Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO
This paper combine the improved PSO algorithm (Analysis of Particle Swarm Optimization Algorithm) with the BP neural network for prediction of Silicon content in hot metal. Firstly, the varying visual mechanism is drawing into the standard PSO through changing the neighbor structure dynamically with...
Saved in:
| Published in: | Functional Materials |
|---|---|
| Date: | 2016 |
| Main Authors: | , |
| Format: | Article |
| Language: | English |
| Published: |
НТК «Інститут монокристалів» НАН України
2016
|
| Subjects: | |
| Online Access: | https://nasplib.isofts.kiev.ua/handle/123456789/121413 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| Cite this: | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO / Yang Kai, Zhijun He // Functional Materials. — 2016. — Т. 23, № 3. — С. 463-467. — Бібліогр.: 8 назв. — англ. |
Institution
Digital Library of Periodicals of National Academy of Sciences of Ukraine| _version_ | 1859819716871192576 |
|---|---|
| author | Yang Kai Zhijun He |
| author_facet | Yang Kai Zhijun He |
| citation_txt | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO / Yang Kai, Zhijun He // Functional Materials. — 2016. — Т. 23, № 3. — С. 463-467. — Бібліогр.: 8 назв. — англ. |
| collection | DSpace DC |
| container_title | Functional Materials |
| description | This paper combine the improved PSO algorithm (Analysis of Particle Swarm Optimization Algorithm) with the BP neural network for prediction of Silicon content in hot metal. Firstly, the varying visual mechanism is drawing into the standard PSO through changing the neighbor structure dynamically with each particles, in order to enhance the local and global searching ability in particle swarm. Afterwards, the improved algorithm is used to optimize the weights and threshold of BP neural network to avoid falling into local extremum. Finally, the prediction model of Si content in hot metal is built based on BP network optimized by Variable neighborhood PSO. The average relative error of the prediction model is 6.7% based on the data from blast furnace.
|
| first_indexed | 2025-12-07T15:24:54Z |
| format | Article |
| fulltext |
Functional materials, 23, 3, 2016 463
ISSN 1027-5495. Functional Materials, 23, No.3 (2016), p. 463-467
doi:http://dx.doi.org/10.15407/fm23.03.463 © 2016 — STC “Institute for Single Crystals”
Target prediction in blast furnace based on BP
network optimized by variable neighborhood PSO
Yang Kai1,2, Zhijun He1
1 School of materials and metallurgy, University of Science and Technology
LiaoNing,
LiaoNing 114051, China,
2 School of Software, University of Science and Technology LiaoNing,
LiaoNing 114051, China
Received June 22, 2016.
This paper combine the improved PSO algorithm ��nal�sis o�� Particle S�arm Optimi�ation ��nal�sis o�� Particle S�arm Optimi�ation
�lgorithm)� �ith the �P ne�ral net�or�� ��or prediction o�� Silicon content in hot metal�� �irstl��� �ith the �P ne�ral net�or�� ��or prediction o�� Silicon content in hot metal�� �irstl���
the var�ing vis�al mechanism is dra�ing into the standard PSO thro�gh changing the neighbor
str�ct�re d�namicall� �ith each particles�� in order to enhance the local and global searching
abilit� in particle s�arm�� ���ter�ards�� the improved algorithm is �sed to optimi�e the �eights
and threshold o�� �P ne�ral net�or�� to avoid ��alling into local extrem�m�� �inall��� the prediction
model o�� Si content in hot metal is b�ilt based on �P net�or�� optimi�ed b� Variable neighbor-
hood PSO�� The average relative error o�� the prediction model is 6��7% based on the data ��rom
blast ���rnace��
Keywords: particle s�arm optimi�ation�� ne�ral net�or���� silicon content�� prediction��particle s�arm optimi�ation�� ne�ral net�or���� silicon content�� prediction���� ne�ral net�or���� silicon content�� prediction��ne�ral net�or���� silicon content�� prediction���� silicon content�� prediction�� silicon content�� prediction���� prediction�� prediction����
Предлагается улучшенный алгоритм PSO для прогнозирования содержания кремния в
горячем металле. Средняя относительная погрешность модели прогнозирования составляет
6,7% на основе данных из доменной печи.
