Energy and Power Engineering, 2009, 90-93
doi:10.4236/epe.2009.12014 Published Online November 2009 (http://www.scirp.org/journal/epe)
Modeling of the Unburned Carbon in Fly Ash
Weiping YAN, Jun LI
Key Laboratory of Condition Monitoring and Control for Power Plant Equipment of Ministry of Education,
North China Electric Power University, Baoding, China
Abstract: Numerical simulation of the content of unburned carbon in fly ash on the 300MW tangentially
pulverized coal fired boiler is performed by the numerical simulation software COALFIRE, which is based
on international advanced TASCFLOW software platform. Firstly, take the result of calculation of number
value as the sample, and then set up the support vector machine model of unburned carbon content on the
boiler. The relative error between the predicted output and measured value is 0.00186%, which proves the
modeling is good for the unburned carbon in fly ash predict.
Keywords: numerical simulation, unburned carbon content, support vector machine
1. Introduction
The unburned carbon in fly ash is the important data
which reflects the combustion efficiency of the coal fired
boiler in thermal power plant, but the content of unburned
carbon in fly ash of the coal fired boiler is influenced by
so many kinds of factors which are difficult to be meas-
ured in real time [1–3].
At present, there are lots of the researches about the
monitor of the content of unburned carbon in fly ash in
real time in domestic and international. Most of them are
concentrated on the technology of the measurement
equipment and software [4–6]. The equipment of the con-
tent of unburned carbon in fly ash has many problems,
such as the block of sample tube, complicated extra
equipment, many influence factors about the precision of
the measurement, high cost of the fabrication and the high
request of maintain, etc. Moreover, the traditional soft
measurement technology is almost all used the artificial
neural network method to build the model, because the
study algorithm of the neural network adopts the experi-
ence risk and minimizes the principle, lacks of quantita-
tive analysis and theory result with complete mechanism.
At the same time, for the reason of it is limited in quantity
to survey samples, it will cause the question of over-
studying and the ability of generalization badly, etc.
Therefore, there is very great limitation exist in the meas-
urement of the content of unburned carbon in fly ash
based on the artificial neural.
As a kind of new statistics learning method, support the
vector machine (Support Vector Machine, SVM) uses the
principle of the structure risk minimize (SRM), which
solved the problem of generalization in the theory of ma-
chine studying effectively. The goal function of the
structure risk minimize (SRM) has suppressed the phe-
nomenon of owed and passed studying effectively, and it
can get the good ability of generalization. The algorithm
is turned into a secondary programming and gets opti-
mum point of global. It has solved local minimum prob-
lem in the neural network. Its topological structure only
relates to support the vector and has reduced the calculat-
ing amount. The computational speed is fast, which is
suitable for online application. The method is suited to
study in little sample. Support vector machine has already
become the important research means between the pat-
tern-recognition and the data excavate field at present
[7,8].
In the limited sample studying system [9], because it
has the wild value in the sample, the statistics study the-
ory of V. N. Vapnik thinks it has caused the reducing of
minimum guaranteed risk value. It also caused low cate-
gory precision in the category questions. Regression will
have the phenomenon of “over fitted” or “over trained”,
and the ability of generalization of the study drops greatly
too. On the condition of training samples is limited, the
selection of training collection influences the ability of
generalization of study machine obviously. That is to say,
the ability of generalization not only relates to the study
algorithm, but also relates to the selection of training col-
lection. When choosing the model sample, the existing
method always uses the unburned carbon in fly ash of the
boiler hot condition as training the sample [4–6]. Because
the combustion operating mode in the pulverized coal
fired boiler is very complicated, involving the course in
many aspects of burning, flowing, conducting heat etc,
receives such as the influence in proportion, nature of the
coal, density and fineness of coal flow, wind velocity,
load of boiler, etc. Additionally, these factors often influ-
ence each other and interweave each other, which have

W. P. YAN ET AL. 91
increased the complexity of the combustion process even
more, thus it causes changeable of the unburned carbon in
fly ash. Adding the restriction of the on site condition, it
is more difficult to get the intact sample of unburned car-
bon in fly ash from the operating mode of the boiler.
Using the method of numerical simulation to calculate
the content of unburned carbon in fly ash of the boiler has
satisfied accuracy, whose variation tendency predicted
can meet the need of combustion adjustment for operation
actually. However, its amount of calculation will con-
sume much time. It will consume a large amount of time
even used for optimize design off line. It makes it diffi-
cult to use in monitoring and optimizing online directly.
To the question of appealing, this paper adopts the
coal-fired pulverized boiler furnace numerical simulation
software COALFIRE, to calculate the content of un-
burned carbon in fly ash off-line under many kinds of
burning operating modes of the boiler. Regard the result
of calculation of number value in advance as the training
sample of the model of vector supporting machine. It
makes the model not only has the precision of the nu-
merical simulation, but also can applies the result in the
online system of content of unburned carbon in fly ash
monitoring and optimizing.
