Professional Documents
Culture Documents
DOI 10.1007/s00521-015-1951-7
REVIEW
Received: 4 October 2014 / Accepted: 5 June 2015 / Published online: 23 June 2015
The Natural Computing Applications Forum 2015
Abstract As a highly complex multi-input and multi-output processes, blast furnace is the major source of sulfur
system, blast furnace plays an important role in industrial dioxide emissions and energy consumptions. It is a com-
development. Although much research has been done in the past plex industrial reactor used to produce hot metal from iron
few decades, there still exist many problems to be solved, such ore for subsequent processing into steel. When a blast
as the modeling problem. This paper adopts support vector furnace runs, most chemical reactions happen among dif-
regression (SVR) to construct the prediction model of blast ferent phases including gas phase, liquid phase and solid
furnace silicon content. To ensure a good generalization per- phase, along with high temperature and high pressure [1].
formance for the given datasets, it is important to select proper This leads to large time delay in reaction process and
parameters for SVR. In view of this problem, a new particle spatiotemporal characteristics, such as it will take about
swarm optimizer called DMS-PSO-CLS is presented to opti- 6–8 h for a cycle of iron-making [2]. Due to the compli-
mize the parameters of SVR. In DMS-PSO-CLS, a new coop- cated process and hostile environment, current study
erative learning strategy is hybridized with DMS-PSO, which focuses on multi-fields, such as modeling, control method,
makes particle information be used more effectively for gener- performance optimizing and metallurgical analysis [3].
ating better-quality solutions. DMS-PSO-CLS takes merits of Researches show that modeling of blast furnace often
the DMS-PSO and the cooperative learning strategy so that both means modeling of its thermal state since the thermal state
the convergence speed and the convergence precision can be represents the hot metal quality and energy utilization state.
improved. Experimental results show that DMS-PSO-CLS can The silicon content in hot metal has a linear approximation
find the optimal parameters of SVR with high speed and the relation with the thermal state. Thus, a great deal of data-
SVR model optimized by DMS-PSO-CLS can achieve a good driven models have been developed and used to predict the
regression precision on the predictive problem of blast furnace. hot metal silicon content since the modeling process needs
only a large set of observations but not prior information
Keywords Blast furnace Hot metal silicon content about the process. Neural networks [4–6], fuzzy logic [7],
Support vector regression Particle swarm optimizer support vector regression [8, 9], state space [10, 11], partial
least squares recursive analysis [12] and evolutionary net-
1 Introduction works [13, 14] have been successfully applied to predict
the silicon content.
Iron and steel making, a typical high-energy-consuming, The support vector machine (SVM) is a powerful tool for
high-emission and high-pollution industry, is a pillar document classifications and pattern recognitions since it is
industry for Chinese national economy. Among all the proposed. Its main idea is to embed the inputs into a feature
space through a high-dimension mapping, so that an opti-
mal decision hyperplane can be found among the high-di-
& Xia Xu mension embedded data points [15]. In order to find a
xuxiajnz@163.com
decision rule with good generalization capability, the so-
1
Institute of Electrical Engineering, Yanshan University, called support vectors (SVs), including a small subset of the
Qinhuangdao 066004, China training data, are selected to support the optimal hyperplane
123
1452 Neural Comput & Applic (2016) 27:1451–1461
[16]. SVM can overcome the overfitting problem and iron and slag are tapped periodically from the bottom.
achieve enhanced generalization performance by minimiz- Meanwhile, flue gases exit from the top of the furnace.
