Neurocomputing 174 (2016) 232237

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

An improved ELM algorithm for the measurement of hot metal

temperature in blast furnace
Haigang Zhang, Yixin Yin, Sen Zhang n
School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China

art ic l e i nf o a b s t r a c t

Article history: This note proposes a modied ELM algorithm named P-ELM subject to how to get rid of the
Received 17 September 2014 multicollinear problem in calculation based on PCA technique. By reducing the dimension of hidden
Received in revised form layer output matrix (H) without loss major information through PCA theory, the proposed P-ELM
13 April 2015
algorithm can not only ensure the full column rank of newly generated hidden layer output matrix (H 0 ),
Accepted 14 April 2015
Available online 19 August 2015
but also improve the training speed. In order to verify the effectiveness of P-ELM algorithm, this paper
establishes a soft measurement model for hot metal temperature in the blast furnace (BF). Some
Keywords: comparative simulation results with other famous feedforward neural network and the ordinary ELM
Extreme learning machine algorithm with its variants illustrate the better generalization performance and stability of the proposed
Principle component analysis
P-ELM algorithm.
Blast furnace
& 2015 Elsevier B.V. All rights reserved.
Hot metal temperature
Soft measurement model

1. Introduction approach to get rid of the multicollinear problem based on PCA

Extreme Learning Machine (ELM), proposed by Huang et al., is a
novel machine learning algorithm for single hidden layer feedfor-
ward neural networks (SLFNs) [1]. In contrast to other famous
neural networks algorithms, the learning parameters, i.e., the
input weights and the bias of hidden nodes in ELM are generated
randomly without human tuning, while the output weights are
determined based on the method of least squares [1,2,6]. Unlike
the traditional feedforward neural networks learning algorithms,
ELM has the fast training speed and gets rid of the opportunity to
converge to local minima. Nowadays, because of its good general-
ization, ELM algorithm has been applied in many aspects like
image segmentation [3], fault diagnosis [4,5], human action recog-
nition, and human computer interface [7].
So far, ELM algorithm has drawn more and more attention.
However, the multicollinear problem in ELM algorithm deterio-
rates its generalization performance in many complex industrial
process. When to calculate the output weights, Huang just adds a
T
threshold

to every
  1 element in the main diagonal of matrix H H
^
H H I=C
T T
H T [1], which links the ELM theory with
other mathematical theory such as matrix theory and ridge
regression. In our previous work, LDLT decomposition to HTH is
employed to overcome the multicollinear problem (R-ELM) [8].
One just needs to set a proper threshold to some singular elements
to matrix D after the decomposition. Here, we present a novel
n
Corresponding author.
E-mail address: zhangsen@ustb.edu.cn (S. Zhang).
theory. Compared with other related algorithms, the proposed
algorithm can make sure that the estimation of output weights is
unbiased. Simultaneously, the proposed ELM algorithm can also
improve the training and testing time in contract to the ordinary
ELM and its variant based on LDLT decomposition.
Iron and steel industry is one of the pillar industries in the
national economy of China, which is also an industry with high
energy consuming and high emission. The blast furnace (BF) is the
rst and also the most important step in the manufacture of iron
and steel [9,11]. Because of the great amount of iron production,
the tiny improvement of automation technology in the blast
furnace production process will bring huge benets. However,
due to lack of industrial data and the unknown mechanism in BF,
the model is a typical black box system [10]. For a long time, the
operators of BF just speculate the process situation relying on the
experience. The thermal situation in BF is the most typical and
important performance indicator to maintain the normal opera-
tion. In the industrial process, the operators often know the
thermal situation inside the BF by monitoring the temperature
of hot metal [12]. Nowadays, the common used method for
measuring the temperature of hot metal is articial contact
measurement. The operators insert the disposable thermocouple
into the molten iron and then the temperature data will be
uploaded to the host computer [11]. This ancient measurement
operation is very inconvenient in the aspects of discontinuous data
acquisition and unsafe operation. On the other hand, this mea-
surement is costly because of the employment of one-time
thermocouple. In this paper, a soft measurement model to the
temperature of hot metal is established based on 5 performance