2. Pre-Numerical Calculation
2.1. Object of Simulation
Object of the simulation is the boiler of DG1025/18.2-, Ⅳ
which was made by the Don Fang boiler factory in China.
2.2. The Method of Numerical Simulation
It chooses the calculate area between the bottom of the
furnace hopper and the angle of flame diverter on the top
of the furnace. It takes the width of the furnace as the X
direction, the depth of the furnace as the Y direction, the
height of furnace as the Z direction. Divide 160 unit areas,
508896 mesh altogether. The numerical simulation adopts
the three-dimensional stable state to calculate. Adopt
standard k–ε equation model to simulate the turbulent
gaseous phase of flow; use the particle track model at
random to solve the solid particle phase problem; adopt
P1 radiate models to solve the radiation Conduct heat;
adopt the single step response model to the release of the
coal pulverized volatile matter; have adopted the motive
force model of spreading to the burning of the coke;
adopt the PDF model of conservation scalar to simulate
the non-premixed combustion and the non-cross mesh
SIMPLE method to solve the flowing field of the gaseous
phase. Having identified the convergence by the residual,
all of the relative error of the calculating amounts (for
instance, the relative error that u, v, w, ε, etc.) must be
smaller than the value of 1.0×10-4. Take it as the conver-
gence criterion.
2.3. The Operating Mode of Calculation
Consider the change of unburned carbon in fly ash with
the change of the nature of coal, coal fineness, boiler load,
distribution of secondary air, withdrawal or input of terti-
ary air, etc.
2.4. Comparison and Analysis of Results of the
Unburned Carbon in Fly Ash
The unburned carbon content in fly ash of outlet of fur-
nace can be received through the numerical simulation.
The actual measurement value of unburned carbon also
can get by the online monitoring system of carbon content
of unburned carbon in fly ash which located in front of
the air pre-heater at the same time. The result of calculat-
ing simulation and hot surveying value of unburned car-
bon content in fly ash are as Figure 1 shows. Figure 1
shows that the actual measurement value is higher than
the result of numerical simulation, the greatest relative
error is 0.07% after revising and the minimum relative
error is 0.02%. Considering the error of the surveying
value, the result of calculation of number value and hot
operating mode surveying value are identical better.
3. The Model of Unburned Carbon in Fly
Ash Based on Support Vector Machine
3.1. The Introduce of Support Vector Regression
The basic problem of regression is to find a function
f
F

(F—Union of function) and make the expected
risk function as follow minimum [10–15].
3.2. Unburned Carbon Content in Fly Ash
Model Based on Support Vector Machine
 The choice of input and output parameter of the
model
According to the condition of the pre-numerical simu-
lation in advance, adopt the flowing as parameter of in-
putting, include the boiler load, rate of powder quality,
The value of actual measurement
The result of numerical simulation
Operating mode
Cfh/%
0
7
2468 10 12 1416
6
5
4
3
2
1
0
Figure 1. The unburned carbon content under various oper-
ating conditions

W. P. YAN ET AL.
92
fineness of pulverized coal, wind velocity of secondary
air and tertiary air, tilt angle of burner, net calorific
power, volatile, ash content and moisture content of coal.
The amount of adjustable parameter is 22. Regard carbon
content of flying ash as the parameter of outputting, and
use the supporting vector machine to set up the carbon
content characteristic model of flying ash.
 Selecting of kernel function
Selecting of kernel function influences the regression
analysis of support vector machine obviously, but does
not have a ripe theory about it at present yet [16]. The
paper selected the radial kernel function [17].
 Constitution of sample union and the selecting of
algorithm parameter
The sample union is structured by the result of pre-
calculation in different operating condition of boiler.
Taking the result of 1~2、4~8、10~14 operating condition
as the training sample, the result of 3、9 operates condi-
tion as test sample. Adopting the radial kernel function to
regression analysis, the precision of model ε is 0.0001.
 The result and the analysis
The trained result shows as follows: the relative error
of the output of predicting is 0.00186％. The predicting
output and the relative error of the test sample of operat-
ing condition 3 and 9 show as Table 1. It is indicating
that the output error of support vector machine model is
very small, and its identical degree of data is quite good.
It is proved that this model has gotten the correct corre-
sponding behavior between the input and output. This
proves that the modeling of the unburned carbon in fly
ash based on numerical simulation and support vector
machine algorithm on the 300MW tangentially pulver-
ized coal fired boiler is successful.