ing the structural risk instead of the empirical risk. Support During the process, iron ores pretreated by agglomera-
vector regression (SVR) is later exploited to cope with the tion into sinter or pellets are fed into the top and charged in
regression estimation and function approximation prob- alternate layer with coke. Preheated compressed air toge-
lems. SVR has shown breakthroughs and excellent perfor- ther with pulverized coal is injected lower down through
mance in many areas [17–21]. Since the quality of SVR tuyeres which are just above the tapholes. The gases ascend
models depends on a proper setting of SVR parameters, the through the charge and reduce the iron oxides to a metallic
key issue for applying SVR is how to set its parameters. phase and a slag phase. The oxygen in iron oxides is
Existing sources on SVR have given some suggestions on removed by a series of chemical reactions. These reactions
appropriate setting of SVM parameters [22, 23]. occur as follows:
Particle swarm optimization (PSO) is a population- 3Fe2 O3 þ CO ! 2Fe3 O4 þ CO2 ; begins at 450 C
based stochastic optimization algorithm which was origi-
Fe3 O4 þ CO ! 3FeO þ CO2 ; begins at 600 C
nally introduced by Eberhart and Kennedy [24]. PSO is
considered to be a robust algorithm in many applications, FeO þ CO ! Fe þ CO2 or FeO þ C ! Fe þ CO;
and a number of PSO variants have been developed [25, begins at 700 C or higher
26]. Multi-swarm technique has attracted increasing
attention during the last decade. It is one of the effective The iron oxides go through these purifying reactions and
approaches maintaining diversity of swarm [27–29]. In finally trickle as liquid iron through the coke to the bottom of
[30], a dynamic multi-swarm particle swarm optimizer the furnace. At the same time, the coke descends to the
(DMS-PSO) is introduced through dividing the whole bottom of the furnace to the level where the preheated air or
population into many small sub-swarms. These sub-swarms hot blast enters the blast furnace. The coke is ignited by this
are regrouped frequently using a certain regrouping hot blast and immediately reacts to generate heat as follows:
schedule, so that particle information can be exchanged C þ O2 ! CO2 þ Heat
among the sub-swarms. However, the frequent regrouping
Since the reaction takes place in the presence of excess
operation results in the deficiency of exploitation. To
carbon at a high temperature, the carbon dioxide is reduced
overcome this drawback, we hybridize DMS-PSO with a
to carbon monoxide as follows:
new cooperative learning strategy here, which can be
called DMS-PSO-CLS. The new strategy makes more CO2 þ C ! 2CO
particle information be exchanged among sub-swarms,
The product of this reaction, carbon monoxide, is necessary
which is not limited to the regrouping operation. Based on
to reduce the iron ore as seen in the previous iron oxide
this, a good balance between the local exploitation and the
reactions. The limestone descends in the blast furnace and
global exploration can be achieved.
remains solid while going through its first reaction as
In this paper, the proposed DMS-PSO-CLS is used to
follows:
search the optimal parameters for SVR. Then, the predic-
tion model of hot metal silicon content in blast furnace can CaCO3 ! CaO þ CO2
be established using the DMS-PSO-CLS-based SVR.
This reaction requires energy and starts at about 875 C.
This paper is organized as follows: In Sect. 2, the iron-
The CaO formed from this reaction is used to remove
making process of blast furnace is presented. This is fol-
sulfur from the iron which is necessary before the hot metal
lowed by a brief review of SVR for the modeling process in
becomes steel. This sulfur-removing reaction is
Sect. 3. The DMS-PSO-CLS-based SVR is described in
Sect. 4. Application and validation of the proposed pre- FeS þ CaO ! CaS þ FeO
diction model, and the detailed analysis about the experi- CaO þ SiO2 ! CaSiO3
mental results are presented in Sect. 5. Finally, conclusions
are presented in Sect. 6. The reactant becomes part of the slag. The slag is also
formed from the remaining silica (SiO2 ), alumina (Al2 O3 ),
magnesia (MgO) or calcia (CaO) that entered with the iron
2 Blast furnace iron-making process ore, pellets, sinter or coke. The liquid slag then trickles
through the coke bed to the bottom of the furnace. It floats
Blast furnace smelting process is a complex dynamic non- on the top of the liquid iron since it is less dense. The
linear process, which has been used as a benchmark problem desired final product, liquid hot metal, is produced at the
for black-box analysis. Figure 1 presents a typical schematic same time, which is further processed to produce steel or
diagram of blast furnace. The end products including molten other alloys [31].
