Professional Documents
Culture Documents
net/publication/233155613
CITATIONS READS
17 18,367
2 authors:
All content following this page was uploaded by Nazan Yalçın Erik on 02 January 2014.
To cite this Article Erk, Nazan Yalçin and Yilmaz, Işk(2011) 'On the Use of Conventional and Soft Computing Models for
Prediction of Gross Calorific Value (GCV) of Coal', International Journal of Coal Preparation and Utilization, 31: 1, 32 —
59
To link to this Article: DOI: 10.1080/19392699.2010.534683
URL: http://dx.doi.org/10.1080/19392699.2010.534683
This article may be used for research, teaching and private study purposes. Any substantial or
systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or
distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss,
actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.
International Journal of Coal Preparation and Utilization, 31: 32–59, 2011
Copyright # Taylor & Francis Group, LLC
ISSN: 1939-2699 print=1939-2702 online
DOI: 10.1080/19392699.2010.534683
INTRODUCTION
Coal is valued for its energy content and, since the 1880s, is widely used to
generate electricity. Steel and cement industries use coal as a fuel for
extraction of iron from iron ore and for cement production. Coal mining
is one of the most important mining activities in the world and there has
been found very large terrestrial areas in different parts of the world [1, 2].
Turkey has also large coal reserves—about 9 gigatons (GT) [3]. The
low-rank coals of Turkey represent the country’s major energy source
with their relatively large geological reserves. The coal-bearing terrestrial
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
The study presented herein aims to predict GCV of the coals from
proximate and ultimate analyses results using a traditional statistical
method (multiple regression [MR]) and two soft computing techniques
(artificial neural networks [ANN] and adaptive neuro-fuzzy inference
system [ANFIS]) and compares the models as to their prediction capa-
bilities. A total of 74 coal samples were collected from various locations
in Turkey and proximate and ultimate analyses performed. These para-
meters were correlated with the determined GCV first and statistically
significant ones were selected. In order to establish predictive models,
both statistical (multiple regression) and soft computing techniques (arti-
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
ficial neural networks and neuro-fuzzy models) were used and prediction
performances were then analyzed.
GCV TM A VM FC C H N O S Ro
Minimum 2.12 6.21 7.87 16.91 0.59 29.49 1.37 0.21 4.23 0.13 0.31
Maximum 27.65 55.37 74.26 60.08 45.09 78.52 6.56 2.54 67.63 9.12 0.57
Average 14.47 26.63 41.23 36.51 22.36 59.69 4.23 1.17 31.36 3.54 0.43
Std. Dev. 6.461 11.655 19.525 10.661 11.618 13.660 1.226 0.582 16.853 2.208 0.012
and they were converted into daf (dry ash free base) by calculation.
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
DATA PROCESSING
Particular attention is required to select the data set having a normal
distribution. In order to characterize the variation of GCV used as an
independent value, descriptive statistics such as minimum, maximum,
mean, mode, median, variance, standard deviation, skewness, and kur-
tosis were calculated using the SPSS Version 10.0 [24] package. Table 3
shows that the independent values show almost normal distribution.
However, it is close to the normal distribution and data are skewed right
and showed a kurtosis (Figure 2). It can be seen that the respective skew-
ness and kurtosis values of 0.028 and 0.931 were very low. In conclusion,
it was evident that the regression analyses will work well in this case.
