Jangho Jo1, Dae-Gyun Lee1, Jongho Kim2, Byoung-Hwa Lee3, Chung-Hwan Jeon1,3. 1. School of Mechanical Engineering, Pusan National University, Busan 46241, Republic of Korea. 2. Chemical Engineering, University of Newcastle, Callaghan, New South Wales 2308, Australia. 3. Pusan Clean Energy Research Institute, Pusan National University, Busan 46241, Republic of Korea.
Abstract
The basic properties of coal influence various procedures of power generation, such as the design of power generation plants, estimation of the current condition of boilers, and total efficiency of power plants. The elemental composition is a needed factor in evaluating the process of chemical conversion and predicting the flow of flue gas and the quality of air in coal combustion. In the past, several relationships have been established using ultimate and proximate analyses. This study aims to predict the elemental compositions of 104 thermal coals used for coal-fired power plants in South Korea using an artificial neural network (ANN) that uses proximate analysis values as input parameters. The ANN-based model was optimized with six activation functions and four hidden layers after evaluating various performance indices, including R 2, mean square error (MSE), and epoch, then additional calculations were derived to compare performances from previous research using the mean absolute error (MAE), average absolute error, and average bias error. It was found that the best topology was established using the Levenberg-Marquardt activation function and 10 hidden layers, resulting in the highest R 2 value and smallest MSE of all topologies tested. As a result, the relative impact on calculation accuracy was derived from ANN hidden layers to analyze prediction accuracies of carbon, hydrogen, and oxygen compositions. Accuracy was improved over previous results by 4.71-0.91% via coal rank division topology optimization. Based on the MAE, the current results are even close in performance to those of adaptive neuro-fuzzy inference systems. They also outperformed previous research models by 5.40 and 7.39% in terms of MAE accuracy. Applicability of the ANN was also analyzed with limitations of the chemical composition of ANNs and present reinforcement measures in the future studies through qualitative analysis comparisons based on Fourier transform infrared spectroscopy. Consequently, the relative effect was derived from the ANN hidden layer weight for specific calculation of the relationship between elemental composition and proximate analysis properties. As a result, it was possible to qualitatively analyze how the proximate analysis value affects the composition of elements and calculate the ratio accordingly. The findings of this study provide an improved and efficient approach to predicting the elemental composition of thermal coal, based on data from South Korean power plants. Also, further research can follow schematics from this study with the applicability and accessibility of the ANN.
The basic properties of coal influence various procedures of power generation, such as the design of power generation plants, estimation of the current condition of boilers, and total efficiency of power plants. The elemental composition is a needed factor in evaluating the process of chemical conversion and predicting the flow of flue gas and the quality of air in coal combustion. In the past, several relationships have been established using ultimate and proximate analyses. This study aims to predict the elemental compositions of 104 thermal coals used for coal-fired power plants in South Korea using an artificial neural network (ANN) that uses proximate analysis values as input parameters. The ANN-based model was optimized with six activation functions and four hidden layers after evaluating various performance indices, including R 2, mean square error (MSE), and epoch, then additional calculations were derived to compare performances from previous research using the mean absolute error (MAE), average absolute error, and average bias error. It was found that the best topology was established using the Levenberg-Marquardt activation function and 10 hidden layers, resulting in the highest R 2 value and smallest MSE of all topologies tested. As a result, the relative impact on calculation accuracy was derived from ANN hidden layers to analyze prediction accuracies of carbon, hydrogen, and oxygen compositions. Accuracy was improved over previous results by 4.71-0.91% via coal rank division topology optimization. Based on the MAE, the current results are even close in performance to those of adaptive neuro-fuzzy inference systems. They also outperformed previous research models by 5.40 and 7.39% in terms of MAE accuracy. Applicability of the ANN was also analyzed with limitations of the chemical composition of ANNs and present reinforcement measures in the future studies through qualitative analysis comparisons based on Fourier transform infrared spectroscopy. Consequently, the relative effect was derived from the ANN hidden layer weight for specific calculation of the relationship between elemental composition and proximate analysis properties. As a result, it was possible to qualitatively analyze how the proximate analysis value affects the composition of elements and calculate the ratio accordingly. The findings of this study provide an improved and efficient approach to predicting the elemental composition of thermal coal, based on data from South Korean power plants. Also, further research can follow schematics from this study with the applicability and accessibility of the ANN.
