Академический Документы
Профессиональный Документы
Культура Документы
PII: S0360-5442(18)31895-4
DOI: 10.1016/j.energy.2018.09.131
Reference: EGY 13825
Please cite this article as: Mutlu AY, Yucel O, An artificial intelligence based approach to predicting
syngas composition for downdraft biomass gasification, Energy (2018), doi: https://doi.org/10.1016/
j.energy.2018.09.131.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
An Artificial Intelligence Based Approach to Predicting Syngas Composi-
tion for Downdraft Biomass Gasification
Ali Yener Mutlu1*, Ozgun Yucel2
1
Department of Electrical and Electronics Engineering, Izmir Katip Celebi University, Izmir, Turkey
2
Department of Environmental Engineering, Izmir Katip Celebi University, Izmir, Turkey
*
aliyener.mutlu@ikc.edu.tr
PT
ABSTRACT: Artificial neural networks and artificial intelligence based regression techniques have been recently applied to
various gasification processes. Although these techniques obtain relatively satisfactory results for predicting gasification
products, most of the proposed models are prone to low number of samples in the training data sets, which also lead to over-
RI
fitting problem. Furthermore, these models may fall into local minima since cross-validation has never been used for predict-
ing gasification products. In this paper, we consider prediction of gasification products as a classification problem by using
machine learning classifiers. Two types of classifiers have been proposed, i.e., binary least squares support vector machine
and multi-class random forests classifiers, for predicting producer gas composition and its calorific value obtained by woody
SC
biomass gasification process in a downdraft gasifier. The proposed approaches have been developed and tested with 5237
data samples using 10-fold cross-validation, where binary and multi-class classifiers achieved over 96% and 89% prediction
accuracy values, respectively.
U
1. INTRODUCTION Limited numbers of kinetic mathematical models, which
evaluate possible homogeneous and heterogeneous
AN
As a result of increasing energy demand and environmental reactions, have been developed for evaluating the effect of
awareness, distributed and decentralized electricity operating conditions on syngas composition and calorific
generation is crucial and gaining importance. Conventional value during gasification process [3], [4]. However, these
power stations are centralized and often require electric mathematical models deal with transport (heat, mass and
momentum) and kinetic equations leading to extensive
M
role for the electric power distribution system. modeling ‘reacting flow systems’ such as combustion
As a promising technology, gasification is an effective chemical reactors and other mixing processes [5], [6]. So
CFD and thermodynamic equilibrium models with limited
TE
fuel or can be further processed to produce electricity and Thermodynamic equilibrium approach is one of the most
to generate heat. Furthermore, gasification converts low- widely studied modeling techniques for simulating biomass
value feedstocks into valuable forms of energy since it gasification and can be divided into two sections: the
enables energy production from non-conventional feeds stoichiometric and non-stoichiometric equilibrium
C
like forest waste, agricultural waste, poultry waste, and methods. The non-stoichiometric method is based on Gibbs
municipal solid waste [2]. free energy minimization [9]–[11]. On the other hand, in
AC
Fixed bed gasifiers that produce syngas are the oldest and the stoichiometric method, the equilibrium is determined
widely studied reactors due to their easy construction and by using the equilibrium constants for some particular
simple operation [3]. The fixed bed gasification systems are reactions including water gas shift and methanation [12]–
classified as updraft, downdraft and crossdraft gasifiers. [15]. Both techniques are simple and fast but the
Downdraft gasifiers have the advantage of yielding lower equilibrium condition in the gasifiers is never achieved,
tar content (1-2 g/Nm3) compared to other fixed bed especially for fixed bed gasifiers.
gasifiers. For this reason, they do not require high cost gas Artificial neural networks (ANN) are widely used for
cleaning units and are suitable for small-scale applications. signal processing, function approximation, simulation, and
In a downdraft gasifier, it has been proved that a number of recognition of patterns [16]. However, ANNs have been
chemical reactions take place while solid fuels move rarely used for biomass gasification since they require data
together with air in downward direction. Mathematical sets consisting of large number of samples [17]. Hence,
models that describe dynamic behavior of fixed bed contrary to fluidized beds, the use of ANNs for modeling
gasifiers can facilitate designing cost effective processes to downdraft biomass gasification is very limited [18]–[20].
reduce time and expense of complicated and time
consuming experiments.