Прогнозування змісту кремнію в гарячому металі за допомогою алгоритму
PSO. Ян Кай, Шун Хі
Пропонується покращений алгоритм PSO для прогнозування вмісту кремнію в гарячому
металі. Середня відносна похибка моделі прогнозування становить 6,7% на основі даних з
доменної печі.
1. ������������������������
The silicon content in blast ���rnace hot met-
al is an important index to meas�re the q�alit�
o�� pig iron and metall�rgical technolog��� The
change of silicon content directly reflects the
stabilit� o�� the process[1��2]�� There��ore�� real-
i�ing the silicon content in hot metal and it’s
var�ing tendenc� in real time�� and ma��ing ac-
curate forecasts, have significances to conduct
operating ��or temperat�re o�� blast ���rnace�� re-
d�ce ��ocal ratio and the cost o�� pig iron�� red�ce
fluctuation of furnace condition, and ultimately
achieve energ� saving and red�cing cons�mp-
tion��
�eca�se the traditional prediction model
based on the prod�ction mechanism o�� hot
metal has certain limitations�� Recentl��� Man�
scholars use artificial intelligence methods
to predict the q�alit� o�� hot metal [3-6]�� This
paper combine the improved PSO algorithm
�ith the �P ne�ral net�or�� ��or prediction o��
Silicon content in hot metal�� The concept that
each particles has certain visual field is firstly
proposed in this paper�� �ll the particles in the
visual field of a given particle form a neighbor-
464 Functional materials, 23, 3, 2016
Yang Kai, Zhijun He / Target prediction in blast furnace based on ...
hood structure of the particle. The visual field
o�� each particle is changing along �ith itera-
tion�� leading to the var�ing neighborhood o��
particles�� in order to improve the abilit� o�� the
particle swarm to find the optimal solution.
On this basis�� the improved algorithm is �sed
to optimi�e the �eights and thresholds in �P
ne�ral net�or�� to avoid ��alling into local extre-
m�m�� Later the prediction model based on �P
net�or�� optimi�ed b� Variable neighborhood
PSO is to be trained and validated b� real data
in a steel compan���
2. PSO alg����hm
a�� ��’s �mp��veme��
2.1 The s�a��a�� PSO
A group of n particles is flying at a certain
speed in D-dimensional searching space�� Each
particle is a potential sol�tion ��or the optimi-
�ation problem to be solved�� Speed determines
their flight distance and direction. All the par-
ticles have a fitness value that is determined
b� the ���nction o�� the optimi�ation�� Particles
�pdate their position and speed b� ��ollo�ing
two extreme values in the process of flight,
which separately represent their own flight ex-
perience and the swarm flight experience. One
extreme val�e is the optimal sol�tion that the
particle itsel�� has ��o�nd so ��ar�� called the in-
divid�al extreme val�e�� denoted b� pbest�� The
other extreme val�e is the optimal sol�tion that
the �hole pop�lation has ��o�nd so ��ar�� called
the global extreme val�e�� denoted b� gbest�� The
�pdate ��orm�la ��or particle position and veloc-
it� is:
v wv c r pbest xid
k
id
k
id id
k+ = + - +1
1 1( )
c r gbest xgd id
k
2 2 ( )- �1)�
x x vid
k
id
k
id
k+ += +1 1 �2)�
Where i n= 1 2, , , �� n is the si�e o�� the
s�arm; d = 1��2��…��D�� ��D is the n�mber o�� dimen-
sions�� k is the n�mber o�� iterations; c c1 2, are t�o
positive constants�� called cognitive and social
parameter respectivel�; r r1 2, are t�o random
