ISIJ International, Vol. 52 (2012), No. 1, pp.

96–100

Identification and Control of Blast Furnace Gas Top Pressure

Recovery Turbine Unit

Ping WU and Chun-Jie YANG*

State Key Lab. of Industrial Control Technology, Zhejiang University, Hangzhou, 310027 China.
E-mail: Stephen.pingwu@gmail.com, cjyang@iipc.zju.edu.cn
(Received on July 14, 2011; accepted on September 1, 2011)

Blast Furnace (BF) Gas Top pressure Recovery Turbine (TRT) unit is the most popular energy-saving
equipment in iron making process. It is desired to control top pressure of BF precisely and reliably after
installing TRT. The dynamic mathematical model of top pressure with TRT is the key point of controller
design. However, it is difficult to establish the dynamic mathematical model of TRT from first principles.
In this paper, the dynamic mathematical model of No. 1 TRT in Hangzhou Iron & Steel Group Company is
first estimated by subspace identification method. For safety reasons, we adopt closed-loop identification
method. Then a model predictive controller is proposed based on the identified model. Results show that
the identified model can capture the dynamic properties of top pressure process accurately. Moreover,
the model predictive controller gives excellent performance.

KEY WORDS: iron making; blast furnace; top pressure recovery turbine unit; top pressure; dynamic mathe-
matical model; closed-loop subspace identification; model predictive control.

pressure involves not only safety of iron making process but

1. Introduction
Iron and steel making industry is one of the most energy-
intensive industries in the world.1) Iron making process con-
sumes about 70% of the total energy in this industry. The
energy consumption of iron making process is comprised of
39% for blast furnace, 11.9% for coking, 3.51% for balling,
5.55% for sintering, 12.5% for power, 7.77% for rolling
steel, 17.5% for electric furnace, and 2.22% for converter
furnace.2) These statistical data show that blast furnace is the
highest energy consumption unit in the industry. On the oth-
er hand, there is the greatest opportunity for energy saving
in blast furnace.
During iron smelting process, blast furnace gas (BFG)
with high pressure and temperature is produced in blast fur-
nace. In conventional practice, the energy of BFG was wast-
ed by pressure reduction at septum valves. Equipping TRT
unit is the best way to recover the pressure and thermal ener-
gies of BFG. A top pressure recovery turbine is installed in
the downstream of gas-cleaning equipment for a blast fur-
nace. The work generated by the turbine is transferred to the
generator and converted to electric power. A TRT unit could
recover about 20% of electricity attached to the BF includ-
ing air blowers. The experts point out that the blast furnaces
of capacity greater than 1 000 m3 should be equipped with
TRT.3) In china, the blast furnaces larger than 2 000 m3 were
equipped with TRT units at the end of 2007.
The top pressure is controlled by opening or closing the
1st stage stator blades of the turbine in TRT. The stability
of top pressure is of key importance. The small fluctuation
of top pressure is advantageous to the operation of BF. Top
also reduction of raw materials. However, to our knowledge,
there have been no literatures focusing on modeling and
high-precision control of top pressure process in TRT. The
dynamic mathematical model of top pressure process in
TRT is too difficult to set up from first principles. Subspace
identification method (SIM) has been widely accepted for
the identification of linear systems due to its numerical com-
putation advantages and applicability to MIMO systems in
process control society.4) Compared to the prediction error
method (PEM),5) subspace identification uses numerically
robust algorithms such as SVD decomposition. SIM does
not need nonlinear optimization. Hence, they are computa-
tionally faster than PEMs. The model is parameterized by
the full state space model in SIM, and the model order is
decided upon in the identification procedure. The parame-
terization problem may occur in PEM. Moreover, the state
space model is more available in model predictive controller
design. In practice, it is desirable that identification experi-
ments are carried out under the closed-loop for safety rea-
sons or quality restrictions. Several new SIMs are developed
for closed-loop identification since the traditional SIMs
(e.g., CVA, N4SID and MOESP) are biased. Among these
closed-loop identification, Chiuso6) proposed a closed-loop
subspace identification method called PBSID_OPT (Opti-
mized Predictor Based Subspace IDentification). Chiuso7)
proved the excellent performance of PBSID_OPT compared
with several SIM algorithms in closed-loop conditions. It is
shown that PBSID_OPT could derive similar performance
to PEM in simulations.8,9) In this paper, we adopt
PBSID_OPT method to build the dynamic mathematical