http://dx.doi.org/10.1016/j.neucom.2015.04.106
0925-2312/& 2015 Elsevier B.V. All rights reserved.
 H. Zhang et al. / Neurocomputing 174 (2016) 232237 233

indexes and the proposed P-ELM algorithm is employed to train is widely used, where H H T H  1 H T . So
the sampling data and build a reliable feedforward neural
network model. ^ H T H  1 HT T 4
In this paper, a modied ELM algorithm named P-ELM is It is apparent from (4) that one can not obtain correct and
proposed based on PCA technique. Compared with the previous satisfactory result if the matrix HTH is singular despite that some
methods, the proposed P-ELM algorithm can deal with the multi- mathematical softwares like Matlab have the corresponding
collinear problem under the condition that the estimation of methods to deal with the inverse of singular matrix. Next some
output weights is unbiased. The improved P-ELM algorithm shows properties of the solution of ELM algorithm will be discussed. In a
a better performance in dealing with interfered data obtained complex industrial production environment, the data will often be
from complex industrial process. In order to verify the effective- interfered by external noise. Then the model (2) should be
ness of the proposed P-ELM algorithm in real industrial applica- modied as
tion, a soft measurement model subject to hot metal temperature
H T 5
in BF is presented. Finally, simulation results using the real
 
industrial data show that P-ELM algorithm has better stability where white noise is considered here and A N 0; . 2

and generalization performance compared with other famous After adding the interference of external noise, the solution of
feedforward neural network. It is worth mentioning that [13] also ELM will also be modied as
employed PCA theory to the ordinary ELM algorithm. However, in PN~   PN~

1
1 H i H i i i i 1 H i i
T T
this paper, PCA technique is used to deal with the hidden layer ^ HT H H T T i P
N~ T
PN~ T
6
output matrix rather than the original sampling data in [13]. i 1 Hi Hi i 1 Hi Hi
The following sections are organized as follows: Section 2 is the
Next we analyze the result from three aspects: expectation (E),
main part to describe the improved P-ELM algorithm. Some basic
variance (V) and mean square error (MSE). Then one can get
theory about BF will be presented in Section 3 and simulation


results are given in Section 4. Section 5 summarizes the conclusion  E ^
of this paper.

  
2 XN~
2 2 1
 V ^ E ^  E ^ P ~ 2
N T
i 1 Hi Hi i1 i
2. The improved extreme learning machine algorithm
 1 
T
 1 2 2 X N~
1
 MSE ^ E ^  ^  
PN~
In the last few years, ELM algorithm has received very wide N~ N~ H T
H N~
i 1
i
i1 i i
range of applications and development because of its fast train 7
speed and good generalization performance. In this section, a
review of the ordinary ELM algorithm is presented and some where i is the ith eigenvalue of H H [8].
T

properties of the solution will be discussed. And then the improved
P-ELM algorithm is proposed to overcome the restrictions.
2.1. The theory of ordinary ELM algorithm
Suppose there are N arbitrary samples xi ; t i , where xi
xi1 ; xi2 ; ; xin T A Rn denotes an n-dimensional feature of the ith
sample and t i t i1 ; t i2 ; ; t im  A Rm denotes the target vector. The
mathematical model of SLFNs with N~ hidden nodes is as follows:
N~
X N~
X This subsection is the main part of the note. Here the improved
i g i xk i gwi ; bi ; xk t k ; k 1; 2; ; N
i1 i1
where wi and bi are the learning parameters which will be
determined randomly. i is the output weight matrix connecting
the ith hidden and the output nodes. gwi ; bi ; xk is a nonlinear
piecewise continuous function which satises ELM universal
approximation capability theorems.
The above N~ equations can be written in matrix form as
H T
where
0 1
gw1  x1 b1 gwN~  x1 bN~
B C
HW; B; X @ A data distribution. In addition, w2 stands for another direction of
gw1  xN b1 gwN~  xN bN~
NN~
called the hidden layer output matrix. According to the theory of
Least Square, the output weight can be estimated as
^ H T
where H is the MoorePenrose generalized inverse of H [19].
There are several methods to calculate the MoorePenrose gen-
eralized inverse. Here singular value decomposition (SVD) method
As we all know, the learning parameters of hidden nodes in the
model of ELM algorithm are generated randomly without human
tuning. In most case, the number of hidden nodes is far less than
~
that of the samples NN. So H is usually not a square matrix and
H T H may not always be nonsingular. That is to say when
HT H is
multicollinear,