ACCEPTED MANUSCRIPT
In addition to the ANNs, other artificial intelligence (AI) 2. EXPERIMENTAL SETUP
based approaches have been developed for predicting
system outputs effectively by learning and mining system The experiments were conducted on a 10 kW downdraft
characteristics from limited experimental data [21]. fixed-bed gasifier developed by All Power Labs Inc.
Therefore, AI based techniques can contribute discovering Experimental system design for the modular gasification is
underlying rules and characteristics of downdraft biomass shown in Fig. 1. In a downdraft gasifier, the solid fuels
gasification. For instance, the performance of least-squares move together with air in the downward direction. The air
support vector machine (LS-SVM) approach is found better suction provided by venturi ejector enables air to enter the
than the ANN approach in the prediction of biochar yield reactor via an air flow meter. The biomass undergoes
from cattle manure pyrolysis [22]. However, AI based several different processes while passing through four
classifiers and regressors except for the ANNs have never different zones based on operation: drying zone, pyrolysis
PT
been used in the prediction of syngas composition. In the zone, combustion zone, and reduction zone. Each stage
literature, input parameters of ANN models include involves a specific combination of complex homogeneous
biomass moisture (MC), biomass content of ash, C, H and and heterogeneous chemical reactions.
O, gasification temperature (Tg) and equivalence ratio (ER) Produced syngas is cleaned through cyclone separator to
RI
[18]–[20], [23]–[25]. However, in our preliminary studies remove the particles and other condensable materials are
we found that these parameters, which depend on fuel cleaned through the packed bed filter and cooled further.
composition, did not change the accuracy of predicting Finally, the flue gas is burnt out with a swirl burner. A
gasification products using the proposed AI based small gas stream before the burner is brought via a small
SC
approaches. Consequently, the temperature distribution has clean-up unit to a gas analyzer. A portable infrared syngas
more influence on the syngas composition compared to the analyzer, Wuhan Cubic Syngas Analyzer Gas board 3100P,
fuel parameters. Therefore, in this study we prefer to use was used for the measurement of the concentrations of CO,
temperature distribution, equivalence ratio (ER) and fuel CO2, CH4, H2 and O2 in the sample gas simultaneously
U
flow rate (FR) parameters to train machine learning while heating value (calorific value) is calculated
classifiers. automatically by using NDIR and Thermal Conductivity
AN
In this paper, due to being capable of deciphering subtle Detector. The heating value (HHV) is the amount of heat
patterns in noisy and complex datasets, we employ random that is released when a unit mass of fuel is fully combusted.
forests and support vector machine (SVM) based classifiers
to predict syngas composition that is released during the
gasification process. The calorific values with discrete
M
Real
25 Quantized
AC
Quantization Error
20
Concentration (%)
15
10
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Number of Data Points
Figure 2. Concentration values, quantized levels, i.e., classes, and quantization errors for H2 collected from 5 different experiments.
ACCEPTED MANUSCRIPT
During the experiments, temperatures at 6 different heights quantization error is negligible. Furthermore, the feature
inside the gasifier were measured as T0, T1, T2, T3, T4 , T5 vectors for all samples are assigned class labels such that
using a group of 6 in-line K-type thermocouples with 8 mm training data sets consisting of 37, 20, 13, 24 and 19 dis-
diameter. Then real-time values of temperatures, pressures tinct classes are obtained corresponding to distinctive and
and flow rates were monitored/recorded using an Arduino discrete CO, CO2, CH4, H2 and HHV levels, respectively.
based system shown in Fig. 1. However, the collected
3.2. Preprocessing
temperature and pressure values are noisy and fluctuate due
to the process dynamics. Although, there are instantaneous Once the extracted features are assigned labels to define
differences due to channelling and bridging phenomena, classes for the discrete levels of CO, CO2, CH4, H2 and
the gas composition for the same temperature values and HHV, machine learning classifiers will employ distance
the same air fuel ratios should be fixed. Therefore, it is measures to compute similarities and dissimilarities among
PT
expected that the average value and trend of the gas these classes to predict biomass gasification products. For
composition will be the same when the experiment is this purpose, distances between feature vectors need to be
repeated. For this reason, it can be deduced that the computed by using measures such as Euclidean (L2), city-
repeatability of the experimental processes in this study is block (L1) and Mahalanobis distance measures [26].