Electric power is an irreplaceable and
essential resource in modern
society, and it is utilized in innovative industries and nearly all
aspects of life. Electric power can be generated from several sources,
and coal-fired power plants are the most popular globally owing to
their relatively low costs and simple fuel supply chains.[1] In these power plants, coal-fired boilers use
a precise amount of air to generate steam by combusting thermal coal
based on its estimated elemental composition. Proximate and ultimate
analyses are necessary to obtain gross and elemental level compositions
that are required for determining operating conditions. The ultimate
analysis derives the elemental composition of coal but requires a
longer analysis time than proximate analysis, which estimates its
moisture, volatile matter, fixed carbon, and ash content. Although
most properties of coal are well-understood, it remains difficult
to estimate the precise nature of a sample owing to its stochastic
composition and varying characteristics. It has been reported that
the systematic transformations resulting in coalification are likely
to permit correlations even when we do not understand the cause and
effect relationship between composition and sample properties.[2] However, to obtain some semblance of accuracy,
it has been proposed to apply artificial neural networks (ANNs) to
the problem. ANNs are used in various industrial and academic fields
to solve nonlinear problems, and they occupy a large body of research,
including that of determining coal properties.[3−8] Saptoro et al. designed a feed-forward neural network to predict
the elemental composition of coal and configured several ANN topologies
with various hidden layers, learning rates, and iterations. After
an optimum topology was found, the model was further improved.[6] Yi et al. predicted elemental compositions and
rankings of various coal samples and evaluated their correlation factors,
including the average absolute error (AAE), average bias error (ABE),
and R2 values.[9] Lawal et al. compared various methods for predicting the elemental
composition of coal and reported that ANN-model optimization can improve
performance with purposeful topology construction and appropriate
activation function diversification.[10] Jorjani
et al. used an ANN for desulfurization prediction of coal, and it
will be pretended as an advantage of this research, which is about
chemical composition tracing.[11] Basic analysis
from the result of the current manuscript can be utilized in various
chemical reactions. Another method is the gradient boosted regression
stress method applied by Samadi et al., but this technique has the
disadvantage of having difficulty in modification such as addition
of layers or computational connection based on predicted physical
properties.[12] In the study of K. Said et
al., the proximate analysis-based combustibility prediction was shown
through an ANN, but it can be said that it is difficult to provide
a basis for future research by quantifying the values shown in these
results.[13] Notably, prediction performance
and calculation efficiency are the two evaluation criteria that must
be considered when developing a prediction model based on a feed-forward
neural network for process control purposes. A high level of criteria
can be satisfied by employing a suitable training function, which
leads to a reasonable prediction with higher accuracy and efficiency
through training. Subsequently, performance evaluation was conducted
to build an even more optimized ANN model that used weights for prediction,
which is a unique property of ANNs.[6] Although
many researchers have leveraged ANN models to predict the elemental
composition of coal, the relative impact of proximate analysis has
yet to gain attention. It is expected that doing so would help fine-tune
the use of a variety of gaseous fuels (e.g., ammonia and methane)
at coal-fired power plants, optimistically resulting in the creation
of important new guidelines to achieve optimal reaction pathways.[9] Even if it is difficult to build an advanced
ANN such as the multilayer perceptron because of relative impact application
based on a single hidden layer topology, the future study that is
based on the proposed methodology from this manuscript can be combined
with other technical models due to the applicability and compatibility
of the ANN. There are several studies concerning the prediction of
the elemental composition of solid fuels from the proximate analysis
by applying machine learning algorithms. The performances of the models
are better than traditional empirical correlations but still need
to be improved with more advanced approaches.[14] By comparing the differences in chemical composition through Fourier
transform infrared (FT-IR) spectroscopy analysis, the limitations
in ANN-based calculations were analyzed, and the application direction
of the ANN in the future studies is discussed by analyzing the cause
of this phenomenon.[15]In this research,
an ANN model is trained using proximate analysis
inputs (i.e., moisture, volatile matter, fixed carbon, and ash content)
to predict elemental coal composition (i.e., carbon, hydrogen, and
oxygen) with element composition division, coal rank division, and
performance evaluation to reach a higher accuracy than that of extant
studies. Here, the combined results of the prediction mechanism and
the derived proximate analysis products of 104 South Korean thermal
coals are used as training data for the ANN model to predict elemental
composition via numerical calculations. Combining FT-IR analysis and
ANN calculation results can provide a basis for accuracy and logical
development in future studies on analysis based on elemental composition
using ANNs, as well as a link between chemical composition and thermal
properties.
Methods
Sample Data Preparation
Accuracy of a mathematical
model requires both certain quality and quantity of data that include
a variety of coal types.[16] The use of coal
of various ranks not only ensures accuracy during the processing but
also enables flexibility in situations such as extrapolation. When
a high-dimensional multilayer application technique is applied in
the neural network, learning accuracy can be further improved, but
it can be weakened by interference such as extrapolation in the operation
condition for nonlinear materials such as coal, and it can be overcome
through a simplified structure. Thermal coals utilized for coal-fired
power plants in South Korea were selected as samples to establish
a feasible prediction model for future applications. The samples were
ground in a mill (RS 200, Retsch GmbH, Haan, Germany) and dried at
40 °C for 12 h. The particle sizes were 75–90 μm,
which were sieved using a shaker (AS 200, Retsch GmbH). A 5 g sample
was used to conduct proximate analysis according to the ASTM D3172
standard using a thermogravimetric analyzer (TGA701, Leco Co., St.
Joseph, MI, USA). Carbon, hydrogen, nitrogen, and sulfur contents
were measured using an elemental analyzer (Leco-TruSpec Micro CHNS,
LECO Co., St. Joseph, MI, USA). Oxygen content is calculated by employing
the rest. For each sample, the proximate and ultimate analyses were
repeated three–five times, and the average calculation results
are used for further research.[17]Figure shows a van
Krevelen diagram with atomic O/C and H/C ratios based on whole-sample
elemental compositions. A total of 104 coal samples were classified
into sub-bituminous and bituminous types, depending on their elemental
compositions.[18] The complete data set is
provided in Appendix A. When using the first
plot-based coal properties as given, estimation is nonlinear, and
elemental relationships are not well-understood. The thermal composition
of a proximate analysis and the elemental composition are nonlinear,
but the linear assumption of empirical formulae may lead to erroneous
estimations.[19] Thus, it is difficult to
equate elemental composition and proximate analysis properties.[9] This is why researchers originally decided that
an ANN might offer better relativity and trend recognition.