4. Conclusions
The unburned carbon content in fly ash of tangentially
pulverized coal is influenced by so many factors. The
factors also influence each other. Therefore, it is difficult
to predict and control the unburned carbon content in fly
ash. By the pre-calculating, the precision result of un-
burned carbon content in fly ash can be gotten. After it
is fixed by the practice measurement in different boiler
operating condition, it can be taken as the training test
Table 1. The comparison between the predict value of sup-
port vector machine model and pre-calculation value of
unburned carbon content in fly ash
Operating
condition
The result of
Numerical
simulation
（％）
The result of
model of
support vector
machine（％）
The absolute
value of relative
error（％）
3 5.18 5.152 0.541
9 5.23 5.214 0.306
sample of the support vector machine model. The simu-
lating result shows that the predict result and the result of
practice measurement of unburned carbon content in fly
ash of boiler concordances are better. It provides the ef-
fective method for applying the numerical simulation
result for the online monitor of boiler burn state parame-
ter, whose character is to get accurate result but consume
enormous time during computational process. If com-
bining the overall situation seeking excellent algorithm,
it can make the boiler reach optimum operating mode,
and improve security and economy of the boiler.
REFERENCES
[1] M. H. Fan and R. C. Brown, “Precision and accuracy of
photo acoustic measurements of unburned carbon in fly
ash [J],” Fuel, Vol. 80, No. 11, pp. 1545–1554, 2001.
[2] K. Styszko-Grochowiak, J. Golas, H. Jankowski, et al.,
“Characterization of the coal fly ash for the purpose of
improvement of industrial on-line measurement of un-
burned carbon content [J],” Fuel, Vol. 83, No. 13, pp.
1847–1853, 2004.
[3] A. K. Ouazzane, J. L. Castagner, A. R. Jones, et a1.,
“Design of an optical instrument to measure the carbon
content of fly ash [J],” Fuel, Vol. 81, No. 15, pp.
1907–1911, 2002.
[4] H. Zhou, H. B. Zhu, T. H. Zeng, et a1., “Artificial neural
network modelling on the unburned carbon in fly ash
from utility boilers [J],” Proceedings of the CSEE, Vol.
22, No. 6, pp. 96–100, 2002. (in Chinese)
[5] X. T. Fang and N. Y. Ye, “A system forpredicting the
anburned carbon of the fly ash from utility boilers based
on BP artificial neural netwoks [J],” Jouunal of Huazhong
Uruversity of Science & Technology, Nature Science
Edition, Vol. 31, No. 12, pp. 75–77, 2003. (in Chinese)
[6] M. Sebastia,I. F. dez Olmo, and Angel Irabien, “Neural
network prediction of unconfined compressive strength of
coal fly ash cement mixtures [J],” Cement and Concrete
Research, Vol. 33, No. 8, pp. 137–145, 2003.
[7] G. Y. Zhang and J. Zhang, “Fuzzy SVM-based multilevel
binary tree classifier for fault diagnosis of hydroturbine
speed regulating system [J],” Proceedings of the CSEE,
Vol. 25, No. 8, pp. 100–104, 2005. (in Chinese)
[8] Y. C. Li, T. J. Fang, and E. K. Yu, “Study of support
vector machines for short-term load predicting [J],”
Proceedings of the CSEE, Vol. 23, No. 6, pp. 55–59,
2003. (in Chinese)
[9] Y. Wang, Z. H. Zhou, and A. Y. Zhou, “The apply of
machine study [M],” The Publish House of Tsinghua
University, Beijing, pp. 1–27, 2006.
[10] A. Smola and B. Scholkopf, “A tutorialon support vector
regression [R],” Royal Holoway College, London,1998.
[11] V. N. Vapnik, “The nature of statistical learning theory
[M],” Springer, New York, 1999.
[12] G. Z. Li, M. Wang, H. J. Zeng (translate), N. Cristianini,

W. P. YAN ET AL.
93
J. Shawe-Taylor (write), “Introduction of support vector
[M],” The Publish House of Electric Industry, Beijing, 2004.
[13] X. G. Zhang (translate), Vapnik (write), “Theory of
statistic study [M],” The Publish House of Electric
Industry, Beijing, 2004.
[14] X. G. Zhang, “Introduction to statistical learning theory
and support vector machines [J],” ACTA Automatica
Sinica, Vol. 26, No. 1, pp. 32–43, 2000.
[15] X. D. Wang and J. Q. Wang, “A survey on support vector
machine traing and testing algorithms [J],” Computer
Engineering and Applications, Vol. 40, No. 13, pp. 75–78,
2000.
[16] S. S. Keerthi and C. J. Lin, “A symptotic behaviors of
support v machines with Gaussian kernel [J],” Neural
Computation, pp. 1667–1689, 2003.
[17] C. L. Wang and H. Zhou, et a1., “Support vector machine
modeling on the unburned carbon in fly ash [J],” Pro-
ceedings of the CSEE, Vol. 20, No. 25, pp. 72–76, 2005.
(in Chinese)