RI
expected to be high. However, feature vectors consisting of multiple elements
generally lie in a very high dimensional space and if any of
The recorded airflow rate values were used in calculation
the features has broad range of values, the distance
of Equivalence Ratio (ER), which is a crucial operating
calculated in the high dimensional space will be dominated
variable in biomass gasification, and expresses the amount
SC
by this particular feature. Hence, the features should be
of external air supplied to the gasifier. C, H and O contents
scaled such that each feature equally governs the distance
of biomass are calculated with elemental analysis of
measure that is used for classification. In this paper, we
biomass. Among the features that are used in our study, the
− min
employ linear scaling to unit range:
=
measurement uncertainty for the carbon (C) is ≤ 0.45% and
max
− min
U
for the hydrogen (H) is ≤ 0.10%. For the oxygen, the (1)
element of the feature vector for the nth sample, min
AN
and FR, it is ≤ ±1%. In addition, the measurement range for
lence ratio (ER), gasification temperature and pressure are 3.3.1. Random Forests
important parameters that significantly affect gasification. For the development of multi-class aggregated decision
As we introduced in section 2, the gasification experiment trees, ensemble-learning methods such as boosting and
has input variables listed as C, H, O, T0, T1, T2, T3, T4 , T5 , bagging have been widely studied [27]. Since the decision
C
ER and FR. However, C, H and O that come from ele- tress are known to be sensitive to small disturbances in the
mental analysis from biomass involve redundant infor- given training set, random forests (RF) classifier has been
mation since the changes in these variables are either con- proposed by Breiman by using bagging, which employs
AC
stant or too low to discriminate between different classes, multiple bootstrapped training sets of decision trees to
i.e., discrete levels of syngas composition. Therefore, we increase randomness [28]. RF classifier splits each node
sional feature vector, , for the nth sample in the data set.
predictors at that particular node. Studies have shown the
effectiveness of RF in classifying multi-class hyper-
spectral data [29], modelling language [30], which consists
In order to construct classes corresponding to discrete lev-
of identifying large number of classes, and diagnosing
els of gasification products for the CO, CO2, CH4, H2 and
Alzheimer’s disease [31]. In this study, we propose a RF
HHV, we quantize concentration values to the nearest inte-
classification based approach consisting of 100 decision
gers prior to classification. For instance, we illustrate con-
trees for the prediction of gasification product levels.
centration values and corresponding quantized levels for
the H2 obtained from the five experiments in Fig. 2, where 3.3.2. Least-Squares Support Vector Machine
vertical dashed lines separate different experiments and
ACCEPTED MANUSCRIPT
The support vector machine (SVM) is a supervised learning once. In this paper, we use 10-fold cross-validation since
technique that forms an optimal hyperplane to separate data the most unbiased generalization error for machine learning
belonging to different classes [32]. While conventional problems is achieved when k is equal to ten [38].
classifiers, such as the ANNs, are prone to local minima
4. RESULTS
existence due to gradient descent learning, the SVM
exploits inequality type constraints to optimize quadratic
The performance of the proposed machine learning ap-
function of variables. On the contrary, the least squares
proaches using the RF and the LS-SVM classifiers in pre-
formulation of SVM, referred to as LS-SVM, computes
dicting discrete levels of gasification products is evaluated
separating hyperplane by solving a set of linear equations
by computing commonly used evaluation measures for
with equality type constraints [33]. Furthermore, the LS-
binary classification referred to as classification accuracy
SVM offers relatively fast model computation using
HIJ: + HK+
PT
Lagrangian multipliers [34]. Thus, in this paper we also
G = × 100
HIJ: + HN+g + MIJ: + MK+
predicting gasification product levels using the LS-SVM (7)
HIJ:
The SVM classifies an N – dimensional test input, , into I=
classifier.
HIJ: + MIJ:
RI
(8)
HIJ:
= sign +
N =
one of two separate classes by defining a decision function:
HIJ: + MK+
where maps the input space into a higher dimensional
(2) (9)
SC
and is a bias term [35]. In order to compute the I×N 10
M1 = 2 ×
parameters and , the LS-SVM solves the optimization I+N
(
U
problem: where TPos and TNeg represent the amount of correctly
.
!,",#$ 2 2
(3) respectively. On the other hand, FPos and FNeg
AN
,/0
correspond to the amount of incorrectly predicted false
positive and false negative events, respectively. The
1,
, +
= 1 − +, , = 1,2, … , 3
with equality constraints
accuracy, precision, recall and F1-score values statistically
, , 1, $.
forms 3 training input-output pairs,
(4) assess the performance of a classifier by computing the
,/0
1, = ±1 represents the class label of , and & =
M
+0 , +- , … , +.