Figure 1
Van Krevelen
diagram of thermal coals used in South Korea.
Van Krevelen
diagram of thermal coals used in South Korea.
ANN Model Structure
An ANN is capable of determining
the relationships between various data sets.[20] ANN models with the shape of a single neuron, likened to a simplified
biological neuron system, are mainly used.[21] In this ANN structure, the weight of the signal is given to the
neuron cell, and then, all are combined to form a firing spike signal
to activate the activation transfer function when the summing value
exceeds the threshold. In this study, the sigmoid function was used
as such an activation function.[21] The nntool modification code was used to generate the appropriate
ANN topology.[22]Figure shows the brief layout of an ANN that is
developed for this research. The input layer is set up with each proximate
analysis property. The output layer is set up with ultimate analysis
that is paired with proximate analysis as a sample. The random data
collection function is selected as the sample divide function. Among
the samples, 70, 15, and 15% were randomly selected for each training,
validation, and testing sets, respectively. The selection of these
samples was performed independently for each operation to increase
the accuracy of the ANN. Since the classification of data sets in
each operation proceeds differently, Appendix A did not directly classify for each section of the training, test,
and validation. The separation ratio of these samples is considered
to be the ratio for balancing between overfitting and validation accuracy,
which is performed repeatedly per each training process. Under the
hidden layer, each input layer data set is a tangential sigmoid activation
function that was used for training using a hidden layer with weight
and bias.[6] Then, the ANN was evaluated
for overfitting and training accuracy using regression graph training,
validation, and testing with R2 and a
data point distribution.[22] When an input
data set is introduced to the neural network, the synaptic weights
between the neurons are stimulated, and signals propagate to an output
layer based on weight and bias. Depending on how close the output
layer is to the predicted output layer, the weights and biases between
the layers are modified for improvement to provide an output layer
of more accurate results to the expected outcome with each step.[23] The training in an ANN model is critical for
accuracy and efficiency because it ensures that an accurate prediction
of the output value is achieved. We used various algorithms that are
widely known for adequately training ANN models: Levenberg–Marquardt
(L–M), Broyden–Fletcher–Goldfarb–Shanno
quasi-Newton, resilient backpropagation, scaled conjugate gradient,
conjugate gradient with powder/Beale resort.[24−29]
Figure 2
Schematic
layout of an ANN.
Schematic
layout of an ANN.The input layer consisted of moisture, volatile
matter, char, and
ash according to proximate analysis, and the hidden layer was set
to 10. Initially, the output layer was organized into three elements
(i.e., carbon, hydrogen, and oxygen); however, as observed in Figure , the pilot result
showed several problems, such as the exaggeration of R2 and the sharing of the same fitting function. For this
reason, the model was retrained using separate applications to obtain
more accurate weights and biases.[30] The
elements in the pilot run included hydrogen, oxygen, and carbon, respectively,
from the bottom left, which is distinguishable by comparing the data
distribution from the fitting results derived through subsequent learning.
It is possible to confirm that each proximate analysis value was learned
for elements and that the target values were influenced most under
the derivation.[22] The regression line indicates
accuracy, and the straight line, where Y and T values are the same, represents the most accurately predicted
result.[31] Various parameters can be cited
as components that can be an important role in the development of
ANN. Typically, the number of samples, selection function, and so
forth, are involved in this process, and it is divided into training–test–validation
sets by the selection function, and the ANN is configured according
to the learning function based on these data.[14,32]
Figure 3
ANN
regression result for three elemental compositions (carbon,
hydrogen, and oxygen) of 104 South Korean thermal coals (or data sets).
ANN
regression result for three elemental compositions (carbon,
hydrogen, and oxygen) of 104 South Korean thermal coals (or data sets).
ANN Performance Evaluation
It has been suggested that
a performance index should be gathered with each function to choose
an optimal function for ANN learning. Under the learning process,
the errors of both training and testing data sets are reduced with
each step. This procedure is repeated in the feed-forward backpropagation
ANN until the resulting errors have reached the threshold level specified
by the system’s error function, such as the root mean squared
error (RMSE).[10] The R2, optimal learning point arrival epoch, mean square error
(MSE), mean absolute error (MAE), AAE, and ABE were selected for this
case.[7,9,33,34] The performance of the model on the training and
testing data, MSE and MAE, indicate that it has both good predictive
ability and generalization performance.[35] The AAE indicates the degree of closeness between the predicted
and measured results, whereas the ABE represents the degree of underestimation
and overestimation. The R2 value used
for nonlinear regression shows how well the correlations mean after
removing the effect of the mean of the dependent variable.[9] These indices are calculated using eqs –5. We evaluated the performance based on the R2, MSE, and epoch as a pilot development, then we calculated
the MAE, AAE, and ABE for further comparison with previous research.where P and M represent the predicted and measured target values, respectively,
and n represents the number of data points used for
regression. The higher the R2 value, the
lower the MSE, and the smaller the epoch value, the greater the optimal
performance index. R2 in the nonlinear
regression indicates how strongly the dependent variable correlates
after excluding the average.[9] The epoch
is a presentation of the training set (input and/or target) vectors
to a network and the calculation of new weights and biases.[34] The functions used for performance evaluation,
shown in Table , provided
by the neural network toolbox in MATLAB were applied.[10,36]
Table 1
Various Training Algorithms and Performance
Evaluation Indicesa
training algorithm
performance
index
Levenberg–Marquardt
R2, MSE, epoch
Broyden–Fletcher–Goldfarb–Shanno quasi-Newton
resilient backpropagation
scaled conjugate gradient
conjugate gradient with powder/Beale resort
Mean square error (MSE).