© 2012 ISIJ 96
 ISIJ International, Vol. 52 (2012), No. 1

model of No.1 TRT in Hangzhou Iron & Steel Group Com- sampled data are acquired under closed loop. The model
pany under closed loop. A model predictive controller is identified under closed loop can obtain better performance
implemented by using this identified model. in control design scheme.
Section 2 is devoted to illustrating top pressure process in
TRT. In Section 3, PBSID-OPT method is briefly described.
3. Closed-loop Subspace Identification Method
The model predictive controller design is included and
experimental results are demonstrated in Section 4. Section In the following section, the PBSID_OPT algorithm is
5 states our conclusions. briefly illustrated. A stochastic linear system can be written
in the following innovation form,
2. Process Descriptions x(k + 1) = Ax(k ) + Bu(k ) + Ke(k ) ...............(1a)
There are two types of turbines: radial and axial turbines. y(k ) = Cx(k ) + Du(k ) + e(k ) ................. (1b)
Currently, axial turbines are widely used in TRT because
they are more appropriate for dealing with a large flow vol- where y(k ) ∈ \ ny , x(k ) ∈ \ n and u(k ) ∈ \ nu are the system
ume. A dry gas clean-up system raises the turbine inlet tem- output, state and input, respectively. A,B,C,D,K are system
perature compared to the wet type, increasing the power matrices with appropriate dimensions. The innovation
recovery by about 25–30%.10) In this paper, we will focus e(k ) ∈ \ ny is white noise and independent of past input and
on the dry type TRT unit with axial turbine. output data. The system described by (1) can also be repre-
The schematic flow diagram of this process is shown in sented in the predictor form,
Fig. 1. After dust is collected by gas-cleaning equipment,
clean BFG is led to the turbine through an inlet butterfly x(k + 1) = Ax(k ) + Bz(k ) .....................(2a)
valve and an inlet goggle valve. Normally closed bypass y(k ) = Cx(k ) + Du(k ) + e(k ) ................. (2b)
valves and normally open emergency shutoff valve are used
for emergency case. The mechanical work generated by tur- ⎡u(k ) ⎤
where z(k ) = ⎢ ⎥ , A = A − KC and B = [ B − KD K ].
bine is transferred to a generator for electric power. BFG ⎣ y( k ) ⎦
with low pressure and temperature enters into gas line Based on state space description(2), an extended state
through the outlet goggle valve. space model with stack vector can be formulated,
The operation of TRT can be divided into two phases:
y f (k ) = Γ f x(k ) + G f z f −1 (k ) + D f u f (k ) + e f (k ) ..... (3)
start-up phase and normal running phase. In the start-up
phase, all preparations such as hydraulic sub-system must be where f denotes the future horizon similar as other litera-
checked. Then a signal for starting TRT is sent to BF oper- tures. The matrices are defined as following form,
ator. If BF operator agrees the request, the inlet butterfly
valve and emergency shutoff valve will be opened. The ⎡C ⎤ ⎡D 0 " 0 ⎤
speed of turbine is controlled by the inlet butterfly valve and ⎢ ⎥ ⎢0 ⎥
CA ⎥ D " 0
top gas pressure is controlled by the septum valves. Until the Γf = ⎢ Df = ⎢ ⎥
⎢ # ⎥ ⎢ # # % # ⎥
turbine reaches the required rotate speed, the septum valves ⎢ f −1 ⎥ ⎢ ⎥
are closed slowly. Then the start-up phase turns into the nor- ⎢⎣CA ⎥⎦ ⎣0 0 " D ⎦
mal running phase. In the normal running phase, top gas
pressure is fully regulated by the 1st stage stator blades of ⎡ 0 0 " 0⎤
the turbine. The top pressure is controlled by opening or ⎢ ⎥
CB 0 " 0⎥
closing the 1st stage stator blades of the turbine. We will Gf = ⎢
⎢# ⎥
focus on the top pressure control system in this paper. ⎢ ⎥
The dynamic mathematical model of TRT in normal run- ⎢⎣CA f − 2 B CA f −3 B " CB ⎥⎦
ning phase is considered in this paper. Here, we consider the
opening of 1st stage stator blades of the turbine as the input The input and output stack vectors are arranged in the fol-
variable and top pressure as the output variable. The lowing form:
⎡ y( k ) ⎤ ⎡ z( k ) ⎤
⎢ y(k + 1) ⎥ ⎢ z(k + 1) ⎥
y f (k ) = ⎢ ⎥ , z (k ) = ⎢ ⎥
⎢ # ⎥ f −1
⎢ # ⎥
⎢ ⎥ ⎢ ⎥
⎣ y(k + f − 1) ⎦ ⎣ z( k + f − 2) ⎦
From (2), it is straightforward to derive the following
relation,
x(k ) = L p z p (k ) + A p x(k − p) ................... (4)

