Accepted Manuscript

An artificial intelligence based approach to predicting syngas composition for

downdraft biomass gasification

Ali Yener Mutlu, Ozgun Yucel

PII: S0360-5442(18)31895-4
DOI: 10.1016/j.energy.2018.09.131
Reference: EGY 13825

To appear in: Energy

Received Date: 21 July 2018

Revised Date: 15 September 2018
Accepted Date: 19 September 2018

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
AN
As a result of increasing energy demand and environmental
awareness, distributed and decentralized electricity
generation is crucial and gaining importance. Conventional
power stations are centralized and often require electric
M
Limited numbers of kinetic mathematical models, which
evaluate possible homogeneous and heterogeneous
reactions, have been developed for evaluating the effect of
operating conditions on syngas composition and calorific
value during gasification process [3], [4]. However, these
mathematical models deal with transport (heat, mass and
momentum) and kinetic equations leading to extensive

energy to be transmitted over long distances. On the

contrary, localized and small-scale electricity generation efforts for formulation. Therefore, finding the solution of
from renewable energy sources such as biomass, biogas, these models is a time consuming and iterative process.
solar power and wind power increasingly play an important Computational Fluid Dynamics (CFD) can also be used for
D

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

thermochemical conversion process, which transforms

solid biomass into combustible gases, i.e., mixture of
hydrogen, methane, carbon monoxide, carbon dioxide, light
hydrocarbons and char [1]. The produced gaseous mixture,
referred to as syngas, can then be used directly as a gaseous
EP
number of reactions have also been developed for biomass
gasification but these models are more appropriate to
describe fluidized bed gasifiers rather than fixed bed
gasifiers [7], [8].

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
widely studied reactors due to their easy construction and
simple operation [3]. The fixed bed gasification systems are
classified as updraft, downdraft and crossdraft gasifiers.
Downdraft gasifiers have the advantage of yielding lower
tar content (1-2 g/Nm3) compared to other fixed bed
gasifiers. For this reason, they do not require high cost gas
cleaning units and are suitable for small-scale applications.
In a downdraft gasifier, it has been proved that a number of
chemical reactions take place while solid fuels move
together with air in downward direction. Mathematical
models that describe dynamic behavior of fixed bed
gasifiers can facilitate designing cost effective processes to
reduce time and expense of complicated and time
consuming experiments.
the stoichiometric method, the equilibrium is determined
by using the equilibrium constants for some particular
reactions including water gas shift and methanation [12]–
[15]. Both techniques are simple and fast but the
equilibrium condition in the gasifiers is never achieved,
especially for fixed bed gasifiers.
Artificial neural networks (ANN) are widely used for
signal processing, function approximation, simulation, and
recognition of patterns [16]. However, ANNs have been
rarely used for biomass gasification since they require data
sets consisting of large number of samples [17]. Hence,
contrary to fluidized beds, the use of ANNs for modeling
downdraft biomass gasification is very limited [18]–[20].
 ACCEPTED MANUSCRIPT
In addition to the ANNs, other artificial intelligence (AI)
based approaches have been developed for predicting
system outputs effectively by learning and mining system
characteristics from limited experimental data [21].
Therefore, AI based techniques can contribute discovering
underlying rules and characteristics of downdraft biomass
gasification. For instance, the performance of least-squares
support vector machine (LS-SVM) approach is found better
than the ANN approach in the prediction of biochar yield
from cattle manure pyrolysis [22]. However, AI based
classifiers and regressors except for the ANNs have never
2. EXPERIMENTAL SETUP
The experiments were conducted on a 10 kW downdraft
fixed-bed gasifier developed by All Power Labs Inc.
Experimental system design for the modular gasification is
shown in Fig. 1. In a downdraft gasifier, the solid fuels
move together with air in the downward direction. The air
suction provided by venturi ejector enables air to enter the
reactor via an air flow meter. The biomass undergoes
several different processes while passing through four
different zones based on operation: drying zone, pyrolysis

been used in the prediction of syngas composition. In the
literature, input parameters of ANN models include
biomass moisture (MC), biomass content of ash, C, H and
O, gasification temperature (Tg) and equivalence ratio (ER)
PT
zone, combustion zone, and reduction zone. Each stage
involves a specific combination of complex homogeneous
and heterogeneous chemical reactions.
Produced syngas is cleaned through cyclone separator to

[18]–[20], [23]–[25]. However, in our preliminary studies
we found that these parameters, which depend on fuel
composition, did not change the accuracy of predicting
gasification products using the proposed AI based
RI
remove the particles and other condensable materials are
cleaned through the packed bed filter and cooled further.
Finally, the flue gas is burnt out with a swirl burner. A
small gas stream before the burner is brought via a small

approaches. Consequently, the temperature distribution has
more influence on the syngas composition compared to the
fuel parameters. Therefore, in this study we prefer to use
temperature distribution, equivalence ratio (ER) and fuel
SC
clean-up unit to a gas analyzer. A portable infrared syngas
analyzer, Wuhan Cubic Syngas Analyzer Gas board 3100P,
was used for the measurement of the concentrations of CO,
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

levels are used to train both multi-class classifiers, i.e.,

random forests (RF), and binary classifiers, i.e., least-
squares SVM, which achieved over 89% and 96%
classification accuracies, respectively. Furthermore, the
D

precision, recall and F1-scores indicate the ability of the

proposed approaches in predicting the syngas composition.
TE

To the best of our knowledge, this is the first study that

utilizes machine learning classifiers for the estimation of
gasification products. Our study proves that machine
learning based approaches offer great potential in
predicting the amount of products released when energy Figure 1. Fixed bed downdraft gasification system and data
EP