123
Neural Comput & Applic (2016) 27:1451–1461 1453
Hot metal temperature is the most important sign to For a given training set D ¼ fðxi ; yi Þ; i ¼ 1; . . .; ng,
judge the thermal state of blast furnace, which can be where xi is the ith input vector, yi is the corresponding
evaluated from some measurable indicators. Among these output and n is the total number of exemplars. We consider
indicators, silicon content in hot metal is identified as the the case can be written as a nonlinear mapping as follows:
most important one. It has been shown that silicon content f ðxÞ ¼ wT uðxÞ þ b ð1Þ
has a linear approximation relation with the thermal state
[2]. The control of a blast furnace system often means to where uðxÞ maps data points xi into a higher-dimensional
control the hot metal temperature and silicon content. Thus, space. In SVR, the goal is to find a function f ðxÞ that has at
modeling of silicon content is crucial for blast furnace. most e deviation from the actually obtained output yi .
The objective function and constraints for e-SVR can be
described as follows, so that the data which lie outside the
e-insensitive band are penalized.
3 Support vector regression X n
1
min k w k2 þC ni þ ni
w;b 2
SVR is to construct a so-called optimal separating hyper- i¼1
8
plane in a high-dimensional (even infinite-dimensional) T ð2Þ
< yi ðw uðxi Þ þ bÞ 6 e þ ni
>
feature space by maximizing the margin among the nearest T
s:t: w uðxi Þ þ b yi 6 e þ ni
training data points [8]. The selection of kernel function >
:
ni ; ni > 0; i ¼ 1; . . .; n
and homologous kernel parameters is very important for
the accuracy of predictive results. Among various types of where C [ 0 is the parameter which gives a trade-off
support vector regression, the most commonly used is e- between the model complexity and training error, e [ 0
SVR which finds a regression hyperplane with an e-in- controls the admissible uncertainty on the data points,
sensitive band. For convenience, in the sequel, the term ni and ni are slack variables, ni denotes the training error
SVR is used to stand for e-SVR. above e, whereas ni denotes the training error below e.
123
1454 Neural Comput & Applic (2016) 27:1451–1461
To solve Eq. (2), Lagrange multiplier method can be The process can be converted into a mathematical
used with turning the Lagrange into the following dual form: problem as follows. Suppose that sz particles are used to
Xn X
n search the solution. The ith particle in the D-dimensional
max
y a
i i þ a i e ai þ ai space is represented as xi ¼ ðx1i ; x2i ; . . .; xdi ; . . .; xDi Þ, where
a;a
i¼1 i¼1
xdi 2 ½xmin ; xmax , d 2 ½1; D. The velocity corresponding to
1X n X n
the ith particle is vi ¼ ðv1i ; v2i ; . . .; vdi ; . . .; vD
ai ai ai ai K xi ; xj i Þ, where
2 i¼1 j¼1 d
vi 2 ½vmin ; vmax . The velocity and location of the ith par-
8P n ð3Þ
> ticle are updated through the formula below:
>
> ai ai ¼ 0
>
>
< j¼1 vdi vdi þ c1 rand1di ðpbestid xdi Þ þ c2 rand2di
s:t: 0 6 ai 6 C ðgbestd xdi Þ ð5Þ
>
>
>
> 0 6 ai 6 C
>
: xdi ¼ xdi þ vdi ð6Þ
i ¼ 1; . . .; n
where c1 and c2 are the acceleration constants. c1 repre-
where ai and ai are Lagrange multipliers and K xi ; xj is a
sents the weight that the ith particle tracks its own histor-
kernel function which can be chosen from the following
ical optimum value pbesti . c2 represents the weight that the
functions:
ith particle tracks the whole group’s optimum value gbest.
• Gaussian radial basis function (GRBF) kernel: pbesti and gbest are updated all the time according to each
0 particle’s fitness value. rand1di and rand2di are two random
kx x k2
kðx; x0 Þ ¼ exp ; variables in the range [0, 1].