 some eigenvalues will tend to zero, while V ^ and
^
MSE will become larger. And the solution is not convincing.
2.2. The introduction for improved P-ELM algorithm
1 P-ELM algorithm is presented to overcome the multicollinear
problem using principle component analysis (PCA) technique.
PCA is a useful statistical technique that has found application in
elds such as face recognition and image compression, and is a
common technique for nding patterns in high dimension data
[14,16]. For a series of data with high dimension, PCA is a powerful
tool for identifying patterns and expressing the data in such a way
as to highlight their similarities and differences by reducing the
number of dimensions [15].
2 Fig. 1 presents the distribution of sample data with two
dimensions where we can see that most of the data distribute
along the direction of w1 . Here w1 is called the rst principle
component direction which can character the main information of
the second principle component which means less important
information like external disturbance. The main purpose of the
theory of PCA is to represent the sample data using the main
principle components. In the theory of mathematical statistics, the
rst principle component direction is the one with maximum
3
variance of the data distribution [16]. So the sort of principle
components is based on the size of variances of sample data in
different directions. Next a brief theoretical derivation of PCA is
presented [17].
234 H. Zhang et al. / Neurocomputing 174 (2016) 232237

x2

w2
w1

x1
Fig. 1. The distribution of sample data.

Fig. 2. The model of P-ELM algorithm.

mn
For a set of real data H A R where with m samples and n
indicators, we want to nd some principle components to present
the information of sample data as more as possible. So algorithm can get rid of the multicollinear problem after the PCA
transform and the calculation of inverse of H 0 is reasonable and
T
T HW 8
less time-consuming. Then the output estimated weight ^ will be
where T is the principle matrix after reducing the dimension of modied as
ordinary H and W is the unit orthogonal transformation matrix.
1
Considering the theory of maximum variance, the problem can ^ H 0 T H0 H0 T
T
16
be transformed into
max covT In fact, the improved P-ELM algorithm equals to adding an
T additional layer to the model of ordinary ELM like in Fig. 2. PCA
s:t: W W I 9
technique is employed to transform H to an orthogonal matrix H 0
Then it equals to with lower dimension. According to the mathematical statistical
theory, it means no less of information in hidden layer matrix if
max hHW; HW i
one chooses the transformation matric G consisted with the
s:t: WT W I 10 eigenvectors of all the nonzero eigenvalues. Additionally, in most
Applying Lagrangian algorithm, one can get case, one can select the number of principle components based on
the following cumulative contribution rate (CCR) equation:
L W T H T HW  W T nW  I 0 11 Pp
i
where is Lagrangian parameter. CCR Pi~ 1 Z 17
Then
N
i1 i
8
> L where is the articial threshold like 90% and p is the number of
>
< 2H T HW 2W 0
W selected principle components [16].
12
>
> L Then recalculate the performance indicators under the situa-
: W T nW  1 0
tion of selecting all the principle components:


So one can get  E ^
H T HW W 13
  
2 X
2 2 p
1 p 2
where we can see that is a diagonal matrix composed of the  V ^ E ^  E ^ Pp 2 r
T
i 1 Hi Hi i1 i
min
eigenvalues of H T H and W is the transform matrix consisted with

