Professional Documents
Culture Documents
Abstract: A Lyapunov-based MIMO state feedback controller is developed for slab temper-
atures in a continuous, fuel-fired reheating furnace. Following an early lumping approach, the
computationally simple tracking controller is designed for a nonlinear, switched dynamic model
that captures both conductive and radiative heat transfer. The controller modifies reference
trajectories of furnace temperatures and is part of a cascade control scheme. Given some
nonrestrictive conditions, exponential stability is ensured, even under non-steady state operating
conditions. The capabilities of the controller are demonstrated by means of an example problem.
+ Top T + Top Top Chen et al. (2008); Ezure et al. (1997); Hollander and
Tz,1 ···
zone 1 z,2 zone 2 zone Nz Zuurbier (1982); Steinboeck et al. (2011b) proposed open-
Refractory loop temperature control schemes for continuous slab re-
wall Wj +
Slab jend Slab j Tz,N z heating furnaces. In most cases, the controller defines ref-
y Dj qj+ Slab jstart erence trajectories of furnace temperatures, which are then
z (feedback) controlled by some subordinate controllers.
Pusher − − qj− − However, these open-loop approaches may be inadequate
Tz,1 Tz,2 Skids Tz,N z if the plant exhibits significant uncertainties.
System
boundary Bottom Bottom Bottom Dahm and Klima (2002); Doss et al. (1992); Ezure et al.
zone 1 zone 2 ··· zone Nz (1997); Facco et al. (1990); Honner et al. (2001); Knoop
and Moreno Pérez (1994); Leden (1986); Marino et al.
Fig. 1. Continuous slab reheating furnace (not to scale, (2004); Pedersen and Wittenmark (1998); Shenvar (1994);
symbols defined in Sec. 2). Staalman (2004); Vode et al. (2008); Wang et al. (2004);
Westdorp (1988); Yoshitani et al. (1994) reported on feed-
back control for slab temperatures, which is usually real-
1.1 Control Task ized as a cascade control scheme. Especially for non-steady
state operating conditions, trajectory tracking control may
Stimulated by economic reasons and increasing demands be required. Doss et al. (1992); Leden (1986); Rixin and
in terms of product quality and diversity, furnace control is Baolin (1992); Staalman (2004); Veslocki et al. (1986);
Vode et al. (2008) described two-degrees-of-freedom con- and times when the row of slabs is pushed forward are
trol of reheating furnaces. In most cases, the feedback law summarized in the series (tsl ) with l ∈ N. Whenever slabs
is based on control errors in terms of slab temperatures. ¯ and Ns are updated.
enter or leave the furnace, J, J,
Pedersen and Wittenmark (1998) developed a stabilizing
nonlinear feedback control law for a single furnace zone. In the coordinate system shown in Fig. 1, the slab j has
the geometric dimensions Dj and Wj along the directions
The control task may be complicated by the fact that the y and z, respectively. The local coordinate y is 0 at the
slab temperature profile cannot be measured. Therefore, mid-plane of the respective slab. The furnace volume is
estimation algorithms are required for monitoring the slab separated into Nz = 5 zones below the slabs and Nz = 5
temperatures (cf. Fitzgerald and Sheridan (1972); Leden zones above the slabs.
(1986); Wick and Köster (1999); Wild et al. (2007); Wild
(2010)). Temperature feedback may alternatively be ob- 2.1 Heat Conduction Inside the Slabs
tained from measurements of the slab surface tempera-
tures (Doss et al. (1992); Honner et al. (2001); Marino The temperature profile Tj (y, t) inside the slab j is de-
et al. (2004); Roth et al. (1986); Shenvar (1994); Staalman scribed by the 1-dimensional heat conduction problem
(2004); Veslocki et al. (1986); Wang et al. (2004)), e. g., by ∂Tj ∂ 2 Tj
means of radiation pyrometry, but the achievable accuracy ρj c j = λj y ∈ (−Dj /2, Dj /2), t > tj,0
is uncertain. ∂t ∂y 2
with boundary conditions qj∓ (t) = ∓λj ∂Tj /∂y|y=∓Dj /2
1.3 Motivation and Method and the initial condition Tj (y, tj,0 ) = Tj,0 (y). The mass
density ρj , the specific heat capacity cj , and the thermal
Many published furnace temperature controllers are de-
conductivity λj are considered as constant. Here, qj− (t)
signed for steady-state operation or are based on relatively
simple mathematical models, e. g., models that neglect and qj+ (t) are the heat fluxes into the bottom and the
the dynamic interaction between furnace zones. Moreover, top surface of the slab. Any quantities associated with the
most publications do not feature a systematic proof of sta- bottom and the top half of the furnace are labeled by the
bility. This work aims at filling this gap with a Lyapunov- superscripts − and + , respectively.