. Using the Lagrangian multipliers 5 =
classifier.
60 , 6- , … , 6.
,
Prediction of Gasification Outcomes Using RF and LS-SVM Classifiers
100
.
95 RF
D
90 LS-SVM
75
,/0
70
+ +, $
(5)
TE
65
60
55
50
45
40
the LS-SVM classifier is defined as [36]:
.
35
= : ; <) 6, 1, =
, + >
,
30
25
20
EP
15
(6)
,/0
10
B ∥C
Gasification Outcomes
A
,
, = + CDC , kernel
∥
radial basis function (RBF), =
?
Figure 3. Prediction of syngas composition using Random Forest (RF)
3.3.3. Cross-Validation
Once the extracted and scaled features are given to the Precision is defined as the proportion of true positive
proposed classifiers, one can divide the entire data set into events to the events classified as positive in a classification
training and testing data sets to analytically measure the task, whereas recall measures the proportion of true
classification accuracy. For this purpose, cross-validation positive events to the sum of true positive and false
techniques, such as the holdout, leave-one-out or k-fold negative events. Furthermore, F1-score utilizes both the
methods can be employed [37]. However, the k-fold cross- precision and the recall to obtain a single score, F1, with
validation is the most preferred approach that randomly the range [0,1], which results in 1 for perfect precision and
PT
accuracy over 90%. rectly identified and how many instances were confused
with other classes [40]. Each column of the matrix repre-
Confusion Matrix Using LS-SVM for H 2 (A = 96.3911 %) sents corresponding predicted class label whereas each row
N/A
represents actual class label. Fig. 4 (a) and (b) show the
RI
6 90
7
8
confusion matrices obtained by the LS-SVM and the RF
Actual Class Labels (Concentration %)
80
9
11
classifiers, respectively, for predicting H2 levels. The rest
70
12
13
of the confusion matrices computed for the CO, CO2, CH4
60
and HHV are presented in the supplementary figures due to
SC
14
15
16 50 space limitation.
17
18 40
19
20 30 Table 1. Precision, Recall and F1-score values for the prediction
21
22 20
of gasification products, CO, CO2, CH4, H2 and HHV, using the RF
U
23
24
and the LS-SVM classifiers.
10
25
26
0
AN
N/A 6 7 8 9 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Predicted Class Labels (Concentration %) CO CO2 CH4 H2 HHV
(a)
RF RF RF RF RF
Confusion Matrix Using RF for H 2 (A = 83.3302 %)
6
LS- LS- LS- LS- LS-
80
8
9
11 70
50
20
21
R
± 0.44 ± 0.36 ± 0.24 ± 0.32 ± 0.28
22 20
23
24
Table 1 shows mean and standard deviation values of the As one can see from the matrices, majority of the predicted
precision, recall and F1-scores provided by both classifiers. samples from each gasification product lies on the diago-
AC
One should note that the precision values are either equal nals of the confusion matrices, which means that the gasifi-
or greater than the recall values for each classifier, which cation product levels are successfully predicted with accu-
indicates that the amount of false negative events is greater racy over 83% and 96%, in particularly for the H2, using
than the amount of false positive events. Furthermore, the the RF and the LS-SVM, respectively. We obtain similar
LS-SVM performs better than the RF in predicting results for the remaining gasification products as seen in
gasification products since it provides higher P and R the supplementary figures. Furthermore, we observe from
values for each product, which results in higher F1-scores. the confusion matrices that misclassification occurs espe-
In addition, the LS-SVM is able to achieve F1-scores over cially for the tails of the matrix diagonals, i.e., for the low
0.8 for the CH4, H2 and HHV, whereas the RF could and high values of class labels, which is consistent with the
provide F1-score over 0.8 only for the CH4. The distribution of the data set. One can see from the Fig. 5,
performance of both classifiers in predicting the products, describing the histograms of the datasets for each gasifica-
CO and CO2, seems to decrease due to the fact that they tion product, that low and high values of class labels consti-
could achieve P, R and F1-score values that are less than
ACCEPTED MANUSCRIPT
tute only a minority portion of the whole data set and the performance of the equilibrium models decreases with ER
proposed machine learning classifiers’ performance slightly value and gasification temperature. On the contrary, the
diminish only for this minority portion. performance of the proposed machine learning classifiers
does not decrease with the ER and gasification temperature
800
Histogram of the data sets since these methods are not based on any chemical reaction
700 and they do not require reaching equilibrium conditions.