 1 
T
 1 2
the corresponding eigenvectors and the columns of W are the  MSE ^ E ^  ^   Pp T
p p
principle components of the sample data. Additionally, the sort of i 1 Hi Hi
importance of principle components is based on the values of
2 Xp
1 2
eigenvalues of H T H. r 18
p i 1 i min
Next the technique of PCA is applied in the calculation of ELM
algorithm. For the hidden layer output matrix H A RN  N~ with N~ As we can see from the above three indicators, P-ELM algo-
hidden nodes rithm can deal with the multicollinear problem under the premise
H 0 HG 14 of unbiased estimation. Compared with our previous work in [8],
N~ p this improvement is more reasonable and accurate. In addition, it
where G : R is the transform matrix consisting with the cor-
is easy and timesaving to calculate the inverse of H 0 in view of the
responding eigenvectors of nonzero eigenvalues of H T H and H 0 :
Eq. (15) . Then the improved P-ELM algorithm can be summarized
RNp is the new hidden layer output matrix with lower dimension.
as following:
p is the new dimension and also equals to the number of nonzero
eigenvalues. After the transform, one can get
2 3 Step 1: Assign arbitrary learning parameters W and B.
1 Step 2: Calculate the hidden layer output matrix H.
6 7
H 0 H 0 GT H T HG 4
T
5 15 Step 3: Select the number of principle components for H and
p calculate the transformation matrix G.
pp 0
Step 4: Calculate the additional hidden layer

output  1 matrix H .
where i a 0, i 1; 2; ; p. It appears that the improved ELM ^
Step 5: Calculate the output weights: H H 0T 0 0T
H T.
 H. Zhang et al. / Neurocomputing 174 (2016) 232237 235

Fig. 3. The iron blast furnace.

3. Blast furnace and the soft measurement model for Table 1

temperature of hot metal The distribution of sampling data.

No. Variable Mean Minimum Maximum Standard deviation

The purpose of a blast furnace is to convert iron oxides into liquid
iron called hot metal through chemical and physical methods. The input
BF is a huge reaction vessel and it is costly to start and stop a 1 Blast volume 4127.2 1225.8 4301.9 409.7
furnace. Thus the BF will operate for several years continuously once 2 Bast temperature 1179.5 1125.1 1273.5 7.5
3 Blast pressure 335.2  0.18 362.2 48.8
started, while just some short pauses are needed for maintenance.
4 Furnace top pressure 185.3  0.02 224.3 28.7
The process of iron-making in a BF is presented in Fig. 3. The iron 5 Oxygen enrichment 71.6 0 85.6 15.5
ore, coke and limestone are poured into the BF from the top in the
output
way of bell-less feeding system. The layers of iron ore and coke are 1 Temperature 1178.7 799.5 1599.0 345.2
distributed alternately and it takes about 8 h for the raw materials to
descent to the bottom of the furnace. The iron ore contains a wealth
of iron oxides and it is dumped into the BF in the form of raw ore, In order to verify the effectiveness and robust of the proposed P-
pallets or sinter. After a series of reduction reactions, the iron ore is ELM algorithm, we employ P-ELM in two popular datasets: Abalone
transformed into liquid iron drained from the iron notch in the (regression) and Wine (classication). More details can be referred
bottom of the BF at regular intervals. The physical and chemical in the appendix and are omitted here for sake of brevity.
reactions in the BF occur under high temperature condition. So the
preheated gas is blown into the bottom and then ascends to the top 4.1. Experimental preparation
during several seconds to support high-temperature environment.
There are some gas outlet ports at the top of BF for the cooled The experimental data are obtained from NO.2 BF in Wuhan Iron
furnace gas [9,18]. And the chemical reactions during iron-making and Steel Company. All the data have been de-noising processing
process are also shown in Fig. 3. before entering the soft measurement model. From Table 1, one can
In the production process of iron-making in BF, enough and see not only dramatic changes often happen in the data from one
stable thermal situation is a prerequisite for steady running. It is index, but also furthermore the big differences exit among the ve
the most effective and direct way to monitor the furnace status indexes. So all the data must be standardized. Taking into account
information via the temperature of hot metal. Need to add that the special nature of the steel industry, it can satisfy the need of
there are three iron notches taking turn to work in the bottom of industrial production to sample one set of data per minute. There
BF. In this paper, the soft sensor technique is employed to detect are two types of activation functions taken into consideration:
the hot metal temperature and a neural network model based on (1) the sigmoidal additive activation function Ga; b; x 1=1
ELM algorithm is established. The inputs are 5 variables related exp  a  x b, where the input weights and biases are randomly
with the hot metal temperature closely, while temperature is the generated from the range
 1; 1; (2)
 the Gaussian RBF activation
only output. Table 1 lists the information of the corresponding function Ga; b; x exp  bJ x a J 2 , where the centers are ran-
neural network model data. Fig. 4 is the distribution of the domly generated from the range  1; 1, and the impact widths b
experimental data over a period of time. are chosen from the range 0; 1. Additionally, all the experiments
have been conducted in Matlab 7.8.0 running on a desktop PC with
AMD Athlon II X2 250 processor, 3.00 GHz CPU and 2G RAM. All the
4. Simulation results simulation experiments are carried out over 50 trials and the
average values are adopted.
This section presents the simulation results subject to the
proposed P-ELM algorithm applied in the soft measurement model 4.2. Simulation results
for hot metal temperature in BF. In addition, some comparative
experiments are carried out in order to illustrate the better This subsection presents some primary simulation results.
generalization performance and stability of P-ELM algorithm. Fig. 5 is the graphical representation for the regression of hot
236 H. Zhang et al. / Neurocomputing 174 (2016) 232237