���nctions in the range [0��1]�� W is called inertia
�eight�� �sed to coordinate the global and local
optimi�ation abilit� o�� PSO algorithm�� W �s�-
all� decreases linearl� �ith the increase o�� the
n�mber o�� iterations�� In this paper�� the �pdat-
ing ��orm�la o�� the inertia �eight is:
w
w w w
=
= - ´ +
-
( )max min min
maxiter curiter
maxiter
�3)�
�here w wmax min, are the maxim�m and mini-
m�m val�es respectivel� o�� the inertia �eight
��� maxiter��c�riter represent the maxim�m and
c�rrent iterations n�mber respectivel���
2.2 Va��able �e�ghb��h��� s�������e
Artificial fish swarm algorithm(AFSA) is
firstly proposed by Li Xiaolei in 2002[7]. In
AFSA, each artificial fish has vision and can
consider a step ��or�ard to the direction better
than c�rrent position�� The vision mechanism o��
artificial fish is firstly lead into particle swarm
optimi�ation a��ter inspired b� the optimi�ation
process in ��S��� In this paper�� each particle is
assigned a vision �ith a limited range�� denoted
b� vis�al�� The other particles �ithin vis�al o��
a given particle ��orm the neighborhood o�� it�� In
the standard PSO�� ever� particle is a����ected b�
both individ�al and pop�lation experience in
sol�tion space�� In order to promote the capa-
bilities o�� in��ormation exchange and sharing�� a
ne� component is added into the velocit� ��or-
m�la �hich represents the attraction gener-
ated ��rom the neighborhood optim�m�� The ne�
��orm�la is as ��ollo�:
v wv c r pbest xid
k
id
k
id id
k+ = + - +1
1 1( ( )α
( )( )) ( )1 2 2- - + -α nbest x c r gbest xid id
k
gd id
k �4)�
Where α is a random n�mber in the range
[0��1] and nbestid represents the best neighbo�r-
hood optim�m ��o�nd b� particle xi: so ��ar�� The
�pdate ��orm�la o�� neighbo�rhood optim�m
�ith xi is as ��ollo��� ��or ever� particle xj:
I��
x x Visualj i- < fitness x fitness nbestj i( ) ( )<
then
nbest xi j= �5)�
In order to improve the local and global
searching abilit� in standard PSO�� In this pa-
�ig�� 1�� Vis�al change Process
Functional materials, 23, 3, 2016 465
Yang Kai, Zhijun He / Target prediction in blast furnace based on ...
per�� �e propose the idea that all particles have
gradient vision. The field change of particle vi-
sion is not happened in each iteration�� b�t in
certain q�antit� interations and �as the lad-
der-li��e distrib�tion�� The advantage o�� this de-
sign is that particles can search optimal val�e
in relative fixation neighborhood field and avoid
missing a better search path d�e to ��req�entl�
shi��t �ith vision�� In the beginning iteration��
the neighborhood o�� a particle is itsel���� When
the iterative process is end�� the neighborhood
o�� a particle is all the s�arm�� The vision o�� a
particle is defined as:
Visual fix curiter step= ´( , )δ �6)�
Where c�riter is the c�rrent iterative n�m-
ber�� δ is interval times be��ore next change o��
vision, fix is a reminder function, step repre-
sents each growth of the visual field. As shown
in Fig.(1), the neighborhood field is solid circle
�hen the vision o�� a particle is eq�al to Vis�-
al1�� ���ter iterations�� the vision o�� the particle
is �pdate to Vis�al2 and the corresponding
neighborhood filed is dashed circle. It satisfy
that Vis�al2 = Vis�al1+step��
2.3 M��a���� �pe�a���
If the optimal fitness value of the algorithm
does not change thro�gh certain times o�� s�c-