Mean square error (MSE).Calculations were performed to derive the optimal
training algorithm
through trial and error, and the average value of 16 performance indices
was obtained by excluding the highest and lowest R2 and MSE values derived from 20 learning results.As shown in Figure , the ANN was validated in the direction of reducing MSE in the learning
process, and it was confirmed that the MSEs of the learning, verification,
and testing data sets gradually converged to the same point, effectively
ending the learning procedure to prevent overfitting. Based on this
training process, the validation data set and test data set selected
separately from the training data set undero the same process to verify
that this ANN learning process is accurate (in this manuscript, with
a low MSE). In the ANN model, the effect of each input factor on the
output parameters is represented by the weight and bias vectors between
neurons in different layers. We can derive the relative impact of
each input parameter with a weight vector using the proper equation.[30] Based on the learned ANN, the relative impact
from the input layer to the output layer was calculated using eq . By deriving the weight
of the hidden layer, it becomes possible to calculate the influence
on the elemental composition based on proximate analysis properties.[30]
Figure 4
Performance of the learning procedure to carbon
with training,
validation, and testing data sets.
Performance of the learning procedure to carbon
with training,
validation, and testing data sets.Relative impact Ir is
calculated based
on the weight differences between hidden input and output layers. W1 represents the weight from the ith neuron of the input layer to the jth
neuron of the hidden layer, W2 represents the weight from the jth neuron of the
hidden layer to the ith neuron of the output layer,
and n represents the number of hidden layers. The
numerator represents the weight collection of each hidden layer, and
the denominator separates each parameter for input. Thus, the relative
impact from a hidden layer could be calculated to derive an approach
for the singular node source to the whole layer.[30] Due to this application of relative impact, the relationship
between input and output based on the design of the ANN close to the
black box model may also be quantified, thereby opening up the possibility
as a grey box model. Understanding the relationship between the input
data set and output data set of nonlinear data will result in improved
results through techniques such as FT-IR spectroscopy, which stands
for strengthening applicability of the ANN, for reinforcing the prediction
of nonlinearity and anisotropy of coal.[21] Through this series of processes, the elemental composition was
quantitatively predicted through the accumulation of weight hidden
layers based on industrial analysis (Figure ).
Figure 5
Brief flowchart of ANN application.
Brief flowchart of ANN application.
FT-IR Spectroscopy
After this ANN-based analysis, there
is an inherent limitation to predicting a similar elemental analysis
for similar proximate analysis values for ANN-based calculations.
Several methods within model development such as ANFIS or gradient
boost regression have been tried to overcome this issue. We would
like to present a breakthrough plan for this limitation through the
analysis of the chemical structure of coal based on FT-IR spectroscopy
and suggest a direction for what form of analysis should be conducted
in future studies following the technique of this paper. For the additional
experiment, two coal samples were chosen for similar proximate properties
for analysis comparison, as shown in Table , but their elemental compositions may be
diverse due to their source and formation process.
Table 2
Basic Analysis Properties of the FT-IR
Sample for ANN Analysis
YP
CP
proximate properties (air dry basis, wt %)
inherent
moisture
0.415
0.595
volatile matter
10.84
17.44
fixed carbon
80.9
73.7
ash
7.85
8.12
ultimate analysis (dry and ash-free basis, wt %)
carbon
89.73
88.02
hydrogen
4.01
4.64
nitrogen
2.97
3.5
oxygen
2.82
3.65
sulfur
0.47
0.2
All the samples were crushed to pass through a 200-mesh
sieve (particle
size of <74 lm) and then characterized using the conventional analyses.
Proximate and ultimate analyses of the selected samples have been
conducted through the same procedure from ANN database development.[17] Spectra were collected using a Thermo Scientific
Nicolet IS-50 spectrometer outfitted with a diamond crystal attenuated
total reflection (ATR) accessory. Spectra were recorded in the ATR
mode and were corrected using the ATR correction of the OMNIC software.
All the spectra were acquired between 4000 and 550 cm–1 with 64 accumulations and a spectral resolution of 4 cm–1. Deconvolution routines were sometimes used to enhance resolution
in some spectral ranges (self-Fourier deconvolution of OMNIC).[37]
Results and Discussion
Performance Evaluation
The effects of moisture, volatile
matter, fixed carbon, and ash to the output of carbon, hydrogen, oxygen,
MSE, R2, and epoch values derived from
six functions after 16 rounds of training are summarized in Table . As a result of performance
evaluation, the LM algorithm was selected as the optimal algorithm,
which reached prominent performance, as shown in Table .