where L p = ⎡⎣ B AB " A p −1 B ⎤⎦ and z p = ⎡⎣ z(k − 1)T z(k − 2)T "

T
z (k − p)T ⎤⎦ .

Fig. 1. Schematic flow diagram of TRT. Combine (4) with (3),

97 © 2012 ISIJ
 ISIJ International, Vol. 52 (2012), No. 1

Fig. 2. Experiment data.

y f (k ) = Γ f L p z p (k ) + Γ f A p x(k − p) + G f z f −1 (k ) + D f u f (k ) + e f (k )
.......................................... (5)
Suppose that p is chosen large enough so that A p ≈ 0 .
And assume D = 0 for convenience,
y f (k ) = Γ f L p z p (k ) + G f z f −1 (k ) + e f (k ) ........... (6)
From (6), we can obtain the following row-wise equation
by exploiting the structure of G f . For i=1,2,...f,

y(k + i − 1) = Γ f (i )L p z p (k ) + G f (i )zi −1 (k ) + e(k + i − 1) ... (7)

where Γ f (i ) = CAi −1 and G f (i ) = ⎡⎣CA B

f −2
CA f −3 B "
CB⎤⎦ .
If we estimate Γ f (i )L p from (7) using least square meth-
Fig. 3. Singular values of SVD decomposition.
od, Γ f L p can be constructed from Γ f (i )L p . Then we obtain,

Γ f X (k ) = Γ f L p Z p (k ) ........................ (8) PBSID_OPT algorithm:

where X(k) and Zp(k) are block-Hankel versions of x(k) and ⎡0.93191 0.17499 0.052944 ⎤ ⎡-0.03676 ⎤
Zp(k).5) In order to estimate X(k), we perform SVD decom- A = ⎢⎢0.34989 0.17269 -0.025735⎥⎥ , B = ⎢0.21576 ⎥
⎢ ⎥
position on Γ f X (k ) and choose the order n from the singu- ⎢⎣0.91217 -2.0
0381 0.87209 ⎥⎦ ⎢⎣0.53113 ⎥⎦
lar values plots. Since X(k) is derived, the system matrices
A,B,C,D,K are estimated by least square algorithm.4,6) C = [-1.7417 -0.036573 -0.035935 ] , D = 0,

4. Experiment Results
The No.1 blast furnace at Hangzhou Iron & Steel Group
Company has a volume of 1 250 m3. The TRT unit of No. 1
blast furnace is a dry-type top pressure recovery turbine
plant. The output power of TRT is in range from 5 625 kW
to 6 100 kW.
We performed an identification experiment for the TRT
under closed loop. The data were collected from the exper-
iment with 1 second sample period. In order to include dif-
ferent working conditions of blast furnace such as tapping
and charging, the data were collected for 2 hours. Hence,
7 200 data were obtained. The first 3 600 data are used for
estimation and the left 3 600 data for validation. The sam-
pled data are displayed in Fig. 2.
⎡-0.68928 ⎤
K = ⎢⎢-0.32728 ⎥⎥
⎢⎣-3.7913 ⎥⎦
Figures 4(a) and 4(b) show the predicted output and the
real output using estimation and validation data. It is clear
that the model predicts well in output. Figures 4(c) and 4(d)
show the errors between the predicted output and the real
output. As these figures shown, the identified model has
enough accuracy to describe the dynamic characteristics of
top pressure process.
As an evaluation of the identified model, we use the coef-
ficient of determination comparing with the classic PEM,
N