355 12004250
76 190
350
74
1195
185 345
72

70 340 11904200
180
68 335
1185
66 175
330

64
325 11804150
170
62
320
60
165 1175
315
58

56 160 31011704100
20 40 60 80 100 120 140 160 180 200
time

Fig. 4. The distribution of sampling data. (From top to bottom: (a) Furnace top pressure. (b) Blast pressure. (c) Blast volume. (d) Blast temperature. (e) Oxygen enrichment.)

metal temperature applying the proposed P-ELM algorithm. 500 1600

sets of sampling data are employed in this experiment. When the 1500
hot iron just ow out, the temperature is relative low (about 1400 The third temperature
measurement
700 1C) because of the cool iron groove. Then the temperature will 1300
Temperature

1# iron notch
gradually rise until about 1500 1C. One can see that the proposed 2# iron notch 3# iron notch
1200
P-ELM algorithm can well track the variation tendency of hot The second temperature
1100
metal temperature. In addition, the performance of the proposed measurement
P-ELM algorithm with sigmoid function is more stable than with 1000
RBF
RBF hidden nodes. 900
Sigmoid
For the continuous production process in BF, it is very essential 800
The first temperature
to realize online condition monitoring, which can provide the real- 700 measurement
time operating condition information to the operators. The pro- 2 4 6 8 10 12 14 16
posed P-ELM algorithm is applied in continuous measurement for
Time
the hot metal temperature through batch learning. We employ P-
ELM algorithm to verify 10 batches of data. Then Fig. 6 shows the Fig. 5. The regression results for hot metal temperature.
simulation results in the aspect of batch error using the corre-
sponding ELM algorithm with sigmoidal function, where one can
60
see that P-ELM algorithm has less error compared with other ELM ELM
algorithm like OP-ELM and R-ELM. P-ELM algorithm can restrain OPELM
the multicollinear problem well and present more accurate pre- 50 RELM
dictions for hot metal temperature. PELM
Temperature error

On-line monitoring requires not only accurate measurement,

but also satisfactory response time. Table 2 lists the comparison
results in terms of some performance indicators. In this simulation
experiment, the number of hidden nodes in P-ELM algorithm is
reduced to 58 from 80. Obviously, it can avoid the inuence of 30
multicollinear problem in calculation within shorter training time.
The proposed P-ELM algorithm runs more stably in view of SD and
CV indexes. The theory of PCA is simple and convenient in
implement and P-ELM algorithm is very easy in programming.
Comparing with other common feedforward neural network (BP
and SVM), ordinary ELM and its variants (OP-ELM and R-ELM), the 10
1 2 3 4 5 6 7 8 9 10
proposed P-ELM has better generalization performance and more
Batch Data
stability in dealing with industrial data.
Fig. 6. Comparison results of different ELM algorithms.