2r2
To control the flying velocity, an inertia weight or
• Polynomial kernel: a constriction factor is introduced by Shi and Eberhart
q [33] as
kðx; x0 Þ ¼ ð1 þ hx; x0 iÞ ;
vdi w vdi þ c1 rand1di ðpbestid xdi Þ
• Sigmoid kernel: ð7Þ
0 0
þ c2 rand2di ðgbestd xdi Þ
kðx; x Þ ¼ tanhðahx; x i þ cÞ:
where w usually decreases linearly from 0.9 to 0.4 during
where r; q; a; c are the adjustable parameters of the above the iterative process [34].
kernel functions. The Gaussian radial basis function kernel Substantially, PSO is divided into two versions. The
is used here. above formula is global PSO, and another version is local
By applying the Karush–Kuhn–Tucker theorem, the PSO. For local PSO, each particle adjusts its position and
above constrained optimization problem can be solved velocity according to its historical best position pbesti and
[32]. Finally, the regression hyperplane for the underlying the best position achieved so far from its group lbesti . The
regression problem can be given by velocity update strategy is described as follows:
X
f ðxÞ ¼ aþ
i ai kðx; xi Þ þ b
ð4Þ vdi w vdi þ c1 rand1di ðpbestid xdi Þ
ð8Þ
xi 2SV þ c2 rand2di ðlbestid xdi Þ
P
where b ¼ yi þ e xi 2SV aþ
i ai kðx; xi Þ and SV
represents the small set of support vectors with aþ 4.2 DMS-PSO-CLS
i ai 6¼ 0.
123
Neural Comput & Applic (2016) 27:1451–1461 1455
Although DMS-PSO can solve problems efficiently, though the globally optimal region is found, the particles
there still exist some drawbacks. In DMS-PSO, each par- cannot converge rapidly to the globally optimal solution.
ticle in a sub-swarm only learns from its pbest and lbest. DMS-PSO achieves great improvement on global explo-
Information among different sub-swarms cannot be ration but lacks local exploitation. Aiming at this drawback,
exchanged until the population is regrouped. Due to the a new cooperative learning strategy is integrated with DMS-
deficiency of cooperative learning among sub-swarms, even PSO, which can be called DMS-PSO-CLS. It is used to
exchange particle information among different sub-swarms
sufficiently. In this way, the collaborative learning among
sub-swarms is enhanced and the balance between the global
exploration and the local exploitation can be achieved. The
cooperative learning procedure is described as follows:
1. For each sub-swarm, we sort the fitness values of the
particles and select the two worst particles to be updated.
Regroup 2. For each particle’s each dimension, we select two sub-
swarms randomly out of the whole groups.
3. We compare the fitness values of the two sub-swarms’
lbests and select the better one.
4. We use the winner’s lbest as the exemplar to learn
from for the corresponding dimension of the particle to
be updated.
Fig. 2 DMS-PSO’s search process The pseudo-code of DMS-PSO-CLS is shown in Algorithm 1.
123
1456 Neural Comput & Applic (2016) 27:1451–1461
123
Neural Comput & Applic (2016) 27:1451–1461 1457
Step 1 Get the reference X0 ¼ ðx0 ð1Þ; x0 ð2Þ; . . .; x0 ðnÞÞ normalized since the values of silicon content are too
and comparison sequences concentrated.
Xi ¼ ðxi ð0Þ; xi ð1Þ; . . .; xi ðnÞÞ; i ¼ 1; 2; . . .m.
Step 2 Calculate the relational degree. 5.3 Parameters setting
To determine the relational degree between the
reference and comparison sequences, a discrete Fitness function is a very important mark to measure the
function of the relational degree coefficient (the parameters found by DMS-PSO-CLS for SVR. In this
gray relational coefficient) is represented by study, it is defined as the leave-one-out cross-validation. In
Dmin þ qDmax addition, PSO and DMS-PSO are also used to optimize the
10i ðkÞ ¼ ; k ¼ 1; 2; . . .; n ð9Þ parameters of SVR. The parameters of the three algorithms
D0i ðkÞ þ qDmax
are set as follows. The population size is set at 20, the sub-
where swarms 5, the maximum iteration 200, the dim 2, and
D0i ðkÞ ¼j xi ðkÞ; x0 ðkÞ j ð10Þ ranges of C and r2 are [0.1, 100] and [0.01, 1000].