CO
600
CO2
Table 2. RMSE and NRMSE values for the prediction of gasification
500
Frequency
CH4
H
products, CO, CO2, CH4, H2 and HHV, using the RF and the LS-SVM
400 2
HHV
classifiers.
300
PT
100
RF RF RF RF RF
0
0 5 10 15 20 25 30 35 40 45
Concentration (%) and HHV (MJ/m 3)
LS- LS- LS- LS- LS-
SVM SVM SVM SVM SVM
RI
Figure 5. Histograms of the datasets for each syngas composition and 5.1443 2.8069 1.7960 5.0412 2.3283
HHV values RMSE
1.5196 0.6403 0.3482 0.5949 0.3803
Lastly, the samples that are predicted inaccurately are often 0.4126 0.2917 1.5860 0.3507 0.2453
SC
NRMSE
classified into neighboring classes, leading to very low 0.2403 0.0778 0.2049 0.0587 0.0477
values of root mean square error (RMSE) and normalized
RMSE (NRMSE) values that are computed as:
5. CONCLUSION
,/01
∑Z , − 1, -
N3VW = X
3
U
(11)
In this paper, we proposed binary and multi-class machine
learning classifiers to predict discrete levels of syngas
AN
composition. The proposed classifiers are fed with the
selected features, T0, T1, T2, T3, T4, T5, ER and FR, from
1 1, − 1,
. -
five different experiments with 5237 samples in total to
KN3VW = [ ) \ ]
3 1,
predict CO, CO2, CH4, H2 and HHV values. Both the LS-
M
,/0
(12) SVM and the RF classifiers outperform stoichiometric and
non-stoichiometric models that have previously been used
label and M is the total number of data samples, which is classifier as the LS-SVM is able to achieve classification
equal to 5237 in our study. As one can see from the Table accuracy over 96%. The computed precision, recall, F1-
TE
2, the LS-SVM provides lower RMSE and NRMSE values score, RMSE and NRMSE values prove that the proposed
compared to the RF classifier. However, both classifiers techniques are great candidates for automatically
achieve very low NRMSE values, in particularly for the estimating levels of gasification products accurately, in
CO2, H2 and HHV. On the other hand, previous work uti- contrast to stoichiometric and non-stoichiometric models.
lized stoichiometric and non-stoichiometric models for
EP
RMSE results for the high heating values provided by the the temperature distribution is continuously measurable and
stoichiometric and non-stoichiometric models were com- controllable with the right methods such as electrical
AC
puted as 1.53 and 0.87, respectively. Therefore, the RF and heaters and fuel/air injectors. Therefore, this study offers
LS-SVM classifiers perform better than model-oriented more flexibility in controlling and optimizing composition
approaches in predicting gasification products. Moreover, of the syngas from woody biomass gasification.
equilibrium models have an inherent tendency to underes-
timate yields of methane and overestimate yields of hydro- The future work will focus on machine learning regression
gen [41]. On the other hand, machine learning based tech- techniques for the prediction and control of syngas
niques do not have such a tendency since they do not in- composition obtained by downdraft biomass gasification.
clude any parameter related to the chemicals or chemical We believe that machine learning algorithms could allow
reactions. Moreover, both the stoichiometric and non- researchers and professionals to optimize syngas
stoichiometric equilibrium models suppose that chemical composition without using models that are prone to the
system is on equilibrium. But for low ER values and gasifi- equilibrium condition.
cation temperatures, gasification products leave the gasifier
without reaching the equilibrium condition. Therefore, the ACKNOWLEDGEMENTS
ACCEPTED MANUSCRIPT
This work was supported in part by the Scientific Research [17] T. K. Patra and P. N. Sheth, “Biomass gasification models for
Projects Coordinatorship of Izmir Katip Celebi University downdraft gasifier: A state-of-the-art review,” Renew.
Sustain. Energy Rev., vol. 50, pp. 583–593, 2015.
by grant 2018-GAP-MÜMF-0009. [18] R. Mikulandrić, D. Lončar, D. Böhning, R. Böhme, and M.
