMPC
algorithm for systems coupled with inputs. The algorithm is
y 3 = Η 31 y1 + Η 32 y 2 + Η 33 u 3
. (1)
applied to a study case of level 1 in a reheating furnace with y 4 = Η 43 y 3 + Η 44 u 4
three controllable zones and gives significant results in
tracking performances. In [3], an iterative DMPC algorithm y 5 = Η 54 y 4 + Η 55 u 5
developed for systems coupled with output is presented. A
study case for level 1 of reheating furnace is brought out and Where yi = Ti – Tinom and ui = Pi – Pinom, with Ti and Pi, i=1÷5
shows interesting improvements in energy efficiency and are variations of zone temperatures and zone powers,
tracking performance. This paper shows analysis and industrial respectively. Tinom and Pinom are nominal operating temperature
results obtained with DMPC algorithm developed in [3]. To and power of zone i. Observations on industrial furnaces
ensure the self-contained character of the paper, the algorithm showed that the transfer functions Hij are alike to first order
presented in [3] will be described. functions with delay, and static gain, see Figure 4.
In this paper, the problem of minimizing energy Gij
−τ ij s
consumption for the reheating furnace as well as guaranteeing H ij = e . (2)
slab temperature target is tackled by employing DMPC method 1 + λij s
for level 1. The modeling and system identification of the
The delays are approximated by
furnace zones is presented in section II. Distributed algorithm
is presented in section III. Industrial results are discussed in 1
section IV. In section V, steel plant feedback after industrial e −τs = . (3)
1 + τs
implementation is discussed. Conclusions are drawn in section
VI.
Soaking 1 Heating 2 Heating 1 Preheating
u1
I. MATHEMATICAL MODELING OF FURNACE ZONES y2
y1 y1
y3 y4
y2 y3 y4 y5 1
y5
A. Mathematical modeling of furnace zones y1 u4 u5
y2 u3
u2
The considered furnace in this study has five controllable Soaking 2
zones: preheating, heating 1, heating 2, soaking 1, and soaking
5 1 2 3 4
2. The heating zone is quite long, and it is divided into two u1 u2 u3 u4 u5
zones in length in order to have better temperature control. The
soaking zone is divided into two control zones in width to Figure 3. Couplings of furnace zones.
achieve temperature uniformity of slabs through length.
Furthermore, each zone includes superior and inferior sub-
zones. The power demand of superior sub-zone is calculated
by controllers level 1 and the corresponding power demand of

727

Authorized licensed use limited to: ArcelorMittal Global R&D. Downloaded on August 24,2022 at 11:53:09 UTC from IEEE Xplore. Restrictions apply.
B. System identification II. DISTRIBUTED MODEL PREDICTIVE CONTROL
To identify the transfer functions of (1), online trials need APPLIED FOR REHEATING FURNACE ZONES
to be performed on the furnace with normal production
conditions. The trials consist in a step signal of zone power A. System modeling and predictions
that is applied to one zone while maintaining constant power The furnace zones transfer functions (1) can be converted
on other zones as well as steady production rate, see Figure 4. to state-space representation as follows.
Afterwards, the temperature variation of the considered zone
will give the direct transfer function Hii of (1), with i=1÷5. The  x i (k + 1) = Ai x i (k ) + Bi u i (k ) + ∑ Bij y j (k )

coupling with other zones is obtained from the variation of the  j ≠i . (4)
other zone temperature. The tools used for identification are  y i (k ) = C i x i (k )
Matlab identification toolbox and method of moments. The
coupling with the neighbor zone temperature is identified with Where {Ci, Ai} is observable for i, j=1÷5. From (4), the
least squares method. predicted zone temperature at time instant k+s of zone
temperature based on output at time instant k can be calculated
C. Models validation as below [4].
Data of several days of furnace operation are used for yˆi (k + s / k ) = Ci Ais xi (k )
offline validation of zone models. Figure 5 shows the dynamics min( s , N u )
of furnaces zone for an entire day, and the identified model fits
correctly the experimental data.
+ ∑C A
p =1
i i
s− p
Biui (k + p − 1 / k ) . (5)
An online tuning is needed when the DMPC operates in s

online mode to adapt the model to the condition of the furnace. + ∑∑ Ci Ais − p Bij yˆ j (k + p − 1 /(k − 1))
j ≠i p =1
The static gains in transfer functions of (1) are usually tuned
accordingly to the nominal operating point of the furnace. s=1÷Np, where Np and Nu are respectively the prediction
horizon and the control horizon. ui (k + p − 1 / k ) is the zone
power at time (k+p-1) calculated at time k. The objective of
control is to minimize tracking errors of zone temperatures and
energy consumption of furnace. Therefore, a local performance
index for zone i can be formulated as follows.

