You are on page 1of 7

International Journal of Coal Geology 85 (2011) 6571

Contents lists available at ScienceDirect

International Journal of Coal Geology


j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / i j c o a l g e o

Studies of relationships between Free Swelling Index (FSI) and coal quality by
regression and Adaptive Neuro Fuzzy Inference System
M. Tayebi Khorami a, S. Chehreh Chelgani b, James C. Hower c,, E. Jorjani a
a
b
c

Department of Mining Engineering, Science and Research Branch, Islamic Azad University, Poonak, Hesarak Tehran, Iran
Surface Science Western, Research Park, University of Western Ontario, London, Ont., Canada N6G0J3
Center for Applied Energy Research, University of Kentucky, 2540 Research Park Drive, Lexington, KY 40511, USA

a r t i c l e

i n f o

Article history:
Received 21 June 2010
Received in revised form 30 September 2010
Accepted 30 September 2010
Available online 8 October 2010
Keywords:
Free Swelling Index
Coking coal
Coal petrography
Ultimate analysis
Proximate analysis
Adaptive Neuro Fuzzy Inference System

a b s t r a c t
The results of proximate, ultimate, and petrographic analysis for a wide range of Kentucky coal samples were
used to predict Free Swelling Index (FSI) using multivariable regression and Adaptive Neuro Fuzzy Inference
System (ANFIS). Three different input sets: (a) moisture, ash, and volatile matter; (b) carbon, hydrogen,
nitrogen, oxygen, sulfur, and mineral matter; and (c) group-maceral analysis, mineral matter, moisture,
sulfur, and Rmax were applied for both methods. Non-linear regression achieved the correlation coefcients
(R2) of 0.38, 0.49, and 0.70 for input sets (a), (b), and (c), respectively. By using the same input sets, ANFIS
predicted FSI with higher R2 of 0.46, 0.82 and 0.95, respectively. Results show that input set (c) is the best
predictor of FSI in both prediction methods, and ANFIS signicantly can be used to predict FSI when regression
results do not have appropriate accuracy.
2010 Elsevier B.V. All rights reserved.

1. Introduction
All types of coal undergo chemical changes, but caking coal exhibits
physical changes in addition to chemical changes when heated. (Speight,
1994, 2005; Ghosh and Chatterjee, 2008). The series of physical changes in
caking coals during the heating process are softening, melting, fusing, and
resolidifying. These changes are within a specic temperature range, the
plastic range of coal, and the physical changes that occur within this range
are known as the plastic properties (Speight, 1994, 2005).
Knowledge of the plastic properties of coal is commonly used to
predict coke quality (Pierron et al., 1959; Habermehl et al., 1981; Lloyd et
al., 1990; Riley, 2007). Caking properties are an essential prerequisite for
coking coals (Miller, 2005); subsequently, a number of tests have been
devised to classify the caking properties of coals, including the Roga test,
GrayKing test, and Free Swelling Index (Speight, 1994, 2005).
The Free Swelling Index (FSI), an important property for both
metallurgical and steam coals (Hower and Eble, 1996), is a measure of
the increase in volume of powdered coal when heated under certain
conditions. In this method, 1 g of fresh powdered coal sample (250 m)
is placed in a silica crucible. The coal is leveled by lightly tapping, covered,
and heated to 800 C in a special furnace for 2.5 min. After cooling, the

Corresponding author. Tel.: + 1 859 257 0261; fax: + 1 859 257 0360.
E-mail address: hower@caer.uky.edu (J.C. Hower).
0166-5162/$ see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.coal.2010.09.011

cross sectional prole of the coke button is compared with a series of


standard proles and designated by numbers from zero to nine at an
interval of 0.5 in increasing order of size; coal with zero FSI being the least
swelling. According to this standard, FSI classied to 02, 24 and 49,
which show weakly, medium, and strongly caking ranges (Table 1)
(Speight, 1994, 2005; ASTM D-720, 1999; Riley, 2007). Coals that are low
in rank, such as lignite, or high in rank, such as anthracite do not usually
fuse and show no swelling value (Speight, 1994, 2005; Miller, 2005; Riley,
2007; Ghosh and Chatterjee, 2008), whereas the Free Swelling Index of
bituminous coal enhances as the rank increases (the peaking in medium
volatile bituminous rank range) (Rees, 1966; Speight, 1994, 2005; Hower
et al., 1994; Riley, 2007).
The relationship between coal's plasticity and its physical and chemical
properties was studied in previous works (Schapiro and Gray, 1966; Lloyd
et al., 1984, 1989, 1990; Cole and Williams, 1981; Clark et al., 1984;
Reasoner et al., 1985; Hower et al., 1994; Maroto-Valer et al., 1998;
Speight, 2005). The caking tendency of coals rises dramatically between
25 and 35% w/w volatile matter content (or between 81 and 92% w/w of
carbon in coals, with a maximum at 89% carbon) and then decreases
(Speight, 1994, 2005; Ghosh and Chatterjee, 2008). The amounts of
mineral matter content of the coal are signicant in coke production
because of diluting effect of ash (Miller, 2005); subsequently, the caking
tendency of coal decreases with increasing of mineral matter (Speight,
1994, 2005).
Fluidity is also dependent upon the organic sulfur content in coal
(Hower et al., 1994), which is also important for metallurgical processes.
Clark et al. (1984) measured organic sulfur in coal and semicoke samples

66

M.T. Khorami et al. / International Journal of Coal Geology 85 (2011) 6571

Coal type

Caking range

No.