Beckmann, “Artificial neural network modelling approach
for a biomass gasification process in fixed bed gasifiers,”
Energy Convers. Manag., vol. 87, no. 0, pp. 1210–1223, 2014.
AUTHOR INFORMATION [19] R. Mikulandrić, D. Böhning, R. Böhme, L. Helsen, M.
Beckmann, and D. Lončar, “Dynamic modelling of biomass
Corresponding author gasification in a co-current fixed bed gasifier,” Energy
*Email: aliyener.mutlu@ikc.edu.tr Convers. Manag., vol. 125, pp. 264–276, 2016.
[20] D. Baruah, D. C. Baruah, and M. K. Hazarika, “Artificial neural
network based modeling of biomass gasification in fixed bed
REFERENCES downdraft gasifiers,” Biomass and Bioenergy, vol. 98, 2017.
PT
[1] G. Cau, V. Tola, and A. Pettinau, “A steady state model for [21] D. S. Pandey, I. Pan, S. Das, J. J. Leahy, and W. Kwapinski,
predicting performance of small-scale up-draft coal “Multi-gene genetic programming based predictive models
gasifiers,” Fuel, vol. 152, pp. 3–12, 2015. for municipal solid waste gasification in a fluidized bed
[2] E. Azzone, M. Morini, and M. Pinelli, “Development of an gasifier,” Bioresour. Technol., vol. 179, 2015.
equilibrium model for the simulation of thermochemical [22] H. Cao, Y. Xin, and Q. Yuan, “Prediction of biochar yield from
RI
gasification and application to agricultural residues,” Renew. cattle manure pyrolysis via least squares support vector
Energy, vol. 46, pp. 248–254, 2012. machine intelligent approach,” Bioresour. Technol., vol. 202,
[3] O. Yucel and M. A. Hastaoglu, “Kinetic modeling and pp. 158–164, 2016.
simulation of throated downdraft gasifier,” Fuel Process. [23] J. George, P. Arun, and C. Muraleedharan, “Assessment of
SC
Technol., vol. 144, pp. 145–154, 2016. producer gas composition in air gasification of biomass
[4] C. Di Blasi and C. Branca, “Modeling a stratified downdraft using artificial neural network model,” Int. J. Hydrogen
wood gasifier with primary and secondary air entry,” Fuel, Energy, vol. 43, no. 20, pp. 9558–9568, 2018.
vol. 104, pp. 847–860, 2013. [24] Y. Li, L. Yan, B. Yang, W. Gao, and M. R. Farahani, “Simulation
[5] B. Rahmanian, M. R. Safaei, S. N. Kazi, G. Ahmadi, H. F. Oztop, of biomass gasification in a fluidized bed by artificial neural
network (ANN),” Energy Sources, Part A Recover. Util.
U
and K. Vafai, “Investigation of pollutant reduction by
simulation of turbulent non-premixed pulverized coal Environ. Eff., vol. 40, no. 5, pp. 544–548, 2018.
combustion,” Appl. Therm. Eng., vol. 73, no. 1, pp. 1222– [25] S. Sunphorka, B. Chalermsinsuwan, and P. Piumsomboon,
AN
1235, 2014. “Artificial neural network model for the prediction of kinetic
[6] M. R. Safaie, B. Rahmanian, and M. Goodarzi, “Investigation of parameters of biomass pyrolysis from its constituents,” Fuel,
the coal diameter effect on pulverized coal combustion for vol. 193, pp. 142–158, 2017.
pollutant reduction,” J. Math. Comput. Sci., vol. 12, pp. 143– [26] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification.
151, 2014. 2000.
[7] M. Oevermann, S. Gerber, and F. Behrendt, “Euler- [27] a Liaw and M. Wiener, “Classification and Regression by
M
Lagrange/DEM simulation of wood gasification in a bubbling randomForest,” R news, vol. 2, no. December, pp. 18–22,
fluidized bed reactor,” Particuology, vol. 7, no. 4, pp. 307– 2002.
316, 2009. [28] L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp.
[8] T. M. Ismail, M. Abd El-Salam, E. Monteiro, and A. Rouboa, 5–32, 2001.
D
“Eulerian - Eulerian CFD model on fluidized bed gasifier [29] J. S. Ham, Y. Chen, M. M. Crawford, and J. Ghosh,
using coffee husks as fuel,” Appl. Therm. Eng., vol. 106, pp. “Investigation of the random forest framework for
1391–1402, 2016. classification of hyperspectral data,” IEEE Trans. Geosci.