In order to establish the predictive models among the parameters
obtained in this study, simple regression analysis were performed in
38 N. YALÇIN ERİK AND I. YILMAZ
N Valid: 74
N Missing: 0
Mean 14.4711
Std. Error of Mean 0.7510
Median 14.4750
Mode 9.89
Std. Deviation 6.4607
Variance 41.7410
Skewness 0.028
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
Table 4. Correlation coefficients (R2) obtained from the simple regressions between GCV
with other parameters
Model TM A VM FC C H N O S
Linear .043 .85 .60 .69 .71 .62 .29 .75 .66
Logarithmic .029 .80 .66 .59 .67 .60 .29 .72 .64
Exponential .067 .76 .50 .66 .70 .62 .31 .74 .58
Power .061 .64 .53 .63 .69 .68 .34 .62 .72
Note. Gray-filled cells show the highest correlation coefficients (R2); gray-filled cells with
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
Predictive model R2
the first stage of the analysis. The relations between GCV with other
parameters were analyzed employing linear, power, logarithmic, and
exponential functions. Statistically significant and strong correlations
were then selected (Table 4), and regression equations were established
among GCV with proximate (Set I) and ultimate (Set II) analyses results
(Table 5). All obtained relationships were found to be statistically signifi-
cant according to the Student’s t test with 95% confidence, except nitro-
gen (N). Figures 3 and 4 show the plot of the GCV versus moisture
content (TM), ash (A), volatile matter (VM), fixed carbon (FC), carbon
(C), hydrogen (H), nitrogen (N), sulfur (S), and oxygen (O).
Figure 3. GCV versus moisture content (TM), ash (A), volatile matter (VM), and fixed
carbon (FC) graphs for Set I.
CONVENTIONAL AND SOFT COMPUTING MODELS 41
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
Figure 4. GCV versus carbon (C), hydrogen (H), nitrogen (N), oxygen (O), and sulfur (S)
graphs for Set II.
42 N. YALÇIN ERİK AND I. YILMAZ
For Set I
Constant 2.894 0.695 4.167 .000
VM 0.231 0.022 10.450 .000
FC 0.408 0.022 18.900 .000
For Set II
Constant 3.274 17.450 0.188 .852
C 0.180 0.179 1.001 .320
H 1.056 0.552 1.913 .060
O 2.36 102 0.175 0.135 .893
S 0.932 0.328 2.840 .006
A is excluded because of colinearity.
CONVENTIONAL AND SOFT COMPUTING MODELS 43
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
varðy y 0 Þ
VAF ¼ 1 100; ð3Þ
varðyÞ
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
u1 X
RMSE ¼ t ðy y 0 Þ2 ; ð4Þ
N i¼1
where y and y0 are the measured and predicted values, respectively. The
calculated indices are given in Table 6. If the VAF is 100 and RMSE is 0,
then the model will be excellent. The obtained values of VAF and RMSE
given in Table 7 indicate the higher prediction performances.
Table 7. Performance indices (RMSE, VAF, and R2) for models used
Multiple Regression
Set I 1.645 93.125 .934
Set II 2.966 71.819 .782
Artificial Neural Networks
Set I 1.019 97.767 .981
Set II 1.341 96.584 .966
Adaptive Neuro-Fuzzy Inference System
Set I 0.565 99.383 .994
Set II 0.717 99.053 .989
Note. RMSE ¼ root mean square error; VAF ¼ value account for.
CONVENTIONAL AND SOFT COMPUTING MODELS 45
systems devised via imitating brain activity and have performance char-
acteristics like biological neural networks. ANN has a lot of important
capabilities such as learning from data, generalization, working with
unlimited number of variables [30]. Neural networks may be used as a
direct substitute for auto correlation, multivariable regression, linear
regression, trigonometric, and other statistical analysis techniques [31].
Neural networks, with their remarkable ability to derive meaning from
complicated or imprecise data, can be used to extract patterns and detect
trends that are too complex to be noticed by either humans or other com-
puter techniques. Rumelhart and McClelland [32] reported that the main
characteristics of ANN include large-scale parallel distributed proces-
sing, continuous nonlinear dynamics, collective computation, high
fault-tolerance, self-organization, self-learning, and real-time treatment.
A trained neural network can be thought of as an ‘‘expert’’ in the cate-
gory of information it has been given to analyze. This expert can then
be used to provide projections given new situations of interest and
answer ‘‘what if ’’ questions.