Noncaking/weakly caking
Medium caking
Strongly caking

02
24
49

107
349
352

from Kentucky, and observing a trend of increasing organic sulfur content


with increasing plasticity of coals at the same rank. Cole and Williams
(1981) and Senftle and Davis (1982) found that plasticity was high only
for the coals with high organic sulfur content. Coals with similar maceral
composition and rank show a greater tendency to undertake coke
formation upon as the organic sulfur content increased (Ignasiak et al.,
1978; Yarzab et al., 1980).
Petrographic analysis can provide insight into the thermoplastic
properties of a particular coal, which are of signicant importance in
the coking industry (Miller, 2005). They display different carbonization behaviors, as expected from their different physical and chemical
characteristics (Krevelen, 1993; Stach et al., 1982; Falcon and Snyman,
1986). Schapiro and Gray (1966) classied the macerals into reactive
including liptinite and vitrinite that enhance the uidity of coal, and
inert, such as inertinite that is infusible during carbonization. Free
Swelling Index is also dependent on the maceral composition, with
vitrinite being the most obvious contributor to the swelling properties
in most coals and inertinite group-macerals act as diluent components, deleterious to swelling (Hower and Eble, 1996).
Hower et al. (1994) predicted coking characteristics of a number of
eastern Kentucky coals and found that FSI increased with increasing rank
(determined as vitrinite maximum reectance). FSI can be affected by the
relative amount of fusible components, vitrinite and liptinite, and of
diluent inert components. Maroto-Valer et al. (1998) studied the inuence
of petrography (density fractionation of petrography) on uidity property
of coking coals and obtained a signicant linear relationship (R2 N 0.96)
between uidity and maceral density, nding that uidity decreases with
increasing vitrinite and semifusinite density. Vitrinite reectance, as a
measure of coal rank, is perhaps the most signicant parameter
inuencing the coke making potential of any coking coal (Ghosh and
Chatterjee, 2008).
Some problems associated with the ASTM D-720 FSI method are
the proper heating rate, oxidation or weathering of the coal sample,
and an excess of ne coal in the analysis sample (Speight, 1994, 2005).
For more clarication, a gas burner as a source of heat requires
frequent calibration because gas pressures are not constant, relatively
small changes in gas pressure resulting in rather wide variations in the
temperature attained in a crucible, and thus in the resultant size of the
coke button. Furthermore, the quartz crucibles specied for the test

Frequency

Table 1
The numbers of samples in database regarding their caking ranges.

120

Mean~0.00
Std. Dev.=1.51

100

N=808

80

60

40

20

0
-6.00

-4.00

-2.00

0.00

2.00

4.00

Difference between actual and


predicted FSI
Fig. 1. Normal distribution of the difference between measured FSI values and
estimated FSI values obtained from multivariate regression Eq. (2).

are not standard equipment, and relatively small variations in


dimension and wall thickness from those specied often result in
rather wide variations in the resultant size of the button and the size
specied in the test is not normally used for any other purpose in the
laboratory; So, prediction of FSI from coal analysis data can be useful
(Swartzman and Behnke, 1952).
Using the experimental data, computing techniques have been
applied to many aspects of coal processing that were mentioned in the
references of the paper (Cilek, 2002; Yao et al., 2005; Patel et al., 2007;
Jorjani et al., 2007, 2008; Chehreh Chelgani et al., 2008, 2010). Adaptive
Neuro Fuzzy Inference System (ANFIS) is one of the most popular and
well documented neural fuzzy systems, which has a good software
support (Lot, 1995). Jang et al. (1997) presented the ANFIS architecture
and application examples in modelling a non-linear function, a dynamic
system identication and a chaotic time series prediction. Given its
potential in building neural fuzzy models with good prediction
capabilities (Cai et al., 2003), the ANFIS architecture was chosen for
modelling of this work.
The aim of the present work is prediction of FSI of coal according to
the proximate and ultimate analysis, group-macerals, mineral matter,
and vitrinite maximum reectance (Rmax) of coal using experimental
data obtained at a laboratory level. The multivariable regression and
Adaptive Neuro Fuzzy Inference System (ANFIS) was used to predict. To
our knowledge, this is the rst time that ANFIS has been used to predict
Free Swelling Index of coal by using mentioned input variables.
2. Experimental data

Table 2
The range of variables for 808 Kentucky coal samples.
Variable (%)

Minimum

Maximum

Mean

Standard deviation

Moisture
Ash
Volatile matter
Fixed carbon
Total sulfur
Pyritic sulfur
Sulfate sulfur
Organic sulfur
Carbon
Hydrogen
Nitrogen
Oxygen
Mineral matter
Vitrinite
Inertinite
Liptinite
Rmax