[ ]
Np

Ji (k) = ∑qi,s yˆi (k + s/ k) − yiref(k + s)

2
min
ui (k k),..,ui (k+Nu−1 k)
s=1
Nu−1 . (6)
+ ∑ri,p+1ui (k + p)
2

p=0

Conditions on limited power, limited power variation, and

bounded zone temperatures are expressed as:

u imin ≤ u i (k + p) ≤ u imax , p = 0,..., N u − 1

Figure 4. Industrial trials for identification

∆u imin ≤ ∆u i (k + p ) ≤ ∆u imax , p = 0,..., N u − 1 . (7)
y i
min
≤ yˆ i (k + s k ) ≤ y i
max
, s = 1,..., N p
Preheating
1100 where {qi,s, ri,p+1} are positive weighting factors, yiref(k+s) is
(°C)

1000

900
the set point temperature of zone i at sample time (k+s),
0 5 10 15 20 ∆u i (k + p) = u i (k + p + 1) − u i (k + p) is the power increment
Heating 1 of zone i at instant (k+p). In a vector form, (5) and (6) can be
1400
reformulated as follows.
(°C)

1300

1200
Yˆi (k ) = Ψi xi (k ) + ΦiiU i (k ) + ∑ ΦijYˆj (k − 1) . (8)
0 5 10 15 20 j ≠i
Heating 2
1400
[
Yˆi ( k ) = yˆ iT ( k + 1 / k ) K yˆ iT ( k + N p / k ) ]T
(°C)

( k ) = [u k )]
1300
T
1200
Where Ui T
i (k / k ) K u iT ( k + N u − 1 ,
[ ]
0 5 10 15 20
T
Ψ i = (C i Ai ) T
N
time (h) K (C i A i ) T p

Figure 5. Validation of zone models: red curve is model data, black curve is
measured data.

728

Authorized licensed use limited to: ArcelorMittal Global R&D. Downloaded on August 24,2022 at 11:53:09 UTC from IEEE Xplore. Restrictions apply.
  CiBij 0 K 0 C. Design of observer
 Ci Bi K 0 
 M O M   CAB CiBij O M  Because the only available parameters are powers injected
  Φ = i i ij to furnace zones, and zone temperature, an observer is
Φii =Ci AiN −1Bi
u
K , ij  M
Ci Bi M O 0
.
 M  necessary to estimates states variables of zone model which is
M M  N −1 N −2  needed to predict zone temperature. Pole placement technique
 N −1 N −N  Ci Ai Bij Ci Ai Bij
p p
K CiBij
Ci Ai Bi
P
K Ci Ai Bi 
p u

is applied to the design observer gains Li.

2 2 xi (k +1) = Ai xi (k ) + Bi ui (k ) + Li [yi (k ) − yi (k ) ]+ ∑ Bij y j (k )
min Ji (k) = Yˆi (k) −Yiref(k) + Ui (k) R  j ≠i .
Ui (k)
And Qi i
. (9)  yi (k ) = Ci xi (k )
aiUi (k) ≤bi
(11)
with ai, bi are detailed in appendix A
Qi = blkdiag (q i ,1 ,..., q i , N p ) and Ri = blkdiag ( ri ,1 ,..., ri , N ) are ui(k)
u
yi(k)
positive definite diagonal matrices. The local optimization Controller i Zone i
problem is solved by a quadratic programming (QP).
xˆ i ( k + 1)
B. Iterative algorithm
Observer i yj(k)
In iterative algorithm, local MPC controllers solve their
local optimization problem and communicate the results with
Figure 6. Observer for zone i.
each other in order to obtain the optimal solution for overall
system. The iterative algorithm for outputs coupling systems in Error of estimation is calculated as
[3] applied for reheating furnace can be described as follows,
~
xi ( k + 1) = xi (k + 1) − xi (k + 1)
assuming that communication and computation power is
sufficient. = Ai xi (k ) + Bi ui (k ) + Li [ yi (k ) − yi (k ) ] +
• Step 1. Initialization and communication. The local + ∑ Bij y j (k ) − Ai xi (k ) − Biui (k ) − ∑ Bij y j (k ) .
MPC controller of zone i initializes its control j ≠i j ≠i