Table 3
Performance Evaluation Results of
Each Activation Functiona
Broyden–Fletcher–Goldfarb–Shanno
quasi-Newton (BFG), Levenberg–Marquardt (LM), conjugate gradient
with powder/Beale resort (CGR), resilient backpropagation (RB), scaled
conjugate gradient (SCG), and one-step secant (OSS).The results shown in Table reflect a variety of hidden layer counts
for optimal topology
development based on the LM algorithm. Based on carbon data, 10 hidden
layer topologies showed the best result for MSE, R2, and epoch values. Using carbon data, the MSE was determined
to be significantly lower than the other measures. Oxygen data presented
the best R2 value. Thus, topology was
chosen based on 10 hidden layers from the result showing the best
performance with hidden layer counts of 5, 10, 15, and 20 based on
epoch, MSE, and R2 values.
Table 4
Performance Evaluation Results of
Each Hidden Layer Topology Based on the L–M Algorithma
element
hidden
layer
MSE
R2
epoch
carbon
5
2.165
0.966
7.500
10
1.689
0.968
4.167
15
3.695
0.965
4.333
20
2.905
0.968
5.667
hydrogen
5
0.018
0.910
7.500
10
0.012
0.923
5.500
15
0.017
0.937
5.571
20
0.017
0.926
4.500
oxygen
5
2.061
0.965
6.167
10
1.951
0.967
6.333
15
3.124
0.967
6.167
20
4.109
0.962
4.500
Mean square error (MSE).
Mean square error (MSE).Next, the element content tracing activity of the
ultimate analysis
approximation was discussed. To establish predictive models among
the parameters obtained in the study, a simple regression analysis
can be performed during the first stage. The relationships between
output and other parameters were analyzed employing linear, power,
logarithmic, and exponential functions. Statistically significant
and strong correlations were found to be linear, and regression equations
were established among index parameters.[38] Under these circumstances, a good regression result would occur
if the fitting line (blue) matches the Y = T line (dotted).From the results of element content
tracing, as shown in Figure , both the y-intercept
and slope differed between the target to pilot learning fitting. The
data set distribution was found to be well-aligned to the Y = T line; thus, the elemental composition
derivation based on each element is more accurate than the total derivation.
In the case of hydrogen, the regression plot accuracy and R2 value did not reach those of the other elements.
Thus, it required another type of improvement. The oxygen regression
result meets the high R2 value, but the
regression plot did not match the Y = T line owing to low learning accuracy. Furthermore, the total learning
plot, which contains both re-learning and validation of the test data
set and comparison results of training, is shown in Table . The relative impact derived
from the training of the nonlinear characteristics of coal has meaningful
accuracy.
Figure 6
Total R2 for each element across 104
data sets, (A: carbon, B: hydrogen, and C: oxygen).
Table 5
R2 Values
of the Learning Process with Each Element per Data Seta
element
MSE (-)
MAE (-)
AAE (-)
ABE (-)
total R2 (-)
carbon
2.005
1.019
0.013
0.001
0.964
hydrogen
0.010
0.078
0.014
0.001
0.928
oxygen
1.921
1.035
0.057
–0.009
0.965
Average absolute error (AAE), average
bias error (ABE), mean absolute error (MAE), and mean square error
(MSE).
Total R2 for each element across 104
data sets, (A: carbon, B: hydrogen, and C: oxygen).Average absolute error (AAE), average
bias error (ABE), mean absolute error (MAE), and mean square error
(MSE).Next, the relative impact deduction of the ultimate
analysis approximation
is discussed. The optimal function was chosen based on performance
indexing, and the relative impact was calculated, which is shown as
weight per layer and is set with input and output layers. Proximate
properties were designated with different weights from each element.
First, volatile matter and fixed carbon were found to significantly
affect carbon by 33.96 and 46.08%, which is a relatively high impact
when deriving the composition of carbon. Volatile matter and inherent
moisture properties affect the higher values of hydrogen by 33.41
and 49.71%. Volatile matter and inherent moisture also influenced
oxygen by 30.77 and 58.68%. However, some proximate analysis properties
of non-affected elements had a relative impact, owing to the distribution
of data set properties and calculation biases. From the results of
the optimal learning function, the relative impact per element was
calculated based on ANN training (Figure ).
Figure 7
Relative impact of proximate analysis properties
per element. Carbon
(C), fixed carbon (FC), hydrogen (H), inherent moisture (IM), oxygen
(O), and volatile matter (VM).
Relative impact of proximate analysis properties
per element. Carbon
(C), fixed carbon (FC), hydrogen (H), inherent moisture (IM), oxygen
(O), and volatile matter (VM).To improve MSE and MAE results, an additional technique
based on
sample data set choice may be effective after topology optimization,
and a more precise elemental prediction may be available with better
regression performance in the future.
ANN Model Improvement with Coal Rank Division
For classification
and based on ASTM criteria, coal rank is divided with volatile matter
to improve element prediction.[39] The coal
sample in this study was divided into 41 sub-bituminous coal (SBC)
samples containing 42.0–35.1 wt % volatile matter, 49 high
volatile bituminous coal (HVBC) samples containing 35.0–29.0
wt % volatile matter, and 14 medium volatile bituminous coal (MVBC)
samples containing 28.8–24.1 wt % volatile matter.Regarding
the ultimate analysis approximation, ANN regression accuracy based
on several errors (i.e., MSE, MAE, AAE, ABE, and R2) improved with a certain coal rank. Results are shown
in Table . Accuracy
improved with various contents by coal rank, shown as the bold text.