0.84
0.64
24.80
37.65
0.42
0.01
0.00
0.26
55.95
4.00
0.80
3.60
1.1
21.0
0.9
0.4
0.39

15.56
19.95
44.20
66.46
8.58
6.50
1.08
4.66
84.38
6.60
2.34
21.65
24.8
90.6
52.6
25.8
1.12

4.29
8.80
35.38
51.53
2.11
0.90
0.07
1.13
71.71
5.21
1.50
10.67
10.7
70.1
13.3
6.0
0.79

2.86
4.46
2.79
5.92
1.67
1.06
0.13
0.70
6.30
0.34
0.20
3.30
5.3
11.6
8.3
3.8
0.17

A remarkable mathematical model requires a comprehensive


database to cover a wide variety of coal types. That model should be
Table 3
The percentage of accurate prediction of FSI by non-linear regression and ANFIS
procedure (in testing stage) regarding different caking ranges.
Input sets

Caking
range

Non-linear
regression (%)

ANFIS
(%)

Proximate analysis

02
24
49
02
24
49
02
24
49

0.93
65.61
78.41
0.93
69.34
81.82
29.90
67.62
83.81

5.56
82.76
78.79
20
74.14
100
90
91.84
98.78

Ultimate analysis, mineral matter

Group-maceral analysis, mineral


matter, moisture, organic sulfur, Rmax

M.T. Khorami et al. / International Journal of Coal Geology 85 (2011) 6571

67

Table 4
FSI estimation deviations from target for non-linear regression equations.
FSI deviation
from target

Model (a)
02

24

49

02

Model (b)
24

49

02

24

49

Less than 0.5


Between 0.5 and 1
More than 1

0.00%
12.15%
87.85%

38.68%
30.66%
30.66%

21.88%
18.47%
59.66%

4.67%
15.89%
79.44%

39.26%
30.95%
29.80%

24.15%
23.58%
52.27%

18.69%
30.84%
50.47%

38.68%
32.95%
28.37%

32.39%
26.14%
41.48%

capable for predicting of FSI with a high validity. Data used to test the
proposed approaches are from studies conducted at the University of
Kentucky Center for Applied Energy Research. The database includes
the determined proximate and ultimate analysis, petrography,
mineral matter, and Rmax analysis as well as Free Swelling Index
(FSI) on an as determined basis. More than 800 coal sample analyses
(generally falling within the high volatile bituminous rank range)
were used. Table 1 shows the number of samples regarding their
coking ranges.
3. Adaptive Neuro Fuzzy Inference System (ANFIS)
The main objects of fuzzy systems that decision-making by using
the knowledge about a target, include human knowledge and perform
interfacing (Mohanadas and Karimulla, 2001). Fuzzy inference is the
process of formulating the mapping from a given input to an output by
using fuzzy logic (Jang et al., 1997). In addition, fuzzy logic is all about
the relative importance of precision (Lot, 1995) that is closer to
human thinking and natural language than conventional logical
systems (Zadeh, 1965).
The fuzzy inference system (FIS) based on the concepts of fuzzy set
theory, fuzzy if-then rules, and fuzzy reasoning (Jang et al., 1997) can
be classied into three types: Tsukamoto-type FIS (Tsukamoto, 1979),
Mamdani-type FIS (Mamdani and Assilan 1975), and TakagiSugenotype FIS (Sugeno and Kang 1988).
The methodology of the fuzzy logic controller (FLC) when the
processes are too complex for analysis or the available sources of
information are interpreted qualitatively, inexactly, or with uncertainty,
appears very useful by conventional quantitative techniques (Takagi
and Sugeno, 1974). The major problems in the fuzzy logic control that
have reduced its application are identifying the appropriate number of
rules, the difculty of choice, and design of membership functions to suit
a given problem (Mohanadas and Karimulla, 2001; Castellano and
Fanelli, 1996). Neural networks are capable of learning, but they cannot
interpret imprecise data that can be helpful in making decisions. This

120

learning capability of the neural network can be combined with the


control capabilities of a fuzzy logic system resulting in a neuro-fuzzy
inference system (Jang, 1993; Jantzen, 1998).
An Adaptive Neuro Fuzzy Inference System (ANFIS), based on the
architecture of the TakagiSugeno-type fuzzy inference system, is one
of the most popular neural fuzzy systems (FIS) (Lot, 1995; Demuth
and Beale, 2002). ANFIS developed by Jang (1993) has a good software
support (Lot, 1995) by using input/output data set, constructs a
fuzzy inference system (FIS), and tuning it with a back propagation
algorithm alone, or in combination with a least squares type of
method (Lot, 1995; Jang, 1993; Jang and Sun, 1995; Jang et al., 1997).
Fuzzy clustering is the partitioning of a collection of data into fuzzy
subsets or clusters based on similarities between the data (Passino and
Yurkovich, 1998) that develops a fuzzy estimation model, to predict the
output given the input (Lot, 1995). There are some basic methods of
fuzzy clustering: fuzzy C-means (FCM) clustering method, the most
popular method, and subtractive clustering (Lot, 1995; Grabusts,
2002). The fuzzy C-means clustering algorithm requires a desired
number of clusters (C) and initial guess positions for each cluster center.
Subtractive clustering has an auto-generation capability for determining
the number and initial location of cluster centers in a set of data when
there is not a clear idea how many clusters should be for a given set of
data. This method partitions the data into groups called clusters by
specifying a cluster radius, and generates a Sugeno-type fuzzy inference
system (FIS) with the minimum number of rules according to the fuzzy
qualities associated with each of the clusters. This type of FIS generation
can be accomplished automatically using generate fuzzy inference
system structure (GENFIS2) (Lot, 1995).
4. Results and discussion
4.1. Multivariable regression
Using all parameters of proximate or ultimate analysis in one equation
may increase the bias. For example, proximate analysis contains the