123
1458 Neural Comput & Applic (2016) 27:1451–1461
DMS-PSO-CLS-based SVR
0.6
0.5
0.4
0.3
0.2
0 50 100 150 200 250 300
Sample sequence
It is worth mentioning that each method is repeated 20 SVR is 100 % with 60 % for PSO-based SVR and 40 %
times independently. For the SVR optimized by DMS- for DMS-PSO-based SVR. Obviously, DMS-PSO-CLS can
PSO-CLS, there are 20 successful results with never performance stably.
missing the optimal value, but only 12 and 8 successful From Sect. 3, the proportion’s size of the two stages in
results can be achieved by SVRs optimized by PSO and DMS-PSO-CLS may be influential for the predictive per-
DMS-PSO. The success rate of the DMS-PSO-CLS-based formance. We do the experiments with different
123
Neural Comput & Applic (2016) 27:1451–1461 1459
0.6
0.55
0.5
0.45
0.4
0.35
0.6
0.55
0.5
0.45
0.4
0.35
Sample sequence
0.6
0.55
0.5
0.45
0.4
0.35
Sample sequence
proportions ranging from 0 to 1 with the interval of 0.1. region at very fast speed without trapping into the local
Finally, we found that the predictive results achieved by optimum, while the other two methods may either fall into
different proportions are almost the same although the the minimum or converge slowly.
proportion is 0.1. The only difference is the convergence From the results of comparison, it is obvious that the
rate of DMS-PSO-CLS with the larger proportions the DMS-PSO-CLS-based SVR has better prediction accuracy
slower convergence rate, which almost remain the same and faster convergence speed than the other two methods
when the proportions is [0.1, 0.8]. The results verify that and its prediction results are stable. The reason why DMS-
the proportion’s size of the two stages has little impact on PSO-CLS can optimize SVR best for hot metal silicon
the performance of DMS-PSO-CLS. Figures 8, 9 and 10 content is that DMS-PSO-CLS takes merits of the DMS-
present the search performance of the three heuristic PSO and the new cooperative learning strategy. With the
techniques. From these analysis and figures, it is clearly novel cooperative learning strategy, more information can
seen that DMS-PSO-CLS can lead the particles to optimal be exchanged among the sub-swarms. The new strategy
123
1460 Neural Comput & Applic (2016) 27:1451–1461
0.14
0.135 the industrial problem.
0.13
0.125
0.12 6 Conclusion
0.115
0.11 This paper has developed a SVR model to predict the hot
0 50 100 150 200 250
Sample sequence
metal silicon content of blast furnace, where the two
parameters of SVR are optimized by an improved opti-
Fig. 8 Variation trend of the best and average fitness values of DMS- mization algorithm which is called DMS-PSO-CLS. Sim-
PSO-CLS-based SVR ulation results show that the DMS-PSO-CLS-based SVR
has a good regression performance. We can observe that
0.15 DMS-PSO-CLS can improve the search precision and
the best fitness
0.145
the average fitness simultaneously quicken the convergence speed of DMS-
PSO. It can also be seen that the proposed new optimiza-
0.14
tion algorithm is effective in optimizing the parameters of
0.135
Fitness
0.155
the best fitness
the averarge fitness
0.15
References
0.145
0.14 1. Gao CH, Jian L, Liu XY, Chen JM, Sun YX (2011) Data-driven
Fitess
123
Neural Comput & Applic (2016) 27:1451–1461 1461
9. Jian L, Gao CH, Xia ZQ (2011) A sliding-window smooth sup- 23. Kwok JT (2001) Linear dependency between e and the input
port vector regression model for nonlinear blast furnace system. noise in e-support vector regression. In: Proceedings of interna-
Steel Res Int 82:169–179 tional conference on artificial neural networks, pp. 405–410
10. Phadke MS, Wu SM (1974) Identification of multiinput–multi- 24. Eberhart RC, Kennedy J (1995) A new optimizer using particle