When a data stream is analyzed using a neural network, it is possible
to detect important predictive patterns that are not previously apparent
to a nonexpert. Thus, the neural network can act as an expert. The parti-
cular network can be defined using three fundamental components:
transfer function, network architecture, and learning law [33]. It is
essential to define these components, to solve the problem satisfactorily.
Neural networks consist of a large class of different architectures,
and there are many kinds of ANN models, among which the back propa-
gation (BP) model is the most widely used, and it is an instructive train-
ing model. The basic idea of the back propagation learning algorithm
[34] is the repeated application of the chain rule to compute the influence
of each weight in the network with respect to arbitrary error function E:
46 N. YALÇIN ERİK AND I. YILMAZ
@E @E @Yi @neti
¼ ; ð5Þ
@vij @Yi @neti @vij
where vij is the weight from neuron j to neuron i, Yi is the output, and neti
is the weighted sum of the inputs of neuron i.
Once the partial derivative for each weight is known, the aim of
minimizing the error function is achieved by performing a simple
gradient descent:
@E
vij ðt þ 1Þ ¼ vij ðtÞ r ðtÞ; ð6Þ
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
@vij
where t denotes time, r is the learning parameter.
It is accepted that the most useful neural networks in prediction and
decision algorithms are back propagation and Radial Basis Function
(RBF) networks. In this article, back propagation algorithm, created
by generalizing the Widrow–Hoff learning rule to multiple-layer net-
works and nonlinear differentiable transfer functions, is used. A back
propagation consists of an input layer, several hidden layers, and output
layers. All of those layers may contain multiple nodes [35].
The inputs xn, n ¼ 1, . . ., n to the neuron are multiplied by weights wni
and summed up together with the constant bias term Qi. The resulting xi
is the input to the activation function y. The activation function was orig-
inally chosen to be a relay function, but for mathematical convenience a
hyperbolic tangent (tanh) or a sigmoid function are most commonly
used. Hyperbolic tangent is defined as
ex ex
f ðxÞ ¼ tanhðxÞ ¼ : ð7Þ
ex þ ex
The output of node i become
!
X
k
yi ¼ f wik xj þ Qi : ð8Þ
j¼1
am m
0:j ¼ ðx m Þj ¼ xj ; ð9Þ
CONVENTIONAL AND SOFT COMPUTING MODELS 47
X
si1
cm
i:j ¼ wi:j;k am
i1:k þ bi:j ; i > 0; ð10Þ
k¼1
am m
i:j ¼ Fi:j ðc i:j Þ; i > 0: ð11Þ
!
X
si1
am
i:j ¼ Fi:j wi:j;k am
i1:k þ bi:j ð12Þ
k¼1
am
L:1 ¼ ym ð13Þ
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
1
2 1
2 2
y m t m ¼ ðym tm Þ ¼ em ¼ em ð14Þ
2 2
All data were first normalized and divided into three data sets such
as training (1=2 of all data), test (1=4 of all data), and verification (1=4 of
all data). In this study MatLab 7.1 [36] software was used in neural
network analyses having a three-layer feed-forward network in both of
the two sets. They consist of an input layer (three neurons in Set I, four
neurons in Set II), one hidden layer (seven neurons in Set I, nine neurons
in Set II) and one output layer (Figure 6). In the analyses, learning and
momentum parameters, networks training, and activation (transfer)
function for all layers were respectively adjusted to 0.01, 0.9, trainLm,
and tansig. As in many other network training methods, models and
parameters were used to be able to reach minimum RMS values, and
the network goal was reached at the end of 421 and 537 iterations in
models of Set I and Set II, respectively.
Cross-correlation between predicted and observed values (Figure 7)
indicated that the ANN model constructed is highly acceptable for
prediction of GCV. RMSE, VAF, and R2 values are tabulated in Table 7.
48 N. YALÇIN ERİK AND I. YILMAZ
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
where lA(x) is called the membership function (MF) for the fuzzy set A.