100

Mean~0.00
Std. Dev.=1.36
N=808

Mean~0.00
Std. Dev.=1.04
N=808

80

80

Frequency

Frequency

100

Model (c)

60

60

40

40
20

20

0
-6.00

-4.00

-2.00

0.00

2.00

4.00

Difference between actual and


predicted FSI
Fig. 2. Normal distribution of the difference between measured FSI values and
estimated FSI values obtained from multivariate regression Eq. (4).

-3.00

-2.00

-1.00

0.00

1.00

2.00

3.00

Difference between actual and


predicted FSI
Fig. 3. Normal distribution of the difference between measured FSI values and
estimated FSI values obtained from multivariate regression Eq. (6).

68

M.T. Khorami et al. / International Journal of Coal Geology 85 (2011) 6571

9.00
8.00

Predicted FSI=1.287+0.701 Actual FSI


R2=0.70

Predicted FSI

7.00
6.00
5.00
4.00
3.00
2.00
1.00
0.00
0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00

Actual FSI
Fig. 4. Graphical comparison of experimental FSIs with those estimated by multivariate
regression Eq. (6).

determination of moisture, volatile matter, ash, and xed carbon (ASTM


D-3172D-3175). Fixed carbon depends on relative amounts of moisture,
ash, and volatile matter then the difference of these three values summed
and subtracted from 100, Fixed Carbon = 100 (Moisture + Ash+
Volatile Matter). It is not necessary to use all four parameters since, by
denition; the four parameters are a closed system, adding to 100%
(Hower, 2006).
To prepare linear equations the stepwise variable selection procedure
was used, in which variables are sequentially entered into the model. The
rst variable considered for entering into the equation is the one with
largest positive or negative correlation with dependent variable. The
procedure stops when there are no variables that meet the entry criterion
(SPSS Inc, 2004).
Unlike linear regression, which is restricted to estimate linear models,
non-linear regressions create equations with arbitrary relationships
between independent and dependent variables. In this study, some
input variables have non-linear relationships with FSI. Therefore, with
respect to linear regressions, non-linear regression was used to develop
equations between Free Swelling Index and coal analysis variables.
In order to cover the whole structure of coal samples, three
different input sets of coal analyses were applied as FSI predictors:

Fig. 5. Normal distribution of the difference between measured FSI values and
estimated FSI values obtained from ANFIS (model (a)).

4.1.1. Proximate analysis


By a least square mathematical method, the inter correlations of
moisture, ash, and volatile matter with FSI were 0.52, 0.23, and
0.16, respectively. The results show that higher moisture contents
in coal can result lower FSI. The moisture content is, of course, a coal
rank parameter. Moisture as a diluent and moisture as an indicator of
coal rank are inseparable, providing a good example of the interrelationship of coal quality parameters. The other variables are not
signicant.
A linear relationship between input variables and FSI can be shown
as following equation:
FSI = 10:0970:313M0:091A0:103VM

Table 5
Parameters of ANFIS based on different input models.

FSI = 9:3181:781M0:054A0:022VM + 0:21M 0:008M

R = 0:38 :
2

2
Where M, A, and VM denote the percentage of moisture, ash, and
volatile matter, respectively. The other non-linear equations were
examined, and they did not improve the correlation coefcient.

50

Mean=0.2
Std. Dev.=0.53
N=200

Frequency

40

30

20

10

Model Input variable


no.

Training
No. of
No. of
training data radius rules
epoch

(a)
(b)
(c)

2
3
65

Proximate analysis
Ultimate analysis, mineral matter
Group-maceral analysis, mineral matter,
moisture, organic sulfur, Rmax

In addition, the non-linear relationship as following equation:

(a) Proximate analysis (moisture, ash, and volatile matter);


(b) Ultimate analysis (carbon, hydrogen, nitrogen, oxygen,
and sulfur), forms of sulfur, and mineral matter;
(c) Group-macerals analysis (vitrinite, inertinite, and liptinite),
mineral matter, moisture, vitrinite Rmax, and forms of sulfur.
The statistical parameters of input variables are shown in Table 2.
According to the stepwise least square mathematical method (SPSS
software) the best variables from each combination were selected to
predict FSI.