5. Conclusions
In practical applications, ELM algorithm encounters an important
constrain for the multicollinear problem in calculation of output
weight matrix. In this paper, an improved ELM algorithm named P-
ELM is proposed based on PCA theory. Through reducing the dimen-
sion of hidden layer output matrix (H) without loss of any statistic
information, the newly generated hidden layer output matrix (H 0 ) has
full rank of the columns. This approach can not only solve the
multicollinear problem but also improve training speed compared
with other ELM algorithm. At last, we establish the soft measurement
model for the temperature of hot metal in iron-making process of the
BF in order to verify the effectiveness of P-ELM algorithm. The
experimental results show that P-ELM algorithm has more stability
in handling the contaminated industrial data and learning speed can
meet the on-line implementation requirements in the industrial eld.
 H. Zhang et al. / Neurocomputing 174 (2016) 232237 237

Table 2 [3] C. Pan, D.S. Park, Y. Yang, H.M. Yoo, Leukocyte image segmentation by visual
Simulation results. attention and extreme learning machine, Neural Comput. Appl. 21 (6) (2012)
12171227.
Algorithms Nodes MSE Training Testing SD CV [4] P.K. Wang, Z.X. Yang, C.M. Vong, J.H. Zhong, Real-time fault diagnosis for gas
time time turbine generator systems using extreme learning machine, Neurocomputing
Training Testing 128 (2014) 249257.
[5] X.F. Hu, Z. Zhao, S. Wang, F.L. Wang, D.K. He, S.K. Wu, Multi-stage extreme
learning machine for fault diagnosis on hydraulic tube tester, Neural Comput.
BP 80 0.1851 0.1239 5.1821 0.0310 0.0354 0.0501
Appl. 17 (4) (2008) 399403.
SVM - 0.0998 0.0979 0.3940 0.0198 0.0305 0.0436
[6] H.G. Zhang, S. Zhang, Y.X. Yin, An improved ELM algorithm based on EM-ELM
ELM 80 0.1015 0.0998 0.0055 0.0011 0.0344 0.0491 and ridge regression, Lecture Notes in Computer Science 8261 (2013) 756763.
OP-ELM 65 0.0995 0.1003 0.0804 0.0030 0.0213 0.0304 [7] G.B. Huang, H. Zhou, X. Ding, R. Zhang, Extreme learning machine for
P-ELM 80-58 0.0990 0.0967 0.0122 0.0011 0.0201 0.0287 regression and multiclass classication, IEEE Trans. Syst. Man Cybern.-Part B
Cybern. 42 (2) (2012) 513529.
[8] H.G. Zhang, S. Zhang, Y.X. Yin, A novel improved ELM algorithm for a real
industrial application, Math. Probl. Eng. (2014) http://dx.doi.org/10.1155/2014/
Table 3 824765.
The distribution of simulation data. [9] V.R. Radhakrishnan, K.M. Ram, Mathematical model for predictive control of
the bell-less top charging system of a blast furnace, J. Process Control 11 (5)
Dataset #Attributes #Classes #Training data #Testing data (2001) 565586.
[10] V.R. Radhakrishnan, A.R. Mohamed, Neural networks for the identication and
control of blast furnace hot metal quality, J. Process Control 10 (6) (2000)
Abalone 8 3000 1177
509524.
Wine 13 3 120 58
[11] G.M. Cui, J. Li, Y. Zhang, Z.D. Li, X. Ma, Prediction Modeling Study for Blast
Furnace Hot Metal Temperature Based on T-S Fuzzy Neural Network Model,
Iron Steel 48 (1) (2013) 1115.
Table 4 [12] P.D. Burk, J.M. Burgess, A coupled gas and solid ow heat transfer & chemical
Simulation results (sigmoid hidden nodes) reaction model, in: Iron Making Conference, Chicago, 2006, pp. 773781.
[13] A. Castao, F. Fernandez-Navarro, C. Hervas-Martnez, PCA-ELM: a robust and
pruned extreme learning machine approach based on principal component
Dataset Algorithm #Nodes Training time (s) RMSE or accuracy
analysis, Neural Process. Lett. 37 (3) (2013) 377392.
[14] M. Kirby, L. Sirovich, application of the Karhunen-Loeve procedure for the
Training Testing characterization of human faces, IEEE Trans. Pattern Anal. Mach. Intell. 12 (1)
(1990) 103108.
Abalone ELM 50 0.0978 0.0767 0.0723 [15] R.P. Good, D. Kost, G.A. Cherry, Introducing a unied PCA algorithm for model
OP-ELM 50 0.1532 0.0699 0.0711 size reduction, IEEE Trans. Semicond. Manuf. 23 (2) (2010) 201209.
R-ELM 50 0.1329 0.0721 0.0719 [16] L. Yan., A PCA-based PCM data analyzing method for diagnosing process
P-ELM 50 0.1321 0.0701 0.0700 failures, IEEE Trans. Semicond. Manuf. 19 (4) (2006) 404410.
Wine ELM 30 0.0022 99.98% 95.70% [17] I.T. Jolliffe, Principal Component Analysis, Second Edition, Springer-Verlag,
OP-ELM 30 0.0030 99.95% 96.01% New York, 2002, ISBN: 978-0-387-22440-4.
R-ELM 30 0.0031 99.94% 95.29% [18] X.L. Tang, L. Zhang, X.D. Hu., The support vector regression based on the chaos
P-ELM 30 0.0025 99.96% 97.02% particle swarm optimization algorithm for the prediction of silicon content in
hot metal, Control Theory Appl. 26 (8) (2009) 838842.
[19] D. Serre, Matrices: Theory and Applications, Springer, New York, 2002.