output transfer function and noise model of a blast furnace swarm theory. In: Proceedings of the sixth international sympo-
from closed-loop data. IEEE Trans Autom Control AC–19(6): sium on micro machine and human science, pp. 39–43
944–951 25. Ratanavilisagul C, Kruatrachue AB (2014) A modified particle
11. Saxén H, Östermark R (1996) State realization with exogenous swarm optimization with mutation and reposition. Int J Innov
variables—a test on blast furnace data. Eur J Oper Res 89:34–52 Comput Inform Control 10(6):2127–2142
12. Bhattacharya T (2005) Prediction of silicon content in blast fur- 26. Dhanraj AV, Nanjundappan D (2014) Design of optimized PI
nace hot metal using partial least squares (PLS). ISIJ Int controller with ideal decoupler for a non linear multivariable
45:1943–1945 system using particle swarm optimization technique. Int J Innov
13. Pettersson F, Chakraborti N, Saxén H (2007) A genetic algo- Comput Inform Control 10(1):341–355
rithms based multi-objective neural net applied to noisy blast 27. Zhang JZ, Ding XM (2011) A multi-swarm self-adaptive and
furnace data. Appl Soft Comput 7:387–397 cooperative particle swarm optimization. Eng Appl Artif Intell
14. Pettersson F, Saxén H, Deb K (2009) Genetic algorithm-based 24:958–967
multi-criteria optimization of ironmaking in the blast furnace. 28. Niu B, Zhu YL, He XX, Wu H (2007) MCPSO: a multi-swarm
Mater Manuf Process 24:343–349 cooperative particle swarm optimizer. Appl Math Comput
15. Saxén H, Gao CH, Gao ZW (2012) Data-driven time discrete 185:1050–1062
models for dynamic prediction of the hot metal silicon content in 29. Zhao SZ, Liang JJ, Suganthan PN, Tasgetiren MF (2008)
the blast furnace. IEEE Trans Ind Inform 9(4):2213–2225 Dynamic multi-swarm particle swarm optimizer with local search
16. Zhao YP, Sun JG (2011) Multikernel semiparametric linear for large scale global optimization. In: Proceedings of the IEEE
programming support vector regression. Expert Syst Appl congress on evolutionary computation, pp. 3845–3852
38:1611–1618 30. Liang JJ, Suganthan PN (2005) Dynamic multi-swarm particle
17. Musicant DR, Alexander F (2004) Active set support vector swarm optimizer. In: Proceedings of the IEEE congress on swarm
regression. IEEE Trans Neural Netw 15(2):268–275 intelligence symposium, pp. 124–129
18. Tian J, Gu H (2010) Anomaly detection combining one-class 31. Zeng JS, Gao CH, Su HY (2010) Data-driven predictive control
SVMs and particle swarm optimization algorithms. Nonlinear for blast furnace ironmaking process. Comput Chem Eng
Dyn 61(1–2):303–310 34:1854–1862
19. Mao WT, Mu XX, Zheng YB, Yan GR (2014) Leave-one-out 32. Fletcher R (1987) Practical methods of optimization, 2nd edn.
cross-validation-based model selection for multi-input multi-output Wiley, New York
support vector machine. Neural Comput Appl 24(2):441–451 33. Shi Y, Eberhart RC (1998) A modified particle swarm optimizer.
20. Zhao MY, Ren J, Ji LP, Fu C, Li JP, Zhou MT (2012) Parameter In: Proceedings of the IEEE world congress on evolutionary
selection of support vector machines and genetic algorithm based computational intelligence, pp. 69–73
on change area search. Neural Comput Appl 21(1):1–8 34. Eberhart RC, Shi Y (2000) Comparing inertia weights and con-
21. Tian J, Gu H, Gao CY, Lian J (2010) Local density one-class striction factors in particle swarm optimization. In: Proceedings
support vector machines for anomaly detection. Nonlinear Dyn of the IEEE congress on evolutionary computation, pp. 84–88
64(1–2):127–130 35. Zhao J, Wang W, Liu Y, Pedrycz W (2011) A two-stage online
22. Deng NY, Tian YJ (2004) New method of data mining-support prediction method for a blast furnace gas system and its appli-
vector machine. Science Press, Beijing cation. IEEE Trans Control Syst Technol 19(3):507–520
123