R2 = 0:32 :

0.61
0.9
0.98

4
2
2

0
-2.00

-1.00

0.00

1.00

2.00

Difference between actual and ANFIS


predicted FSI
Fig. 6. Normal distribution of the difference between measured FSI values and
estimated FSI values obtained from ANFIS (model (b)).

M.T. Khorami et al. / International Journal of Coal Geology 85 (2011) 6571

69

4.1.3. Group-macerals analysis, mineral matter, moisture, organic sulfur,


and Rmax
The inter correlations of vitrinite, inertinite, liptinite, and Rmax with
FSI are +0.29, 0.14, 0.27, and +0.55, respectively. The results show
that, within the rank range studied here, increase of vitrinite and Rmax in
coal results in higher FSI.
In this section, the following combinations of different coal analyses
were used:
i.
ii.
iii.
iv.
v.

Fig. 7. Normal distribution of the difference between measured FSI values and
estimated FSI values obtained from ANFIS (model (c)).

The normal distribution of difference between FSI predicted from


Eq. (2), and actual determined amounts of FSI is shown in Fig. 1. For the
examined data set, the percentage of accurate prediction of FSI using
proximate analysis (Eq. (2)) in three caking ranges is illustrated in
Table 3. In addition, the FSI estimation deviations from targets are
presented in Table 4. The results show that proximate analysis cannot be
a good predictor of FSI by using multivariable regression procedure.

Group-macerals, mineral matter


Group-macerals, Rmax, mineral matter
Group-macerals, Rmax, mineral matter, Sorg
Group-macerals, Rmax, mineral matter, moisture
Group-macerals, Rmax, mineral matter, Sorg, moisture.

Finally, the combination v was found to be the best predictor of


FSI. So as an input set, group-macerals (vitrinite, inertinite, and
liptinite), Rmax, mineral matter, Sorg, and moisture were used. By using
a stepwise regression method, the following linear and non-linear
equations were developed:
FSI = 1:332 + 6:882R max 0:278Lip0:150M0:052MM
R = 0:62:

FSI = 18:69 + 111:753R max 0:303Lip0:775M0:59MM


2

+ 0:838Sorg + 0:041M
+ 0:003Lip

4.1.2. Ultimate analysis and mineral matter


The inter correlations of carbon (C), hydrogen (H), nitrogen (N),
oxygen (O), organic sulfur (Sorg), pyritic sulfur (Spyr), sulfate sulfur(Ssul),
and mineral matter (MM) with FSI value were determined to be +0.49,
0.06, +0.21, 0.59, 0.10, 0.09, 0.24, and 0.23, respectively.
The results show that higher oxygen content in coal can result in lower
FSI and in higher carbon content results in higher FSI. The other variables
are not signicant.
The best linear and non-linear relationships between these
variables and FSI can be shown as the following equations:

+ 0:368Sorg

2
0:162Sorg

2
156:24R max

3
72:474Rmax
2

R = 0:70:

The normal distribution of the difference between FSI predicted


from Eq. (6) and actual determined amounts of FSI is shown in Fig. 3.
For the examined data set, the percentage of accurate prediction of FSI
using Eq. (6) in three caking ranges is illustrated in Table 3. In
addition, the FSI estimation deviations from targets are presented in
Table 4. The results show that Eq. (6) can produce better results than
Eqs. (2) and (4). Fig. 4 shows comparative plots of the FSIs determined
experimentally and estimated by multivariate regression Eq. (6).
4.2. ANFIS prediction

FSI = 29:015 + 0:069O + 0:234MM + 0:538Sorg + 0:406C


2

+ 0:380Spyr

R = 0:41

FSI = 19:6580:118O + 0:435MM + 1:343Sorg 1:021C + 0:29Spyr


2

0:013O + 0:001O + 0:01C 0:006MM 0:225Sorg


2

R = 0:49:
4

The normal distribution of difference between FSI predicted from


Eq. (4) and actual determined amounts of FSI is shown in Fig. 2. For the
examined data set, the percentage of accurate prediction of FSI using
Eq. (4) in three caking ranges is illustrated in Table 3. The FSI estimation
deviations from targets are presented in Table 4. The results show that
there is no acceptable correlation coefcient between coal ultimate
analysis and FSI.
Table 6
The performance of ANFIS procedure for the input models (a) to (c).
Model

Correlation coefcient
(R2)

(a)
(b)
(c)

0.47
0.82
0.95

To increase accuracy of regression results, ANFIS procedure was


applied. The input models (a) to (c) (Section 4.1) were used as
predictors. A total of 808 sets of data were used in the predictions by
ANFIS; 608 data sets for training and 200 data sets for testing.
In this work, gens2 was used as a stand-alone, fast method for
generating a fuzzy model from data. Hybrid learning algorithm, a
combination of least squares and back propagation gradient, was
applied to identify the membership function parameters of singleoutput, Sugeno-type fuzzy inference systems (FIS). In gens2, the
default input membership function type is gaussmf (Gaussian type
membership function), and the default output membership function
type is linear (Lot, 1995). The fuzzy inference system was built to
determine the training epoch number and radius clustering of data
sets that are signicantly different for each input model.
On the other words, the rule extraction method, in gens2, rst
uses the subclust function to determine the number of rules, and
antecedent membership functions and then uses linear least squares
estimation to determine each rule's consequent equations. This
function returns a FIS structure that contains a set of fuzzy rules to
cover the feature space (Lot, 1995). Each input and output has as
many membership functions as the number of clusters that subclust
has identied (The MathWorks, 1999). Table 5 shows the parameters
that used to build 3 different ANFIS models.
The training can be made more efcient by certain pre-processing
steps. In the present study, all inputs and output data in the training