Acknowledgments
Haigang Zhang received the B.S. and M.S. degree in
Electrical Engineering from University of Science and
This work has been supported by the National Natural Science Technology Beijing in 2009 and 2011 respectively. Now
Foundation of China (NSFC Grant no. 61333002) and Beijing he is a Ph.D. candidate of control systems. His research
interests include neural networks and its application to
Natural Science Foundation (Grant no. 4132065). control of the world.

Appendix

In the appendix, the proposed P-ELM algorithm is employed in

two popular datasets downloaded from UCI websites (http://archive.
ics.uci.edu/ml/). The benchmark problems are described in Table 3.
The abalone dataset is normalized into the range 0; 1, while we
normalize the wine attributes into the range  1; 1. The sigmoidal
additive activation function Gai ; bi ; xi 1=1 exp  ai  xi bi :
is employed and all the experiments are carried out 50 trials, and
all the performance indicators are calculated by the average 50
conclusions.
We compare the proposed P-ELM algorithm with the ordinary ELM
and its variants (OP-ELM and R-ELM). Table 4 is the simulation results
with sigmoid hidden nodes, where one can see that P-ELM shows the
best generalization performance with relatively little training time.
References
[1] G.B. Huang, Q.Y. Zhu, C.K. Siew, Extreme learning machine: theory and
applications, Neurocomputing 70 (13) (2006) 489501.
[2] N.Y. Liang, G.B. Huang, P. Saratchandran, N. Sundararajan, A fast and accurate
online sequential learning algorithm for feedforward networks, IEEE Trans.
Neural Netw. 17 (6) (2006) 14111423.
Yixin Yin received the Ph.D. degree in Electrical Engi-
neering from University of Science and Technology
Beijing in 2002. He is full professor of the school of
automation and electrical engineering in the University
of Science and Technology Beijing. His research inter-
ests include control theory and its applications.
Sen Zhang received the Ph.D. degree in Electrical
Engineering from Nanyang Technological University in
2005. She has been working as postdoctoral research
fellow in National University of Singapore and lecturer
in charge in Singapore Polytechnic. She is currently an
associate professor of the school of automation and
electrical engineering in the University of Science and
Technology Beijing. Her research interests include ELM,
target tracking and estimation theory.