cessive iterations�� the algorithm is li��el� to
converge to a local optim�m o�� the problem��
Paper [14] introd�ce reinitiali�ation �ill occ�r
aimed at arbitrar� one-dimension o�� the c�r-
rent optim�m�� It is �sed to change the c�rrent
searching trajector� so that particles can go o�t
o�� the local optima and improve the acc�rac� o��
optimi�ation��
This paper applies s�ch e����ective and con-
venient operator to the improved PSO���� Let
Rmax be a predefined constant that represents
the allo�ed maxim�m iteration n�mber�� That
is to say, if the optimal fitness value of the al-
gorithm does not change thro�gh Rmax times
o�� s�ccessive iterations�� �e choose a arbitrar�
dimension and replace it b� a random n�mber
�ithin [-1��1] so as to enhance the exploration
abilit� ��or particles and escape ��rom the local
optim�m ��or the algorithm��
3.Op��m�za���� f�� BP �e��al �e�w��k
3.1 BP �e��al �e�w��k
�P ne�ral net�or�� is a m�ltila�er ��or�ard
net�or�� ��or �nidirectional propagation �ith
three la�ers or more than three la�ers�� Incl�d-
ing the inp�t la�er�� hidden la�er and o�tp�t
la�er and it is the most �idel� �sed ne�ral
net�or���� In the �se o�� �P ne�ral net�or���� the
ne�ral net�or�� m�st be given a set o�� learning
samples in advance�� Each ne�ron in the ne�-
ral net�or�� contains a threshold and there is
also �eights bet�een ne�rons�� The inp�t pa-
rameters are trans��erred ��rom the inp�t la�er
thro�gh the hidden la�er to the o�tp�t la�er
�nder the action o�� the corresponding �eights
and thresholds as �ell as the excitation ���nc-
tion�� �P ne�ral net�or�� adj�st the connec-
tion �eights vol�ntaril� thro�gh the train-
ing according to the minim�m error criterion
bet�een the inp�t and o�tp�t data�� Thro�gh
contin�o�sl� iteration�� the nonlinear mapping
between input and output is finally obtained.
��t there are some disadvantages o�� �P
ne�ral net�or���� s�ch as it is eas� to ��all into
local extreme val�e�� slo� convergence rate��
randomicit� o�� parameters selection�� �hich can
easil� lead to the instabilit� o�� the net�or���� In
vie� o�� the shortcomings o�� �P ne�ral net�or����
this paper �se the improved particle optimi�a-
tion algorithm to determine the �eights and
threshods�� and bring abo�t prediction o�� silicon
content in hot metal�� �ig�� 2 is a ��orcasting mod-
el o�� �P ne�ral net�or����
3.2 E������g f�� we�gh�s a��
�h�esh�l�s
This paper �se the variable neighborhood
particle s�arm optimi�ation algorithm�VNPSO)�
to determine the �eights and threshods o�� �P
ne�ral net�or���� In the trainning process�� all the
�eights and threshods in �P ne�ral net�or��
are encoded as real n�mber string and repre-
sented b� a particle�� etc�� x x x xi i i iL= ( , , , )1 2 ��
i N= 1 2, , , �� xij is initiali�ed to the val�e be-
t�een [-1��1]�� L is defined as:
L In Hi Hi Ou Hi Ou= ´ + ´ + + �7)�
Where In�� Hi and O� are the n�mber o��
inp�t la�er�� hidden la�er and o�tp�t la�er re-
spectivel���
In the iterations�� each particle in the s�arm
is decomposed to �eights and thresholds in
ne�ral net�or�� and bro�ght into the ne�ral
network model to compute fitness of the par-
ticle repeatedl��� �ntil achieve the maxim�m
n�mber o�� iterations or obtain the permissible
�ig�� 2�� �orcasting model o�� �P ne�ral net�or��
466 Functional materials, 23, 3, 2016
Yang Kai, Zhijun He / Target prediction in blast furnace based on ...