70

M.T. Khorami et al. / International Journal of Coal Geology 85 (2011) 6571

Table 7
FSI estimation deviations from target for various ANFIS models.
FSI deviation from
target

Model (a)
02

24

49

02

Model (b)
24

49

02

24

49

Less than 0.5


Between 0.5 and 1
More than 1

5.56%
11.11%
83.33%

50.86%
32.76%
16.38%

43.94%
39.39%
16.67%

20.00%
30.00%
50.00%

68.10%
27.59%
4.31%

84.38%
10.94%
4.69%

85.00%
10.00%
5.00%

81.63%
16.33%
2.04%

91.46%
7.32%
1.22%

phase were preprocessed by normalizing the inputs and targets so


that they have means of zero and standard deviations of 1:


Np = Ap meanAps = stdAp :

Where Ap is the actual parameter; meanAps is the mean of the actual


parameters; stdAp is the standard deviation of the actual parameters; and
NP is the normalized parameter, the input data (Demuth and Beale, 2002).
The normal distributions of difference between estimated FSI and
actual values in models (a), (b), and (c) are shown in Figs. 57,
respectively. For the examined data sets, the percentage of accurate
prediction of FSI in three caking ranges is illustrated in Table 3. The
correlation coefcients of FSI prediction using input models (a) to (c)
were 0.46, 0.82, and 0.95, respectively (Table 6). In addition, the FSI
estimation deviations from targets for various models are presented
in Table 7. It was observed that Free Swelling Index prediction using
input model (c) and ANFIS procedure could be acceptable and
satisfactory. Fig. 8 shows graphical comparison of experimental FSIs
with those estimated by ANFIS model (c) in testing stage.
4.3. Comparison between input variables, regression, and ANFIS
Eqs. (1)(6) can predict FSI with R2 of 0.32, 0.38, 0.41, 0.49, 0.62,
and 0.70, respectively. It can be seen that Eq. (6), using input variables
of group-macerals, mineral matter, moisture, organic sulfur, and Rmax
analysis (model (c)), with the minimum deviations from laboratory
estimated FSIs (Table 4), yielded the best prediction of FSI.
The percentage of accurate prediction of FSI by multivariable
regression procedure using input model (c) regarding caking ranges
(02, 24 and 49) were 29.9, 67.62, and 83.81%, respectively (Table 3).
Table 6 shows that ANFIS procedure can predict FSI with R2 of 0.47,
0.82, and 0.95 for input models of (a), (b), and (c), respectively. It can
be seen that model (c), with the highest correlation coefcients (R2)
of 0.95 and minimum deviations from laboratory estimated FSIs
(Table 7), is more suitable than the other predictors. According to
caking ranges the percentage of accurate prediction of FSI using this
input for FSIs 02, and 24 were 90, and 91.84, respectively, and for

Fig. 8. Graphical comparison of experimental FSIs with those estimated by ANFIS model
(c) in testing stage.

Model (c)