First, in the SBC area, carbon and oxygen prediction (MSE and MAE)
both improved. Hydrogen showed less MAE and ABE improvement, but higher
MSE and AAE. Thus, it could be treated with similar accuracy.
Table 6
Element Approximation Results of Three
Rank Groups of Coal Samplesa
coal rank
element
MSE
MAE
AAE
ABE
R2
sub-bituminous
carbon
0.804
0.636
0.021
–0.004
0.958
hydrogen
0.010
0.077
0.035
0.000
0.679
oxygen
1.54
0.856
0.108
–0.002
0.920
high volatile bituminous
carbon
1.601
0.934
0.024
0
0.956
hydrogen
0.008
0.066
0.025
–0.005
0.882
oxygen
1.716
0.897
0.13
0.006
0.96
medium
volatile bituminous
carbon
2.327
1.127
0.097
0.006
0.875
hydrogen
0.016
0.073
0.109
0.006
0.904
oxygen
1.781
0.870
0.551
0.111
0.898
Average absolute error (AAE), average
bias error (ABE), mean absolute error (MAE), and mean square error
(MSE).
Average absolute error (AAE), average
bias error (ABE), mean absolute error (MAE), and mean square error
(MSE).In the HVBC area, all carbon, hydrogen, and oxygen
results showed
less MSE, MAE, and ABE improvement, but similar AAE. Thus, most of
the results show that this ANN is highly optimized for HVBC. This
may be caused by the HVBC-centered fitting providing a higher count
of coal samples.In the MVBC area, oxygen regression prediction
had lower MSE and
MAE, but carbon results showed a higher error. The hydrogen result
showed higher MSE and less MAE; thus, there was not much improvement
for the total coal analysis for hydrogen, but it was more accurate
for oxygen.Then, the current model’s accuracy was compared
with those
found in the literature, also provided as Table . The first focus point was about error improvement.[9] Under certain performance comparisons, accuracy
between reference and present studies was derived, ranging from HVBC
to SBC. The range of the AAE varied by up to 100 times depending on
the element; thus, the AAE is calculated as a percentage based on
reference data (Figure ). With an optimal topology derived using an activation function
and hidden layer diversity, the AAE improved by at least 0.486 to
at most 11.742%. On the other hand, based on R2, regression improved by at least 0.026 to at most 0.29. This
accuracy improvement represents an effective improvement in fidelity
that can be derived from elemental calculation division, even based
on the same coal rank division.[9]
Table 7
Comparison of ANN Model Accuracy with
Previous Research Studiesa
Yi et
al.
present
study
improvement rate (%)
convergence rate (−)
coal rank
ABE
AAE
R2
ABE
AAE
R2
AAE
R2
high volatile bituminous
C
0.00
0.51
0.93
0.000
0.024
0.956
4.71
0.026
H
–0.01
2.52
0.83
–0.005
0.025
0.882
0.99
0.052
O
–1.39
9.72
0.67
0.006
0.130
0.960
1.34
0.290
sub-bituminous
C
–0.01
0.93
0.86
–0.004
0.021
0.958
2.26
0.098
H
–0.07
2.16
0.61
0.000
0.035
0.679
1.62
0.069
O
–1.90
11.85
0.71
–0.002
0.108
0.920
0.91
0.210
Reprinted by permission from Yi,
L.; Feng, J.; Qin, Y.-H.; Li, W.-Y. Prediction of elemental composition
of coal using proximate analysis. Fuel2017,193, 315–321. under a Creative Commons Attribution-Noncommercial
4.0 Unported License. Copyright 2017 Elsevier. Average absolute error:
AAE and average bias error: ABE.
Figure 8
Comparison
of AAE and R2 with results from previous
research data. Reprinted by permission from Yi, L.; Feng, J.; Qin,
Y.-H.; Li, W.-Y. Prediction of elemental composition of coal using
proximate analysis. Fuel2017,193, 315–321. under a Creative Commons Attribution-Noncommercial
4.0 Unported License. Copyright 2017 Elsevier. High volatile bituminous
coal: HVBC and sub-bituminous coal: SBC.
Comparison
of AAE and R2 with results from previous
research data. Reprinted by permission from Yi, L.; Feng, J.; Qin,
Y.-H.; Li, W.-Y. Prediction of elemental composition of coal using
proximate analysis. Fuel2017,193, 315–321. under a Creative Commons Attribution-Noncommercial
4.0 Unported License. Copyright 2017 Elsevier. High volatile bituminous
coal: HVBC and sub-bituminous coal: SBC.Reprinted by permission from Yi,
L.; Feng, J.; Qin, Y.-H.; Li, W.-Y. Prediction of elemental composition
of coal using proximate analysis. Fuel2017,193, 315–321. under a Creative Commons Attribution-Noncommercial
4.0 Unported License. Copyright 2017 Elsevier. Average absolute error:
AAE and average bias error: ABE.Model improvement was also compared. When the MAE
values derived
from this study were compared to those calculated by multiple linear
regression (MLR), ANN, and adaptive-network-based fuzzy inference
system (ANFIS) models from the literature, it was found that our ANN
was not only more precise but also competitive with the ANFIS model,
which is shown as Figure .[10] Furthermore, the MAE value
for the prediction of carbon content from the proposed ANN model was
approximately half that of the previous ANN result, indicating that
our model can reach a good near-ANFIS level of prediction using an
optimization process that employs the appropriate topology and activation
functions. However, the prediction of the hydrogen content of different
ranks based on the amount of volatile matter content was affected
more than that of carbon. Although the MAE value for MVBC was higher,
those for both SBC and HVBC dropped, indicating that regression accuracy
depends on coal rank when hydrogen content is predicted.