sequence of power signal U i(l = 0 ) (k ) = U iopt (k − 1) = Ai [ xi (k ) − xi (k )] + LiCi [ xi ( k ) − xi ( k )]

which is the one found at the previous sample time, l = [ A − L C ]~
i x (k )
i i i

is iteration index, sends out the predicted temperature (12)

vector Yˆi (k − 1) with the other controllers and The estimated state xi (k ) will converge to the real state
receives those from other controllers xi (k ) with a proper selection of gain matrix Li such as
Yˆ (l =0 ) ˆ
(k ) = Y (k − 1) , for j = 1,5 . eigenvalues of (Ai -LiCi) are all inside the unit circle.
j j

• Step 2. Optimization. Each MPC controller solves its

local optimal problem expressed in (9) and obtain the
control sequence of iteration l = l + 1 : U i( l =l +1) (k ) . III. INDUSTRIAL RESULTS
• Step 3. Checking and updating. The predicted In the current situation, ArcelorMittal Florange steel plant
temperature vectors are updated with the computed has two reheating furnaces in operation. In the furnace
control sequence of Step 2 by the formula described in section II (furnace A), DMPC has been
Yˆi (l +1) (k ) = Ψi xi (k ) + ΦiiU i(l+1) (k ) + ∑ ΦijYˆj(l ) (k ) . If industrially implemented and works in full-time. Another
j ≠i furnace (furnace B) working in parallel with the furnace A, is
U i( l +1) (k ) − U i( l ) (k ) ≤ ε i for i = 1.5 , where ε i is a still having PID controllers at the level 1 of its control system.
Two furnaces have identical control system technologies and
convergence precision, or the iteration index l=lmax reheat similar products in the same industrial constraints;
then MPC controllers end the iterations, otherwise therefore their thermal behaviors are quite similar. The
l = l + 1 , MPC controllers exchange the predicted behaviors of PID and DMPC controllers are compared using
temperature vectors of the corresponding zone and go measurements from these two furnaces.
to the Step 2. Parameters of DMPC controllers: prediction horizon
• Step 4. Assignment of optimized control input. Once Np=25, control horizon Nu=25, sampling time Ts=5 seconds. In
the iteration is terminated, from the optimal power Figure 8, the zone temperature behaviors of furnace A and
input sequence U i opt (k ) , local MPC controller of furnace B are compared. It is noticed that with DMPC
controller, zone temperature has less delay and overshooting in
zone i applies ui opt (k ) to the corresponding zone and
comparison to that given by PID controllers. The fact that
go to Step 1 and starts a new optimization for sample DMPC does not overshoot the zone set points leads directly to
time k+1 energy savings. In Figure 7, we can see that DMPC controllers
tend to have less oscillations of power demand than that of PID

729

Authorized licensed use limited to: ArcelorMittal Global R&D. Downloaded on August 24,2022 at 11:53:09 UTC from IEEE Xplore. Restrictions apply.
controllers, this effect reduces energy loss to the exhaust fume IV. FEEDBACK FROM INDUSTRIAL PLANT
and contributes to overall energy saving.
The described new control solution has been implemented
Furnace A is configured to work either with DMPC
in industrial conditions and obtained positive feedback from
controllers or PID controllers; Figure 9 shows that the average
furnace team in Florange steel plant. In addition to obtained
energy consumption of the furnace is reduced by using DMPC
energy savings, it was noticed that furnace operation was more
controllers. The online energy saving of 3% is observed by
stable, especially in transient mode (with high productivity
comparing the energy consumption of one month period with
variation). The global gas demand is less disturbed that leads
PID to that with DMPC.
to more stable gas pressure at supply station. Moreover, the
level 2 of control system, which is in closed loop with level 1
(gas pressure and zone temperature measurements are used to
determine the temperature set points), gives more stable set
points and the rate of automatic mode usage of level 2 is
increased compared with manual mode, see Figure 11.

Figure 7. Power demand of controllers.

Figure 9. Daily energy consumption of furnace A by ton of heated steel with

PID and DMPC compared to nominal value.

Figure 8. Zone temperature with PID and DMPC controller.