FSIs 49 were 98.78% (Table 3). The result of prediction shows higher
accuracy in strongly caking range than the other caking ranges.
According to the above-mentioned results, ANFIS can predict the FSI
with more accuracy than regression. In addition, as input variables the
combination of group-macerals, mineral matter, moisture, organic sulfur,
and Rmax analysis, could be proposed as the most suitable model for the FSI
prediction.
5. Conclusions
Because associated problems with the ASTM D-720 method to determine FSI of coal such as need to the proper heating rate, oxidation or
weathering of the coal sample, and an excess of ne coal in the analysis
sample, prediction of FSI from basic characterization data is valuable.
Three models of: (a) proximate analysis, (b) ultimate analysis and
mineral matter, and (c) group-macerals analysis, mineral matter,
moisture, Rmax, and forms of sulfur, were used to predict FSI by using
stepwise regression method and ANFIS procedure.
The inter correlation between input variables (coal rank parameters)
and FSI showed, with increasing of moisture and oxygen content of coal,
the FSI decreased. Higher amount of carbon, vitrinite, and Rmax contents
of coal results in higher FSI.
The non-linear equations can predict FSI with correlation coefcients
(R2) of 0.38, 0.49, and 0.70, for models (a), (b), and (c), respectively.
The ANFIS procedure can predict FSI with correlation coefcients of 0.47,
0.82, and 0.95 for the input models of (a), (b), and (c), respectively. The
results are much better than multivariable regression.
The percentage of accurate prediction of FSI using input model (c) in
caking ranges of 02, 24, and 49 by ANFIS procedure were 90,
91.84, and 98.78%, respectively.
The results are shown that ANFIS can be applied as a reliable method to
predict FSI using group-macerals analysis, mineral matter, moisture,
organic sulfur, and Rmax analysis of coal as predictors.
References
ASTM D-720, 1999. Test method for Free-Swelling Index of coal. ASTM International
226230.
Cai, C.H., Du, D., Liu, Z.Y., 2003. Battery state-of-charge estimation using Adaptive
Neuro-Fuzzy Inference System (ANFIS). IEEE International Conference on Fuzzy
Systems 10681073.
Castellano, G., Fanelli, A.M., 1996. Simplifying a neuro-fuzzy model. Neural Processing
Letters 4, 7581.
Chehreh Chelgani, S., Hower, J.C., Jorjani, E., Mesroghli, Sh., Bagherieh, A.H., 2008.
Prediction of coal grindability based on petrography, proximate and ultimate
analysis with multiple regression and articial neural network models. Fuel
Processing Technology 89, 1320.
Chehreh Chelgani, S., Mesroghli, Sh., Hower, J.C., 2010. Simultaneous prediction of coal
rank parameters based on ultimate analysis using regression and articial neural
network. International Journal of Coal Geology 83, 3134.
Cilek, E.C., 2002. Application of neural networks to predict locked cycle otation test
results. Minerals Engineering 15, 1095.
Clark, C.P., Freeman, G.B., Hower, J.C., 1984. Non-matrix corrected organic sulfur
determination by energy dispersive X-ray spectroscopy for western Kentucky coals
and residues. Scanning Electron Microscopy II, 537545.
Cole, G., Williams, D., 1981. The Manchester coals and coal-bearing rocks of Eastern
Kentucky. In: Cobb, J.C., et al. (Ed.), Coal and Coal-bearing Rocks of Eastern
Kentucky. Kentucky Geological Survey, KY, pp. 8291.
Demuth, H., Beale, M., 2002. Neural Network Toolbox for use with MATLAB. User's
Guide Handbook, version 4, 154. The MathWorks Inc., Berkeley, California.
Falcon, R.M.S., Snyman, C.P., 1986. An Introduction to Coal Petrography: Atlas of
Petrographic Constituents in the Bituminous Coals of Southern Africa. The
Geological Society of South Africa.

M.T. Khorami et al. / International Journal of Coal Geology 85 (2011) 6571


Ghosh, A., Chatterjee, A., 2008. Ironmaking and Steelmaking: Theory and Practice.
Prentice-Hall of India Private Limited, New Delhi, pp. 141143.
Grabusts, P., 2002. Clustering Methods in Neuro-Fuzzy Modelling. Riga Technical
University. Decision Support Systems Department, Information Technology Institute.
Habermehl, D., Orywal, F., Beyer, H.D., 1981. Plastic properties of coal. In: Elliot, M.A.
(Ed.), Chemistry of Coal Utilization, Second Supplementary vol.. Wiley-Interscience, New York., USA, pp. 317368.
Hower, J.C., 2006. Letter to the Editor; Discussion: Li et al., Prediction of grindability
with multivariable regression and neural network in Chinese coal. Fuel 85,
13071308.
Hower, J.C., Eble, C.F., 1996. Coal quality coal utilization link often understated in
discussions. National Coal Leader 30 (5), 12.
Hower, J.C., Wild, G.D., Eble, C.F., 1994. Coal quality considerations in the future production
of metallurgical coal in eastern Kentucky. Journal of Coal Quality 13, 16.
Ignasiak, B.S., Fryer, J.F., Jadernik, P., 1978. Polymeric structure of coal. Structure and
thermoplasticity of sulphur-rich Rasa lignite. Fuel 57, 578584.
Jang, J.S.R., 1993. ANFIS: Adaptive-Network-Based Fuzzy Inference System. IEEE
Transactions on Systems, Man and Cybernetics 23 (5/6), 665684.
Jang, J.S.R., Sun, C.T., 1995. Neuro-fuzzy modeling and control. Processing of the IEEE 83,
378406.
Jang, J.S.R., Sun, C.T., Mizutani, E., 1997. Neuro-Fuzzy and Soft Computing: A
Computational Approach to Learning and Machine Intelligence. Upper Saddle
River, NJ, Prentice Hall.
Jantzen, J., 1998. Neurofuzzy Modelling, Technical University of Denmark, Department
of Automation. Tech. report no 98-H-874, 2.
Jorjani, E., Chehreh Chelgani, S., Mesroghli, Sh., 2007. Prediction of microbial desulfurization of coal using articial neural networks. Minerals Engineering 20, 12851292.
Jorjani, E., Chehreh Chelgani, S., Mesroghli, Sh., 2008. Application of articial neural
networks to predict chemical desulfurization of Tabas coal. Fuel 87, 27272734.
Krevelen, D.W.Van, 1993. Coal3rd ed. Elsevier, Amsterdam. Chapters 23 and 24.
Lloyd, W.G., Reasoner, J.W., Hower, J.C., Yates, L.P., Clark, C.P., Davis, E., Fitzpatrick, A.,
Iren, A., Jimenez, A., Jones, T.M., Reagles, C.L., Sturgeon, L.P., Whitt, J.M., Wild, G.D.,
1984. Predictors of plasticity in bituminous coals. Final Technical Report, 178.
Western Kentucky University, US Department of Energy. DOE/PC/40793.
Lloyd, W.G., Reasoner, J.W., Hower, J.C., Yates, L.P., 1989. Effects of pressure upon the
isothermal plastic behavior of high-volatile bituminous Kentucky coals. Energy and
Fuels 3, 585589.
Lloyd, W.G., Reasoner, J.W., Hower, J.C., Yates, L.P., 1990. Estimates of uid properties of
high volatile bituminous coals. Fuel 69, 12571270.
Lot, A.Z., 1995. Fuzzy Logic Toolbox for use with MATLAB. User's Guide Handbook,
Version 2, 38. The MathWorks Inc., Berkeley, California, pp. 109112.
Mamdani, E.H., Assilan, S., 1975. An experiment in linguistic synthesis with a fuzzy logic
controller. International Journal of Man-Machine Studies 7, 113.
Maroto-Valer, M.M., Taulbee, D.N., Andresen, J.M., Hower, J.C., Snape, C.E., 1998. The
role of semifusinite in plasticity development for a coking coal. Energy & Fuels 12,
10401046.