Figure 9
Comparison
of the MAE with that in previous research studies. Adaptive-network-based
fuzzy inference system (ANFIS), artificial neural network (ANN), carbon
(C), hydrogen (H), high volatile bituminous coal (HVBC), multiple
linear regression (MLR), medium volatile bituminous coal (MVBC), oxygen
(O), and sub-bituminous coal (SBC).
Comparison
of the MAE with that in previous research studies. Adaptive-network-based
fuzzy inference system (ANFIS), artificial neural network (ANN), carbon
(C), hydrogen (H), high volatile bituminous coal (HVBC), multiple
linear regression (MLR), medium volatile bituminous coal (MVBC), oxygen
(O), and sub-bituminous coal (SBC).As a result of oxygen tracing, the MAE value for
HVBC was reduced
to 0.915, and other ranges of the data set showed improvements. Therefore,
under certain coal rank divisions, performance improvements can become
similar to ANFIS based on regression accuracy.[7,40,41] In Table , the improved data are shown.
Table 8
Comparison of Model Accuracy with
Various Models from Previous Research Studiesa
carbon
hydrogen
oxygen
MAE
AAE
ABE
MAE
AAE
ABE
MAE
AAE
ABE
Shen
14.141
0.25
–0.25
2.011
0.317
–0.113
12.6
2.263
2.263
Parikh
13.151
0.233
–0.233
1.776
0.248
–0.181
12.423
2.205
2.205
Nhuchhen
6.698
0.124
–0.124
5.739
0.894
0.848
7.615
1.017
1.017
Lawal et
al.
MLR
2.658
0.055
0.005
0.439
0.076
0.005
2.307
0.205
0.043
ANN
1.317
0.027
–0.002
0.328
0.055
–0.355
1.353
0.123
–0.038
ANFIS
0.45
0.008
0.008
0.122
0.014
0.014
0.988
0.06
–0.01
present study
SBC
0.636
0.021
–0.004
0.077
0.035
0.000
0.856
0.108
–0.002
HVBC
0.934
0.024
0.000
0.066
0.025
–0.005
0.073
0.109
0.006
MVBC
0.856
0.108
–0.002
0.897
0.13
0.006
0.87
0.551
0.111
Average absolute error (AAE), average
bias error (ABE), adaptive-network-based fuzzy inference system (ANFIS),
artificial neural network (ANN), sub-bituminous coal (SBC), high volatile
bituminous coal (HVBC), mean absolute error (MAE), multiple linear
regression (MLR), and medium volatile bituminous coal (MVBC).
Average absolute error (AAE), average
bias error (ABE), adaptive-network-based fuzzy inference system (ANFIS),
artificial neural network (ANN), sub-bituminous coal (SBC), high volatile
bituminous coal (HVBC), mean absolute error (MAE), multiple linear
regression (MLR), and medium volatile bituminous coal (MVBC).
Relative Impact Deduction
Relative impacts calculated
from SBC, HVBC, and MVBE are shown in Figure . The method of determining the relative
impact of coal composition from the proximate analysis was proposed
as a novel method based on ANN hidden layer weight. As each result
is based on the total data set, the trends of relative impacts fall
within a similar range. For example, the impact of volatile matter
on hydrogen content gradually reduces as the volatile matter content
decreases from 53.41 (SBC) to 23.01 (MVBC). The relative impact of
fixed carbon on carbon increases with HVBC, which had the best approximation
result (also treated as the least error) from the total data set and
other results. Through the derivation of this relative impact, it
can be confirmed that the relative impact ratio of proximate analysis
on elements for each coal rank is different. An ANN prediction of
chemical analysis based on this result can predict that coal rank
division may affect the input data of the overall ANN (in this case,
elemental composition). It has been seen that as the content of the
volatile material decreases, the relative effect of the fixed carbon
on carbon increases, and the relative effect of the volatile material
on hydrogen decreases. If further research is conducted based on these
results, it is suggested that the effect on each element may lead
to a result that further emphasizes the importance of the criteria
for determining rank depending on the results of the proximate analysis.
Figure 10
Relative
impact result with each data set. Carbon (C), fixed carbon
(FC), hydrogen (H), high volatile bituminous coal (HVBC), inherent
moisture (IM), medium volatile bituminous coal (MVBC), oxygen (O),
sub-bituminous coal (SBC), and volatile matter (VM).
Relative
impact result with each data set. Carbon (C), fixed carbon
(FC), hydrogen (H), high volatile bituminous coal (HVBC), inherent
moisture (IM), medium volatile bituminous coal (MVBC), oxygen (O),
sub-bituminous coal (SBC), and volatile matter (VM).
Limitations and Improvement of ANN Application through FT-IR
Experiments
In the case of such access through the ANN, the
same output can be obtained for the same input. As a representative
example, for similar proxy properties, the operation will proceed
to show similar ultimate analysis results. Through the comparison
of analysis results through FT-IR spectroscopy, we can present different
approaches to these predictions and present directions in ANN-based
elemental analysis (Figure ).