precision�� The optimal particle is the training
res�lt o�� �P ne�ral net�or����
The individual fitness fitnessi in s�arm is
defined as the mean square error of given sam-
ple:
fitness
y o
Mi
j
M
j j=
-=å 1
2( )
�8)�
Where M is the N�mber o�� samples ��or test-
ing�� are predictive and act�al val�e correspond-
ing to the ith sample respectivel���
3. Op��m�za���� s�eps
The ��ollo�ing steps are given to optimi�e
the �eights o�� �P ne�ral net�or�� b� the vari-
able neighborhood particle s�arm optimi�ation
algorithm:
Step 1: Initiali�e parameters: the pop�la-
tion si�e N�� the maxim�m iteration n�mber
LoopCount, the acceleration coefficients c1�� c2��
the interval ��req�enc� δ �� increment step �� iner-
tia weight w, Rmax, the constriction coefficient
χ, the max velocity vmax ��
Step 2: Initiali�e position and velocit� o��
all the particles randoml� in the D dimension
space��
Step 3: Evaluate the fitness value of each
particle�� and �pdate the global optim�m posi-
tion gbest �� the neighborhood optim�m position
and the individ�al optim�m position pbest ��
Step 4: Update the inertia �eight according
to ��orm�la �3)���
Step 5: Update the vis�al o�� pop�lation ac-
cording to ��orm�la �6)���
Step 6: �or each particle�� �pdate particle
velocit� according ��orm�la �4)��� �pdate particle
position according ��orm�la �2)���
1) If the current fitness value is better than
pbest �� assign the c�rrent position to pbest��
2) If the current best fitness value is better
than the gbest�� assign the c�rrent best position
to gbest��
�or each particle in a given one’s neighbo�r-
hood, if the current fitness value of the particle
is better than the nbest o�� the given one�� assign
the particle’s c�rrent position to nbest corre-
sponding to the given particle��
Step 7: Identi��� �hether the algorithm has
stagnated d�ring s�ccessive Rmax times itera-
tions�� I�� the algorithm has stagnated�� initiali�e
an� one-dimension o�� the optimal position b�
m�tation operator��
Step 8: Repeat Step 4 - 8 �ntil a stop crite-
rion is satisfied or a predefined number of itera-
tions are completed��
4. P�e������� f�� S� ����e�� �� h��
me�al
This paper create a prediction model ��or si
content in hot metal based on �P net�or�� opti-
mi�ed b� Variable neighborhood PSO��Nine pa-
rameters that related to si content in hot metal
�ere selected as inp�t parameters to ne�ral
net�or�� ��rom iron ma��ing in the blast ���rnace��
The� are air vol�me�� hot blast drang�� press�re
di����erence�� ventilating index �� air temperat�re��
top temperat�re�� Ox�gen enrichment�� coal in-
jection amo�nt�� batch n�nber�� The n�mber o��
hidden la�ers is 10 and the o�tp�t parameter
is si content in hot metal�� N�mber o�� particle
s�arm N=60�� The dimension o�� particle s�arm
is 111 and each o�� them represents a �eight or
threshold�� D�e to the di����erence o�� these pa-
rameters�� the inp�t and o�tp�t parameters are
normali�ed in order to ens�re that these ��ac-
tors are in the same position�� the initial range
o�� the �eights is [-1��1]�� The initial parameters
is set as: N w= =60 0 9, . ,max wmin .= 0 4 c1 2=
�� c2 2= �� loopco�nt=200�� δ =10�� step=0��15��
Prod�ction data o�� ever� da�s that collected
from a steel works in China in the first half
o�� �ear 2012 is selected as training data set��
�ig�� 3�� Comparison bet�een predicted and ac-
t�al val�e
�ig�� 4�� Prediction error
Functional materials, 23, 3, 2016 467
Yang Kai, Zhijun He / Target prediction in blast furnace based on ...