based state feedback controller. Steinboeck et al. (2009, 2010) solved the heat conduc-
The considered mathematical model, which is briefly sum- tion problem using the Galerkin method with the trial
marized in Sec. 2, accounts for the transient, nonlinear, functions hj,1 (y) = 1, hj,2 (y) = 2y/Dj , and hj,3 (y) =
switched MIMO behavior of the system. A cascade control (2y/Dj )2 − 1/3. The temperature profile inside the slabs is
system with two-degrees-of-freedom control is outlined in approximated as Tj (y, t) = [hj,1 (y), hj,2 (y), hj,3 (y)]xj (t),
Sec. 3, and Sec. 4 presents the straightforward design of a where the vector xj (t) = [xj,1 (t), xj,2 (t), xj,3 (t)]T accom-
feedback controller for slab temperatures. The developed modates the Galerkin coefficients. The state vector
control law can cope with time-dependent reference tra- "
1 0 0
#
jectories and captures the thermal interaction within the zj (t) = 1 −1 2/3 xj (t),
furnace interior. The positive effect of the feedback con- 1 1 2/3
troller is demonstrated by means of an example problem
in Sec. 5. which contains the mean temperature T̄j (t), the bottom
surface temperature Tj (−Dj/2, t), and the top surface tem-
2. MATHEMATICAL MODEL perature Tj (Dj/2, t), respectively, represents the tempera-
ture state of the slab j. Choosing the slab surface tem-
Wild et al. (2009); Wild (2010) developed a comprehensive peratures as state variables is convenient because they
analytical model of the considered slab reheating furnace. are needed for analyzing the radiative heat exchange. The
It is based on the energy balance taking into account corresponding initial value problem is obtained as
thermal radiation, bulk flow of gases, heat input from żj (t) = aj zj (t) + b− − + +
j qj (t) + bj qj (t) t > tj,0
the burners, heat losses through furnace walls, the skids,
and the funnel, as well as the slabs, which also represent with
0 0 0 1
" # " #
heat sinks. In the real furnace, the model is implemented 12λj ∓ 1
in a state observer for estimating slab temperatures. In aj = − −5 3 2 , bj = 6±3
ρj cj Dj2 −5 2 3 ρj c j D j 6 ∓ 3
this work, it will be utilized as a simulation environment
for testing the suggested controller. Since the model of (cf. Steinboeck et al. (2010)) and the initial value zj (tj,0 ) =
Wild et al. (2009) is mathematically too complex as to zj,0 = [T̄j (tj,0 ), Tj (−Dj/2, tj,0 ), Tj (Dj/2, tj,0 )]T . Steinboeck
be used for controller design, a simplified mathematical et al. (2009, 2010) proposed a similar but nonlinear model
model published by Steinboeck et al. (2010) will be used that captures the influence of temperature-dependent pa-
in this paper. The model captures thermal radiation in the rameters cj and λj .
furnace and heat conduction within the slabs.
Summarizing the states and the heat inputs of all slabs
Consider that each slab is uniquely identified by an index j ∈ J in the vectors z(t) = [zjTstart (t), . . . , zjTend (t)]T and
j ∈ N and that all slabs j ∈ J = {jstart , . . . , jend } are
currently reheated in the furnace. For a concise notation, q ∓ (t) = [qj∓start (t), . . . , qj∓end (t)]T , respectively, yields the
Ns = |J| and the range J¯ = jstart , . . . , jend will be used. sparse system
Thus, j = J¯ is tantamount to j = jstart , . . . , jend . The ż(t) = Az(t) + B − q − (t) + B + q + (t) (1)
slab j is inside the furnace during the period [tj,0 , tj,1 ], for the whole furnace (cf. Steinboeck et al. (2010)) with
11745
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011
B ∓ = δi,j b∓
A = δi,j aj i=J¯,j=J¯ , j i=J¯,j=J¯ , Plant
and the Kronecker delta δi,j . The components and sizes of controller T̃z∓ Air, Tj
z(t), A, B ∓ , and q ∓ (t) may change at the times (tsl ). Zone fuel
Feed- T̄z∓ Furnace Tz∓
Feedback controller
2.2 Radiation Heat Transfer in the Furnace forward z̃ e
controller
controller − +
The dynamic subsystems describing the slabs are coupled z State
by thermal radiation inside the furnace. Steinboeck et al. observer
(2010) used the assumption of gray-body radiation in a
non-participating gaseous medium and the net radiation Fig. 2. Cascade control of a slab reheating furnace.
method (cf. Baehr and Stephan (2006)) for analyzing the
radiation energy balance. The furnace zone temperatures state z̃j (tj,1 ) upon being discharged from the furnace.