Choose p=8 and f=8 in (5), Fig. 3 shows the singular val- ∑ ( y(k ) − yˆ (k ))2
i =1
ue of SVD decomposition in (8). R 2 =max{1 − N
,0} × 100%
From Fig. 3 , it is obvious that the best choice of system ∑ ( y(k ) − y (k ))2
i =1
order is 3. Then we obtain the following system matrices by

© 2012 ISIJ 98
 ISIJ International, Vol. 52 (2012), No. 1

Table 1. Coefficients of determination of two methods.

Data PBSID_OPT PEM

Estimation 95.798% 95.8%
Validation 95.55% 95.55%

(a)

(b)

Fig. 5. Opening (top) and top pressure (bottom) in scenario 1.

where N is the number of data points. y(k), yˆ (k ) and y (k )

(c)
(d)
Fig. 4. (a) Predicted output and real output using estimation data.
(b) Predicted output and real output using validation data.
(c) Errors between predicted output and real output using
estimation data. (d) Errors between predicted output and
real output using validation data.
are the measured output, predicted output and the mean of
the output for the kth output variables, respectively.
The coefficients of determination on the estimation and
validation data are shown in Table 1.
From Table 1, the identified model by PBSID_OPT has
the similar performance to PEM model. And the coefficients
of determination are over 95%. We can consider that the
identified model by PBSID_OPT is suitable for controller
design.
A model predictive controller based on the identified
model is proposed.11) For the opening of 1st stage stator
blades, the upper limit is 100% and the lower limit is 0%.
We consider two scenarios: (1) Set point is 167 kPa all the
time. (2) Set point is a pulse with 50 seconds period. The
plots of opening and top pressure in scenario 1 are shown
in Fig. 5. From Fig. 5, we can see that the fluctuation of top
pressure is within ± 2 kPa. The fluctuation is within ± 1 kPa
at most times. The controller based on the identified model
can well decrease the fluctuation of top pressure. In scenario
2, the top pressure can track the reference trajectory quickly

99 © 2012 ISIJ
 ISIJ International, Vol. 52 (2012), No. 1

and obtain a small fluctuation as shown in Fig. 6.

5. Conclusions
In this paper, we use a closed-loop subspace identification
method (PBSID_OPT) to identify the dynamic mathemati-
cal model of No.1 TRT in Hangzhou Iron & Steel Group
Company. The identified state-space model can accurately
predict the dynamic behavior of TRT top pressure process.
A model predictive controller is developed based on the
identified model. The controller can well decrease the fluc-
tuation of top pressure and keep the top pressure close to set
point.

Acknowledgements
This work is supported by National High-tech R&D
Program of China (863 Program) (Grant No.
2007AA041406 and No. 2006AA04Z184).

REFERENCES
1) C. B. Xu and D. Q. Cang: J. Iron Steel Res., 17 (2010), 1.
2) Z. C. Guo and Z. X. Fu: Energy, 35 (2010), 4356.
3) S. Zeng and L.Y. Huang: Int. J. Greenhouse Gas Contr., 3
(2009),675.
4) S. J. Qin: Comp. Chem. Eng., 30 (2006), 1502.
5) L. Ljung: System identification – theory for the user, Prentice Hall,
NJ, USA, (1999).
6) A. Chiuso: IEEE Trans. Auto. Contr., 52 (2007), 1795.
7) A. Chiuso: Automatica, 43 (2007), 1034.
8) M. Jansson: Proc. of the 13th IFAC SYSID Symp., Elsevier IFAC Pub-
lication, Rotterdam, NL, (2003), 500.
9) W. Larimore: Proc. of the 7th Int. Conf. on Dynamics and Control of
Process Systems, IFAC, Cambridge, USA, (2004).
10) Stelco: Present and Future Use of Energy in the Canadian Steel
Industry, CANMET, Ottawa, Canada, (1993).
11) R. Kadali, B. Huang and A. Rossiter: Control Eng. Pract., 11 (2003),
Fig. 6. Opening (top) and top pressure (bottom) in scenario 2. 261.

© 2012 ISIJ 100