containing 174 gro�p data�� � total o�� 29 gro�p
data in J�l� is �sed as a test data set�� The com-
parison bet�een predicted and act�al val�es
is sho�n in �ig�� 3�� The model can predict the
change trend o�� Si content in hot metal�� The av-
erage relative error is 6��7%���� the acc�rac� �as
97% �hen the errors �as limited in the range
o�� ±0 1. and the acc�rac� �as 76% �hen the
errors �as limited in the range o�� ±0 05. �� Pre-
diction error is sho�n in �ig���4)���
6. C���l�s���s
The prediction acc�rac� o�� �P ne�ral net-
�or�� is largel� determined b� its relative
�eights and thresholds�� In this paper�� the idea
o�� variable neighborhood particle s�arm op-
timi�ation algorithm is proposed b� introd�c-
ing the vision mechanism�� Then the improved
algorithm is �sed to optimi�e the �eights and
thresholds o�� �P ne�ral net�or�� so as to avoid
��alling into local extrem�m�� �inall��� this meth-
od is applied to the prediction o�� silicon content
in hot metal. Through training and verification
o�� the act�al prod�ction data o�� a steel �or��s��
the average relative error o�� the prediction
model is 6��7%�� the acc�rac� �as 97% �hen the
errors �as limited in the range o�� ±0 1. and the
acc�rac� �as 76% �hen the errors �as limited
in the range o�� ±0 05. ��
A�k��wle�geme��s
This work was financially supported by
scientific research fund of Liaoning provincial
ed�cation department�L2014118)� and The sci-
entific research special fund of university of sci-
ence and technolog� Liaoning�2015TD03)���
Refe�e��es
1. Bi Xuegong, Mathematical model and computer
control o�� blast ���rnace process[M]�� �eijing: Chi-
na Metall�rgical Ind�str� Press�� 1996:1��
2. Liu Xiangguan, Liu Fang. BF Ironmaking
Process Optimi�ation and Intelligent Control
S�stem[M]�� �eijing: China Metall�rgical In�s-
tr� Press�� 2003��
3. Gao Xu-dong, China Metallurgy�� 24�� 24�� 2014
4�� Wang H�a-qiang�� G� Jin-chen�� J. EFEI Univ.
Technol., 31�� 73�� 2008
5�� Y� Zh�o�ing�� Zheng Tao�� Hebei Metallurg. GY��
3�� 38�� 2015
6. Wang Wen-hui, Liu Xiang-guan, Metallurg. In-
dust.Autom. 38�� 33�� 2014
7. Li Xiao-lei,Shao zhi-jiang, Qian ji-xin Syst.Eng.
Theor.Pract.�� 11��32��2002
8�� Hai-mei Z�� Yong-ping W���� Control and Decision,
25�� 20�� 2010��
|
| id | nasplib_isofts_kiev_ua-123456789-121413 |
| institution | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| issn | 1027-5495 |
| language | English |
| last_indexed | 2025-12-07T15:24:54Z |
| publishDate | 2016 |
| publisher | НТК «Інститут монокристалів» НАН України |
| record_format | dspace |
| spelling | Yang Kai Zhijun He 2017-06-14T09:43:36Z 2017-06-14T09:43:36Z 2016 Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO / Yang Kai, Zhijun He // Functional Materials. — 2016. — Т. 23, № 3. — С. 463-467. — Бібліогр.: 8 назв. — англ. 1027-5495 DOI: dx.doi.org/10.15407/fm23.03.463 https://nasplib.isofts.kiev.ua/handle/123456789/121413 This paper combine the improved PSO algorithm (Analysis of Particle Swarm Optimization Algorithm) with the BP neural network for prediction of Silicon content in hot metal. Firstly, the varying visual mechanism is drawing into the standard PSO through changing the neighbor structure dynamically with each particles, in order to enhance the local and global searching ability in particle swarm. Afterwards, the improved algorithm is used to optimize the weights and threshold of BP neural network to avoid falling into local extremum. Finally, the prediction model of Si content in hot metal is built based on BP network optimized by Variable neighborhood PSO. The average relative error of the prediction model is 6.7% based on the data from blast furnace. en НТК «Інститут монокристалів» НАН України Functional Materials Modeling and simulation Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO Article published earlier |
| spellingShingle | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO Yang Kai Zhijun He Modeling and simulation |
| title | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO |
| title_full | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO |
| title_fullStr | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO |
| title_full_unstemmed | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO |
| title_short | Target prediction in blast furnace based on BP network optimized by variable neighborhood PSO |
| title_sort | target prediction in blast furnace based on bp network optimized by variable neighborhood pso |
| topic | Modeling and simulation |
| topic_facet | Modeling and simulation |
| url | https://nasplib.isofts.kiev.ua/handle/123456789/121413 |
| work_keys_str_mv | AT yangkai targetpredictioninblastfurnacebasedonbpnetworkoptimizedbyvariableneighborhoodpso AT zhijunhe targetpredictioninblastfurnacebasedonbpnetworkoptimizedbyvariableneighborhoodpso |