Figure 11
FT-IR spectra of two representative samples—YP
and CP.
FT-IR spectra of two representative samples—YP
and CP.First, FT-IR analysis was performed according to
the technique
and sample described in the Methods section. Before the analysis procedure,
the point where YP is dominant was expressed as a red line, and the
point where CP is dominant was expressed as a blue line, and then,
the composition of the point corresponding to such wavelength was
confirmed and analyzed. These two samples had similar elemental compositions,
as seen in the previous methodological explanation, but the proximate
analysis results were very different, 10.84 and 17.44 wt % in terms
of volatiles and 80.9 and 73.7 wt % in terms of fixed carbon, respectively.
As shown in the previous description, it may be difficult to predict
these differences in analysis values in ANN-based calculations. We
compared the analysis result of the FT-IR spectrum with each range
based on the previous experiment to discuss further approach (Figure ).[42]
Figure 12
IR spectrum between 3200 and 1200 cm–1 wavenumbers
of YP and CP samples.
IR spectrum between 3200 and 1200 cm–1 wavenumbers
of YP and CP samples.The aliphatic properties and elemental arrangements
of each coal
can be confirmed through quantitative FT-IR analysis in the 3200–1200
cm–1 section. The N–H stretching group corresponding
to the 3000–2800 cm–1 section corresponds
to the aliphatic structure, and the aliphatic structure group of CP
coal is larger according to the criteria, and thus, the N–H
group is formed. Considering that such an aliphatic structure is broken
as proximate analysis progress upon thermogravimetry analysis (TGA)
and measured as a volatile material, it can be predicted that the
CP sample will have a higher overall volatile content.[43] Also, the unsaturated ketone range of 1620–1610
cm–1 was measured to be high with the CP sample
compared to that with the YP sample, which is about the C=C
stretching group, which can also be treated as a volatile matter released
on the TGA experiment, supporting this analysis. On the other hand,
the analysis in the 1310–1250 range shows that the composition
of C–O stretching represented by aromatics ester is prominent
in YP, and assuming that these aromatics are decomposed to be delayed
during proximate analysis, it can be also confirmed that the content
of volatile matter will be low in CP. It can also be seen that the
difference in the trend of aliphatic–aromatic appears in the
1600–1800 cm–1 section (Figure ).[15]
Figure 13
IR spectrum between 1250 and 500 cm–1 wavenumbers
of YP and CP samples.
IR spectrum between 1250 and 500 cm–1 wavenumbers
of YP and CP samples.This trend is also noticeable as a contrast in
the lower wavenumber
range (aromatic regions), which can be expected to have affected the
content of the aromatic cluster and could be in proportion to the
lower volatile matter content. As seen in the C–O stretching
ester in 1210–1163 cm–1 and CO–O–CO
anhydride in 1050–1040 cm–1, the aromatic
content is considerable with the YP sample, which has lower volatile
matter content. The lower region of less wavenumber count of 895–885
cm–1 of C=C bending alkene, 840–790
cm–1 of C=C bending alkene, and 690–515
cm–1 of the C–Br stretching halo compound
also supports this logical assumption.Based on these brief
analysis results, calculations of elemental
analysis based on the ANN confirm the importance when the experimental
basis for the target physical properties is supported. In other words,
it may be based on the thermal–chemical criteria for analysis
of target properties, or it may be possible to overcome limitations
and provide complementary information through applicability, which
is a unique nature of the ANN.
Conclusions
This research resulted in the development
of a numerical calculation
method to measure the relative impact of proximate analysis properties
on elemental composition detection using an optimized ANN and was
applied to various thermal coals used for coal-fired power plants
in South Korea. The ANN was optimized with a variety of activation
functions and hidden layers via performance evaluation using the MSE, R2, and epoch. The L–M function was chosen
as the optimal function for calculating relative impact. To estimate
elemental composition from proximate analysis properties, a relative
impact mechanism was used based on the ANN hidden layer weight with
accurate measurements of the MSE, MAE, AAE, ABE, or R2 for comparison with previous research.Performance
was improved by analyzing the coal rank division per
element. Each performance index was then compared to estimate numerical
accuracy, and the differences were measured against several prior
ANN elemental composition tracing models for comparison. As a result,
performance improved over previous results by 4.71–0.91%. Moreover,
based on the MAE, our results nearly reached ANFIS levels while improving
over previous models by 5.40 and 7.39%. Then, we derived the relative
impact from proximate properties to the elemental composition. The
result was varied with each coal rank, so this result suggests future
research to apply each input parameter preprocessed with its own standard
for further application. Also, FT-IR spectroscopy analysis was derived
for further application of the ANN to an elemental composition-based
approach. As a larger amount of aliphatic may be contained in the
coal sample, even the same element analysis result may produce diverse
volatile matter content. Thus, it needs to be considered that the
calculation of element composition based on the ANN needs a basis
for the target data.In summary, this study improved the ANN-based
relative impact approach
to allow the derivation of thermal coal elemental composition using
proximate analysis properties with the optimized neural network topology
and a suitable activation function.
Table 9
Elemental and Proximate Properties
of Each Coal Sample Used for ANN Development
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9
Figure 10
Figure 11
Figure 12
Figure 13