∓
Tz,i (t) (i ∈ {1, . . . , Nz }), which represent a combination Moreover, there are several physical and safety limits on
of local flue gas temperatures and surface temperatures both the slab and the furnace zone temperatures. In view
of the furnace walls, are considered to be homogeneously of the complexity of the control task, a cascade control
distributed within the furnace zone i. They are summa- scheme as outlined in Fig. 2 seems apposite.
∓ ∓
rized in the vector Tz∓ (t) = [Tz,i (t), . . . , Tz,N (t)]T , which
z In accordance with the requirements of other process steps
serves as model input. Steinboeck et al. (2010) derived the like the rolling mill, the superordinate plant controller
simple, static radiative heat exchange model defines the order, the movement, and the desired final tem-
4
q ∓ (t) = Pz∓ (t) Tz∓ (t) + Ps∓ (t)M ∓ (z(t))4 (2) peratures T̃j,end of slabs. The outer control loop generates
with the matrix M ∓ = [δ3i,j+(1±1)/2 ]i=1...Ns ,j=1...3Ns . The reference signals T̃z∓ (t) for the zone temperatures, which
model furnishes the net heat fluxes q ∓ (t) separately for the are feedback controlled by individual zone controllers (in-
bottom and the top half of the furnace. The 4th powers ner loop). In the inner loop, PI controllers regulate the
in (2) are a consequence of the Stefan-Boltzmann law (cf. combustion air and fuel supply to the burners. Since the
Baehr and Stephan (2006)) and have to be applied to each furnace zone temperatures Tz∓ (t) exhibit a fast response
component of the respective vector. characteristic and are measured by thermocouples, they
are a good choice for interfacing between the two con-
The matrices Pz∓ (t) and Ps∓ (t) depend on the geometry trol loops. As usual for cascade control, the assumption
and the radiative properties of the slab surfaces and the Tz∓ (t) = T̃z∓ (t), i. e., an ideal inner loop, is made when
furnace walls. Their computation is particularly simple if designing hierarchically higher controllers.
a 2-dimensional configuration (yz-plane) is assumed (cf.
Steinboeck et al. (2010) for a justification of this assump- The outer control loop is responsible for the slab tem-
tion). Because the slabs change their position, Pz∓ (t) and peratures and features a two-degrees-of-freedom control
Ps∓ (t) are piecewise constant with changes occurring only structure. Its control task may be demanding because of
at the times (tsl ). In the sequel, the argument t is omitted the non-linear, switched dynamics of the system and the
whenever confusion is improbable. fact that sometimes more than 30 slabs are concurrently
reheated, while the number of control inputs is just 2Nz =
Consider that Sj is the base area of slab j. As a con- 10. The feedforward block defines reference trajectories
sequence of the reciprocity relations of radiation heat z̃j (t) (j = 1, . . . , N s with Ns N s < ∞) and T̄z∓ (t).
transfer (cf. Baehr and Stephan (2006)), diag{[Sj ]j=J¯}Ps∓ The block can be realized in form of an iterative planning
is symmetric. By analyzing the monotonicity properties and optimization algorithm or dynamic optimization as
of (2), Steinboeck et al. (2010) showed why the proposed proposed by Steinboeck et al. (2011a,b). Neither the feed-
radiation model is in line with the second law of thermody- forward controller nor the inner control loop will be further
∓
namics. They also demonstrated that Ps∓ = [Ps,i,j ]i=J¯,j=J¯ discussed in this work. It is therefore assumed that useful
is a Hurwitz matrix – a fact that will be utilized for reference signals z̃j (t) and T̄z∓ (t), which are solutions of
controller design. (1) and (2) with the initial values zj,0 (j = 1, . . . , N s ), are
available.
2.3 Assembled Model
The feedback controller aims at minimizing the deviation
Eqs. (1) and (2) constitute a nonlinear state-space model. e(t) = z(t) − z̃(t). The corresponding control law will be
Steinboeck et al. (2010) reported that this model is expo- derived in the following section. Since slab temperatures
nentially stable if Ps− = Ps+ and a rather weak condition can normally not be measured in the furnace, they need to
on the relation between ρj , cj , λj , and Dj are satisfied. be estimated, e. g., by a state observer. In the considered
Therefore, the furnace system can be open-loop controlled furnace control system, an extended Kalman filter devel-
(cf. Steinboeck et al. (2011a,b)). However, it is expected oped by Wild et al. (2007) is used for this purpose.
that the control performance (mainly the reheating quality
of the slabs) can be improved by feedback control, as
4. FEEDBACK CONTROLLER
proposed in the following.