71

Miller, B.G., 2005. Coal Energy Systems, 242. Elsevier publication, Burlington, USA.
Mohanadas, K.P., Karimulla, S., 2001. International Conference on Electrical and
Electronics Engineering Program. aramba Turkey.
Passino, K., Yurkovich, S., 1998. Fuzzy Control. Addison-Wesley, Menlo Park, CA.
Patel, S.U., Kumar, B.J., Badhe, Y.P., Sharma, B.K., Saha, S., Subhasish, B., Chaudhury, A.,
Tambe, S.S., Kulkarni, B.D., 2007. Estimation of gross caloric value of coals using
articial neural networks. Fuel 86, 334344.
Pierron, E.D., Rees, O.W., Clark, G.L., 1959. Plastic Properties of Coal, 269. Illinois State
Geological Survey. Circular.
Reasoner, J.W., Hower, J.C., Yates, L.P., Lloyd, W.G., 1985. Analytical pyrolysis as a tool
for the study of plastic coals. Fuel 64, 12691273.
Rees, O.W., 1966. Chemistry, uses, and limitations of coal analysis. Report of
Investigations, 220. Illinois State Geological Survey, Report of Investigations.
Riley, J.T., 2007. Routine coal and coke analysis: collection, interpretation, and use of
analytical data. ASTM International 8492.
Schapiro, N., Gray, R.J., 1966. A Guidebook for the Geological Society of America and
Associated Societies, Utah USA, 80, p. 55.
Senftle, J.T., Davis, A., 1982. A database for analysis of compositional characteristics of
coal seams and macerals. Final Report Part 1, The Pennsylvania State University,
Coal Research Section, p. 223.
Speight, G.J., 1994. The Chemistry and Technology of Coal. CRC Press, Berkeley
California, pp. 217229.
Speight, G.J., 2005. Handbook of Coal Analysis. Wiley-Interscience, Chichester, West
Sussex, England, pp. 141149.
SPSS, 2004. Version 13. SPSS Inc.. Help Files.
Stach, E., Mackowsky, M.Th., Teichmller, M., Taylor, G.H., Chandra, D., Teichmller, R.,
1982. Stach's Textbook of Coal Petrology, 3rd. ed. Borntraeger, Berlin.
Sugeno, M., Kang, T.G., 1988. Structure identication of fuzzy model. Fuzzy Sets and
Systems 28, 1533.
Swartzman, E., Behnke, G.C., 1952. Coal Using Electric Furnaces and Heaters to
Determine the Free-Swelling Index of Coal. The American Institute of Mining,
Metallurgical, and Petroleum Engineers. http://www.onemine.org.
Takagi, T., Sugeno, M., 1974. Application of fuzzy algorithms for control of simple
dynamic plant. IEE Proceedings 12, 15851588.
The MathWorks, Inc, 1999. Fuzzy Logic Toolbox DEMOS, Modeling Trafc Patterns
using Subtractive Clustering. The MathWorks Inc, Berkeley, California.
Tsukamoto, Y., 1979. An approach to fuzzy reasoning method. In: Gupta, M.M., Ragade,
K.K., Yager, R.R. (Eds.), Advances in Fuzzy Set Theory and Applications. North
Holland, Amsterdam.
Yao, H.M., Vuthaluru, H.B., Tade, M.O., Djukanovic, D., 2005. Articial neural networkbased prediction of hydrogen content of coal in power station boilers. Fuel 84,
15351542.
Yarzab, R.F., Given, P.H., Spackman, W., Davis, A., 1980. Dependence of coal liquefaction
behavior on coal characteristics. Cluster analyses for characteristics of 104 coals.
Fuel 59, 8192.
Zadeh, L.A., 1965. Fuzzy sets. Information and Control 8, 338353.

You might also like