The MF maps each element of X to a membership grade (or a value)
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
5. Layer 5. The single node in this layer is a fixed node labeled R, which
computes the overall output as the summation of all incoming signals.
X
n
E ¼ ðfk fk0 Þ2 ; ð19Þ
k¼1
where fk and f 0k are the kth desired and estimated output, respectively, and
n is the total number of pairs (inputs-outputs) of data in the training set.
In this study, a hybrid intelligent system called ANFIS (the adaptive
neuro-fuzzy inference system) for predicting GCV was also applied.
ANFIS was trained with the help of Matlab version 7.1 [36]; the SPSS
10.0 [24] package was used for RMSE and statistical calculations. Bell
function was used as MF type, and output MF was linear.
According to the RMSE, VAF, R2 values (Table 7) and cross-
correlation between predicted and observed values (Figure 9), the
ANFIS model constructed to predict GCV has a very high prediction
performance.
Figure 9. Cross-correlation of predicted and observed values of GCV for ANFIS model.
54 N. YALÇIN ERİK AND I. YILMAZ
1. The results of the models for prediction of the gross calorific value
showed that the equations obtained from the multiple regression
models have high-prediction performances.
2. In order to predict the gross calorific value, ANN models having
three (for Set I) and four (for Set II) inputs, with one output was
applied successfully and exhibited more reliable predictions than
the regression models.
3. The ANFIS model for prediction of gross calorific value revealed the
most reliable predictions when compared with the multiple regression
and ANN models.
Figure 10. Comparison graphics of RMSE, VAF, and R2 for the models (MR, ANN,
ANFIS).
Figure 11. Graphics showing the variation of the values predicted by MR, ANN, and
ANFIS models from the observed values.
REFERENCES
1. Marschalko, M., M. Fuka, and L. Treslin. 2008. Influence of mining activity
on selected landslide in the Ostrava-Karvina coalfield. ACTA Montanistica
Slovaca 13: 58–65.
2. Marschalko, M., and L. Treslin. 2009. Impact of underground mining to
slope deformation genesis at Doubrava Ujala. ACTA Montanistica Slovaca
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011
14: 232–240.
3. Turkish Lignite Authority (TKI). 2008. Available at http://www.tki.gov.tr (in
Turkish).
4. Yilmaz, I. 2006. Indirect estimation of the swelling percent and a new classi-
fication of soils depending on liquid limit and cation exchange capacity.
Engineering Geology 85: 295–301.
5. Channiwala, S. A., and P. P. Parikh. 2002. A unified correlation for estimat-
ing HHV of solid, liquid and gaseous fuels. Fuel 81: 1051–1063.
6. Schuster, V. F. 1951. Über die berechnung des heizwertes von kohlen aus der
immediatzusammensetzung. Brennstoff – Chemie 32: 19–20.
7. Spooner, C. E. 1951. Swelling power of coal. Fuel 30: 193–202.
8. Mazumder, B. K. 1954. Coal systematics: deductions from proximate analysis
of coal. part I. Journal of Scientific and Industrial Research 13B: 857–863.
9. Mazumder, B. K. 2000. Theoretical oxygen requirement for coal combustion:
Relationship with its calorific value. Fuel 79: 1413–1419.
10. Mason, D. M., and K. N. Gandhi. 1983. Formulas for calculating the calor-
ific value of coal and chars. Fuel Processing Technology 7: 11–22.
11. Given, P. H., D. Weldon, and J. H. Zoeller. 1986. Calculation of calorific
values of coals from ultimate analyses: Theoretical basis and geochemical
implications. Fuel 65: 849–854.
12. Kucukbayrak, S., B. Durus, A. E. Mericboyu, and E. Kadioglu. 1991.
Estimation of calorific values of Turkish lignites. Fuel 70: 979–981.
13. Demirbas, A. 1997. Calculation of higher heating values of biomass fuels.
Fuel 76: 431–434.