3. CASCADE TEMPERATURE CONTROL An exponentially stabilizing control law for the block
feedback controller in Fig. 2 is developed. Its control
The primary objective of controlling a slab reheating fur- performance in terms of the norm ||e(t)||2 will turn out
nace is that the slabs reach their desired final temperature to be superior to that of pure open-loop control.
11746
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011
4.1 Furnace System with Immobile Slabs so that V̄ (e) is negative definite (cf. Steinboeck et al.
(2010)). Since p̄ is negative semidefinite, P A + AT P is
Consider a situation where the slabs do not move, e. g., negative semidefinite if
during the interval (tsl , tsl+1 ), which implies that Pz∓ and [Pi,j (di + dj )]i=J¯,j=J¯ (5)
Ps∓ are constant. Moreover, the following proposition is positive definite. Pre- and postmultiplying (5) by
requires Ps− = Ps+ . This condition is approximately (Ps∓ )T diag{[ρj cj Dj ]j=J¯}−1 and diag{[ρj cj Dj ]j=J¯}−1 Ps∓
satisfied by the considered system because the furnace
geometry is almost symmetric with respect to the mid do not change its definiteness and yield the symmetric
∓
plane of the slabs. matrix − Si Ps,i,j (di + dj ) i=J,j=
¯ J¯
. It is positive definite if
Pjend ∓ ∓
Proposition 1. Given that the slabs are immobile and
Pjend j=jstart ,j6=i Ps,i,j (di +dj ) < −2Ps,i,i di ∀ i ∈ J. The latter
∓
that Ps− = Ps+ as well as j=j Ps,i,j (di + dj ) < condition follows from the Gershgorin circle theorem.
start ,j6=i
∓
−2Ps,i,i di ∀ i ∈ J with dj = λj /(ρj cj Dj2 ) ∀ j ∈ J, the Given the above condition holds, V̄ (e) is negative definite
control law because of the following reasons. For V̄ (e) = 0, the last
T̃z∓ (t) = diag{T̄z∓ (t)} 1 + G∓ e(t)
1/4
, (3a) matrix in (4), which is diagonal, forces certain components
of e to be 0. All remaining components of e are 0 as well
evaluated separately for the bottom and the top half of because (5) is positive definite.
the furnace, with the vector 1 of unity components only,
Hence, according to Lyapunov’s direct method, the system
G∓ is exponentially stable (cf. Vidyasagar (1992)) with respect
(3b)
= diag{g ∓ }(Pz∓ )T (Ps∓ )−T diag{[ρj cj Dj Sj ]j=J¯}M ∓ , to the equilibrium e(t) = 0 ∀ t. Given some initial error
e(t0 ), the control error e(t) will decrease according to
and some user-defined gain vector g ∓ ∈ (R+ )Nz exponen- ke(t)k22 ≤ ke(t0 )k22 exp(−(t − t0 )k3 /k 2 )k2 /k1 ∀ t ≥ t0 with
tially stabilizes the system (1) and (2) with respect to the
reference trajectory z̃(t). k1 = µmin (P ), k2 = µmax (P ),
k3 = µmin − P A − AT P
Proof. Consider the error dynamics
+ 60(G− )T diag{g − }−1 diag{T̄z− }4 G−
ė = ż − z̃˙
+ 60(G+ )T diag{g + }−1 diag{T̄z+ }4 G+
= A + B − Pz− diag{T̄z− }4 G− + B + Pz+ diag{T̄z+ }4 G+ e
+ ζ 3 240[δi,j ρj cj Dj Sj diag {0 1 1}]i=J¯,j=J¯ ,
+ B − Ps− M − + B + Ps+ M + (e + z̃)4 − z̃ 4 ,
where µmin (·) and µmax (·) are the minimum and maximum
the very reasonable assumption z > ζ1 with some small eigenvalue of the respective matrix.