Remote Sens., vol. 43, no. 3, pp. 492–501, 2005.
TE
stoichiometric equilibrium model for the simulation of the [32] V. Vapnik and R. Izmailov, “Knowledge transfer in SVM and
biomass gasification process,” Appl. Energy, 2017. neural networks,” Ann. Math. Artif. Intell., vol. 81, no. 1–2, pp.
[11] J. Han et al., “Modeling downdraft biomass gasification 3–19, 2017.
process by restricting chemical reaction equilibrium with [33] R. Langone, C. Alzate, B. De Ketelaere, J. Vlasselaer, W. Meert,
Aspen Plus,” Energy Convers. Manag., vol. 153, pp. 641–648, and J. A. K. Suykens, “LS-SVM based spectral clustering and
C
systems,” Energy, vol. 130, pp. 86–98, 2017. “Multivariate calibration with least-squares support vector
[13] M. Fani, M. Haddadzadeh Niri, and F. Joda, “A Simplified machines,” Anal. Chem., vol. 76, no. 11, pp. 3099–3105, 2004.
Dynamic Thermokinetic-Based Model of Wood Gasification [35] V. Bajaj and R. Pachori, “Classification of Seizure and Non-
Process,” Process Integr. Optim. Sustain., 2018. seizure EEG Signals using Empirical Mode Decomposition,”
[14] E. Shayan, V. Zare, and I. Mirzaee, “Hydrogen production IEEE Trans. Inf. Technol. Biomed., vol. 16, no. 6, pp. 1135–
from biomass gasification; a theoretical comparison of using 1142, 2011.
different gasification agents,” Energy Convers. Manag., vol. [36] J. A. K. Suykens and J. Vandewalle, “Least Squares Support
159, pp. 30–41, 2018. Vector Machine Classifiers,” Neural Comput., pp. 1–10, 1999.
[15] D. Castello and L. Fiori, “Supercritical water gasification of [37] D. Glotsos et al., “A multi-classifier system for the
biomass: A stoichiometric thermodynamic model,” Int. J. characterization of normal, infectious, and cancerous
Hydrogen Energy, vol. 40, no. 21, pp. 6771–6781, 2015. prostate tissues employing transrectal ultrasound images,”
[16] A. A. A. A. Alrashed, M. S. Gharibdousti, M. Goodarzi, L. R. de Comput. Methods Programs Biomed., vol. 97, no. 1, pp. 53–61,
Oliveira, M. R. Safaei, and E. P. Bandarra Filho, “Effects on 2010.
thermophysical properties of carbon based nanofluids: [38] C. Ambroise and G. J. McLachlan, “Selection bias in gene
Experimental data, modelling using regression, ANFIS and extraction on the basis of microarray gene-expression data,”
ANN,” Int. J. Heat Mass Transf., vol. 125, 2018. Proc. Natl. Acad. Sci., vol. 99, no. 10, pp. 6562–6566, 2002.
ACCEPTED MANUSCRIPT
[39] M. Galar, A. Fernández, E. Barrenechea, H. Bustince, and F. Gasification,” Exergetic, Energ. Environ. Dimens., pp. 323–
Herrera, “An overview of ensemble methods for binary 339, 2017.
classifiers in multi-class problems: Experimental study on
one-vs-one and one-vs-all schemes,” Pattern Recognit., vol.
44, no. 8, pp. 1761–1776, 2011.
[40] S. V. Stehman, “Selecting and interpreting measures of
thematic classification accuracy,” Remote Sens. Environ., vol.
62, no. 1, pp. 77–89, 1997.
[41] E. S. Aydin, O. Yucel, and H. Sadikoglu, “Numerical
Investigation of Fixed-Bed Downdraft Woody Biomass
PT
RI
U SC
AN
M
D
TE
C EP
AC
ACCEPTED MANUSCRIPT
• AI methods estimated syngas composition for downdraft biomass gasification.
• Random Forests and least-squares support vector machine models are developed.
• First study predicting syngas composition using machine learning classification.
• The models have been tested with 5237 data using 10-fold cross-validation.
• The developed LS-SVM classifier provides higher prediction accuracy.
• Temperature distribution significantly affects syngas composition estimation.
PT
RI
U SC
AN
M
D
TE
C EP
AC