14. Cordero, T., F. Marquez, J. Rodriquez-Mirasol, and J. J. Rodriguez. 2001.
Predicting heating values of lignocellulosic and carbonaceous materials from
proximate Analysis. Fuel 80: 1567–1571.
15. Parikh, J., S. A. Channiwala, and G. K. Ghosal. 2005. A correlation for
calculating HHV from proximate analysis of solid fuels. Fuel 84: 487–494.
16. Patel, S. U., B. J. Kumar, Y. P. Badhe, B. K. Sharma, S. Saha, S. Biswas,
A. Chaudhury, S. S. Tambe, and B. D. Kulkarni. 2007. Estimation of gross
calorific value of coals using artificial neural networks. Fuel 86: 334–344.
58 N. YALÇIN ERİK AND I. YILMAZ
20. American Society for Testing and Materials (ASTM). 2004. D3174, Stan-
dard Method for Ash in the Analysis Sample of Coal and Coke from Coal.
In 2004 Annual Book of ASTM Standards, Gaseous Fuels; Coal and Coke,
Vol. 05.06. Philadelphia, PA: ASTM.
21. American Society for Testing and Materials (ASTM). 2004. D3175, Stan-
dard Method for Volatile Matter in the Analysis Sample of Coal and Coke.
In 2004 Annual Book of ASTM Standards, Gaseous Fuels; Coal and Coke,
Vol. 05.06. Philadelphia, PA: ASTM.
22. American Society for Testing and Materials (ASTM). 2004. D3302, Stan-
dard Method for Total Moisture in Coal, in 2004 Annual Book of ASTM
Standards, Gaseous Fuels: Coal and Coke, Vol. 05.06. Philadelphia, PA:
ASTM.
23. American Society for Testing and Materials (ASTM). 2004. D5373, Stan-
dard Test Methods for Instrumental Determination of Carbon, Hydrogen
and Nitrogen in Laboratory Samples of Coal and Coke, in 2004 Annual Book
of ASTM Standards, Gaseous Fuels: Coal and Coke, Vol. 05.06. Philadelphia,
PA: ASTM.
24. SPSS 10.0.1. 1999. Statistical Analysis Software (Standard Version).
Chicago, IL: SPSS Inc.
25. Yilmaz, I., and A. G. Yüksek. 2009. Prediction of the strength and elasticity
modulus of gypsum using multiple regression, ANN, ANFIS models and
their comparison. International Journal of Rock Mechanics and Mining
Sciences 46: 803–810.
26. Alvarez, G. M., and R. Babuska. 1999. Fuzzy model for the prediction of
unconfined compressive strength of rock samples. International Journal of
Rock Mechanics and Mining Sciences 36: 339–349.
27. Finol, J., Y. K. Guo, and X. D. Jing. 2001. A rule based fuzzy model for the
prediction of petrophysical rock parameters. Journal of Petroleum Science
Engineering 29: 97–113.
28. Yilmaz, I., and A. G. Yüksek. 2008. An example of artificial neural network
application for indirect estimation of rock parameters. Rock Mechanics and
Rock Engineering 41: 781–795.
CONVENTIONAL AND SOFT COMPUTING MODELS 59
29. Jin, Y., and J. Jiang. 1999. Techniques in neural-network based fuzzy system
identification and their application to control of complex systems. In Fuzzy
Theory Systems, Techniques and Applications, ed. C. T. Leondes. New York:
Academic Press.
30. Kaynar, O., I. Yilmaz, and F. Demirkoparan. 2011. Forecasting of natural
gas consumption with neural network and neuro fuzzy system. Energy Edu-
cation Science & Technology Part A: Energy Science and Research 26: 221–238.
31. Singh, T. N., R. Kanchan, A. K. Verma, and S. Singh. 2003. An intelligent
approach for prediction of triaxial properties using unconfined uniaxial
strength. Mining Engineering Journal 5: 12–16.
Downloaded By: [yalcýn, nazan][TÜBTAK EKUAL] At: 07:41 24 January 2011