ζ > 0, and the Lyapunov function candidate V (e) = eT P e
with the positive definite matrices Note that the period
36 −3 −3 k k2
" #
T ∆tmin = 2 ln (6a)
P = P = [Pi,j p]i=J¯,j=J¯, p = −3 4 −1 , k3 k1
−3 −1 4 is sufficient to ensure
[Pi,j ]i=J,j=
¯ J¯ = ke(t)k22 ≤ ke(t0 )k22 ∀ t ≥ t0 + ∆tmin . (6b)
− diag{[ρj cj Dj Sj ]j=J¯}(Ps∓ )−1 diag{[ρj cj Dj ]j=J¯}. Clearly, if g − = g + = 0, i. e., in case of open-loop control,
Thus, k3 is smaller, implying that the control error e(t) decreases
slower.
V̇ (e) = eT P A + AT P
+2P B − Pz− diag{T̄z− }4 G− + B + Pz+ diag{T̄z+ }4 G+ e
4.2 Furnace System with Moving Slabs
−60eT [δi,j ρj cj Dj Sj diag {0 1 1}]i=J,j= 4 4
¯ J¯ (e + z̃) − z̃ . Normal furnace operation requires that slabs are regularly
Substitution of (3b) and consideration of e > ζ1 − z̃ yield pushed forward. Because of sparse and uncertain knowl-
edge about ambient conditions before and after the fur-
V̇ (e) ≤ V̄ (e) = eT P A + AT P nace, Steinboeck et al. (2010) suggested żj (t) = z̃˙ j (t) = 0
−60(G− )T diag{g − }−1 diag{T̄z− }4 G− ∀ j 6∈ J. Therefore, the furnace system with moving slabs
(4) cannot be asymptotically stable.
−60(G+ )T diag{g + }−1 diag{T̄z+ }4 G+
−ζ 3 240[δi,j ρj cj Dj Sj diag {0 1 1}]i=J¯,j=J¯ e.
Consider that the reheating process is to be analyzed
for a time interval where N s (Ns N s < ∞) slabs
There are only negative semidefinite matrices in the last are reheated. Despite N s < ∞, the considered interval
three lines of this expression. Hence, it remains to derive may extend to infinity because the slab movement can
conditions for be stopped at some distant point in the future. The
P A + AT P = [Pi,j (paj + aTi p)]i=J,j=
¯ J¯
control errors in terms of the slabs j ∈ {1, . . . , N s } are
summarized in the vector es (t) = [eT1 (t), . . . , eTN s (t)]T .
= [Pi,j (dj + di )p̄]i=J¯,j=J¯
Proposition 2. The assumptions of Proposition 1 are
with dj = λj /(ρj cj Dj2 ) and adopted but for a furnace with moving slabs. Given that
"
6 −3 −3
# tsl+1 > tsl + ∆tmin,l ∀ l ∈ N with ∆tmin,l according to (6a)
p̄ = −60 −3 2 1 for the period (tsl , tsl+1 ), the control law (3) ensures uniform
−3 1 2 stability of the system.
11747
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011
Proof. It follows from the Lyapunov function candidate meany {Tj (y, t)}
V s (es ) = kes k22 and (6) that V s (es (tsl+1 )) ≤ V s (es (tsl )). Open loop
miny {Tj (y, t)}, maxy {Tj (y, t)}
Consequently, the discrete-time error system with the meany {Tj (y, t)}
state trajectory (es (tsl )) is uniformly stable (cf. Vidyasagar Closed loop
(1992)). Moreover, the continuous-time error es (t) remains miny {Tj (y, t)}, maxy {Tj (y, t)}
finite within each interval (tsl , tsl+1 ). Tj /K Constraints Slab j = 10
5. EXAMPLE PROBLEM 1400
1200 1500
The feedback controller is tested in a simulation environ-
ment using the validated mathematical model presented 1000 1470
by Wild et al. (2009); Wild (2010) and an emulator of the 800
1440
inner control loop from Fig. 2. The considered furnace is 600
35 m long and has a nominal throughput of 280 t/h. 400 1410
tj,0 tj,1 t
5.1 Problem Formulation
Fig. 4. Temperature trajectory of slab j = 10.
In this scenario, the furnace contains Ns = 18 slabs. All
slabs are Dj = 400 mm thick and stay in the furnace for
5 h, i. e., tj,1 = tj,0 + 5 h. The corresponding path-time Open loop Closed loop
diagram is shown in Fig. 3. Tj,end,max maxy {Tj (y, tj,1 )} maxy {Tj (y, tj,1 )}
z T̃j,end meany {Tj (y, tj,1 )} meany {Tj (y, tj,1 )}
Zone 5 Tj,end,min miny {Tj (y, tj,1 )} miny {Tj (y, tj,1 )}
Zone 4
Zone 3 Tj /K
Zone 2 1500
Zone 1
tj,0 tj,1 t 1475
1450
Fig. 3. Path-time diagram.
1425
11748
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011
11749