In histogram of slab discharged temperature errors shown

in Figure 10, it can be noticed that DMPC gives almost the
Figure 10. Histogram of errors of product discharged temperature.
same temperature accuracies as that obtained with PID
controllers. Therefore, the energy saving is acquired by The overall furnace stability with MPC reduces the oxygen
reducing losses on the exhaust fumes and not on absorbed ratio inside the furnace, and that leads to less scale formation
energy by the slabs. It must be noted that the PID parameters on the slab surface.
are those used in the industrial plant. These parameters can be
tuned in order to reduce the oscillations showed in Figure 7,
but this adjustment would produce a performance modification
in the slab temperature tracking showed in Figure 10. The
coupling effect of the plant considered in the model and the
prediction of the future behavior, are effectively exploited by
the DMPC algorithm in order to obtain better results than those
obtained with decoupled and non-predictive controllers.

730

Authorized licensed use limited to: ArcelorMittal Global R&D. Downloaded on August 24,2022 at 11:53:09 UTC from IEEE Xplore. Restrictions apply.
  U imax 
 
 −U i min


∑
 Yi max − Ψi x i (k ) − Φ ij Yˆ j (k − 1) 

bi (k ) =  min j ≠i
,
− Yi + Ψi x i (k ) + Φ ij Yˆ j (k − 1)
∑
 j ≠i

 max 
 M i 
 − M imin 
 I N .n 
− I 
u ui

 N .n  u
u imax 
ui
u imin   y imax 
Figure 11. Automatic mode usage of level 2  Φ ii       
ai =  , U max =  M , U imin =  M  . Yi max =  M  ,
− Φ ii  i
  u imax  u imin   y imax 
V. CONCLUSION C       
This paper shows the industrial results of a distributed  −C 
 
model predictive control (DMPC) applied to the temperature
control of a reheating furnace. Conventional decentralized PID  I nui 0 L L 0 
controllers are replaced by DMPC controllers. The plant is − I M 
modelled considering coupling effects between zones and the y   nuii
min
I nui O
 
model is used to minimization of a quadratic criteria in each Yi min =  M  ,C =  0 O O O M 
 
zone. The proposed distributed algorithm is solved online,  y imin   M O O O 0 
considers plant constraints and arrives to an equilibrium point  
 0 L 0 − I nui I nui 
in few iterations [3]. Industrial results show energy saving with  N u . nui × N u . nui

the same tracking performance than the classical control. ∆uimax + ui (k − 1) ∆ui + ui (k −1)
min

 ∆uimax   ∆uimin 
, M imax = , Mi = 
min
.
APPENDIX  M   M 
 ∆uimax   ∆uimin 
A. Quadratic programming formulation
From (4) and (5) the optimization problem is written as
J i (k ) = U iT (k )(Φ Tii Qi Φ ii + Ri )U i (k )
T
ACKNOWLEDGMENT
 
+ 2 Ψi x i (k ) + ∑ Φ ij Yˆ j (k − 1) − Yiref (k )  Qi Φ iiU i (k ) . X.M. Nguyen would like to acknowledge the financial
 j≠i  support of the Group ArcelorMittal.
+ const
(A1)
The QP is given as REFERENCES
[1] Balbis, L., Balderud, J., Grimble, M.J. (2008). Nonlinear predictive
0.5U i (k ) H iU i (k ) + f i U i (k ) .
T T
min
U (k )
(A2) control of steel slab reheating furnace. American Control Conference.
i
1679-1683.
[2] Scattolini, R. (2009). Architectures for distributed and hierarchical
With model predictive control – A review. Journal of Process Control,
volume 19, 723-731.
 
∑ Φ Yˆ (k − 1) − Y
[3] Nguyen, X.M., Rodriguez-Ayerbe P., Dumur D., Mouchette A.,
f i = Φ Tii Qi  Ψi x i (k ) + ij j iref
(k )  . Lawayeb F. (2012). Distributed model predictive control of steel slab
 j≠i  reheating furnace. Proceedings of the IFAC Workshop on Automation in
the Mining, Mineral and Metal Industries, 175-180.
And [4] Zhang, Y. and Li, Sh. (2007). Networked model predictive control based
on neighborhood optimization for serially connected large-scale
H i = (Φ Qi Φ ii + R i ).
T
ii processes. Journal of Process Control, volume 17, 37-50.
[5] Zhang, B., Chen, Z., Xu, L., Wang, J., Zhang, J., Shao, H. (2002). The
Inequalities in (7) can be rewritten as modeling and control of a reheating furnace, Proceedings of the
American Control Conference, 3823-3828.
a i U i ( k ) ≤ bi ( k ) (A3)
Where

731

Authorized licensed use limited to: ArcelorMittal Global R&D. Downloaded on August 24,2022 at 11:53:09 UTC from IEEE Xplore. Restrictions apply.