sources are gasified. collection instruments

Real and Quantized Values of H 2

C

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
inside the gasifier were measured as T0, T1, T2, T3, T4 , T5
using a group of 6 in-line K-type thermocouples with 8 mm
diameter. Then real-time values of temperatures, pressures
and flow rates were monitored/recorded using an Arduino
based system shown in Fig. 1. However, the collected
temperature and pressure values are noisy and fluctuate due
to the process dynamics. Although, there are instantaneous
differences due to channelling and bridging phenomena,
the gas composition for the same temperature values and
the same air fuel ratios should be fixed. Therefore, it is
quantization error is negligible. Furthermore, the feature
vectors for all samples are assigned class labels such that
training data sets consisting of 37, 20, 13, 24 and 19 dis-
tinct classes are obtained corresponding to distinctive and
discrete CO, CO2, CH4, H2 and HHV levels, respectively.
3.2. Preprocessing
Once the extracted features are assigned labels to define
classes for the discrete levels of CO, CO2, CH4, H2 and
HHV, machine learning classifiers will employ distance
measures to compute similarities and dissimilarities among

expected that the average value and trend of the gas
composition will be the same when the experiment is
repeated. For this reason, it can be deduced that the
repeatability of the experimental processes in this study is
PT
these classes to predict biomass gasification products. For
this purpose, distances between feature vectors need to be
computed by using measures such as Euclidean (L2), city-
block (L1) and Mahalanobis distance measures [26].

expected to be high.
The recorded airflow rate values were used in calculation
of Equivalence Ratio (ER), which is a crucial operating
variable in biomass gasification, and expresses the amount
RI
However, feature vectors consisting of multiple elements
generally lie in a very high dimensional space and if any of
the features has broad range of values, the distance
calculated in the high dimensional space will be dominated

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)

to construct scaled features   , where  

is the ith
measurement uncertainty is ≤ 0.6%, whereas for the ER

element of the feature vector for the nth sample, min 

AN
and FR, it is ≤ ±1%. In addition, the measurement range for

and max 

 show the minimum and maximum values of
the temperature values, T0, T1, T2, T3, T4 and T5, is up to
1250 degrees Celcius (0C), where the measurement
uncertainty is ≤ ±2 0C. the ith feature across all samples, respectively. Therefore,

features with the range 0, 1

.
In this study, we perform five different experiments to form the new set of training data set will consist of scaled
M

training data set for the proposed machine learning

algorithms, which consists of 5237 data samples including 3.3. Classification
concentration values, temperatures, fuel flow rate and
In order to predict composition of syngas, we employ both
D

equivalence ratio values.

binary, i.e., least-squares SVM (LS-SVM), and multi-class,
i.e., random forests, learning techniques. Binary
TE

3. METHODS classification is the process of classifying samples of a

given data set into one of two distinct classes, whereas
3.1. Feature Selection multi-class learning techniques classify samples into one of
three of more groups.
The type of gasifying agent, fuel flow rate (FR), equiva-
EP

lence ratio (ER), gasification temperature and pressure are
important parameters that significantly affect gasification.
As we introduced in section 2, the gasification experiment
has input variables listed as C, H, O, T0, T1, T2, T3, T4 , T5 ,
C
3.3.1. Random Forests
For the development of multi-class aggregated decision
trees, ensemble-learning methods such as boosting and
bagging have been widely studied [27]. Since the decision

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

CH4, H2 and HHV levels and construct an 8 × 1 dimen-

select the remaining eight features to classify CO, CO2, using the best predictor obtained from randomly selected

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

evaluate the effectiveness of the scaled features,   , in

(A), precision (P), recall (R) and F1-score:

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)

space,  is a N – dimensional vector consisting of weights

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
.

min %, , & = + )|+, |-

predicted true positive and true negative events,

!,",#$ 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

where proportion of positive and negative events returned by the

+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

7, b, +; 5 = %, , + ) 6, 1,  

,  + b
− 1
85
80
Classification Accuracy (%)

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

,  is the kernel function. In this paper, we use

, 
where =
5
0
CO CO2 CH4 H2 HHV

B ∥C
Gasification Outcomes

A
, 
,  = + CDC , kernel
∥
radial basis function (RBF), =
?
Figure 3. Prediction of syngas composition using Random Forest (RF)

where E controls the width of the RBF function.

C

and Least-Squares Support Vector Machine (LS-SVM) classifiers:

Classification accuracies over 89% and 96% have been achieved for
the RF and LS-SVM approaches, respectively.
AC

3.3.3. Cross-Validation
Once the extracted and scaled features are given to the
proposed classifiers, one can divide the entire data set into
training and testing data sets to analytically measure the
classification accuracy. For this purpose, cross-validation
techniques, such as the holdout, leave-one-out or k-fold
methods can be employed [37]. However, the k-fold cross-
validation is the most preferred approach that randomly
Precision is defined as the proportion of true positive
events to the events classified as positive in a classification
task, whereas recall measures the proportion of true
positive events to the sum of true positive and false
negative events. Furthermore, F1-score utilizes both the
precision and the recall to obtain a single score, F1, with
the range [0,1], which results in 1 for perfect precision and

the classifier is trained using F − 1 subsets, referred to as

divides entire data set into k equal sized subsets. Therefore,
training set, and is tested using the remaining subset,
referred to as testing set. This process is repeated k times,
i.e., k-folds, such that each subset is used as the test set
recall.
In order to estimate the levels of syngas composition, CO,
CO2, CH4, H2 and HHV, we employed the RF and LS-
SVM classifiers and obtained classification accuracy values
for each product. It should be noted that one-vs-one
 ACCEPTED MANUSCRIPT
approach has been used for the training and testing of the 0.75. This could be caused by the higher number of classes
LS-SVM classifier since the LS-SVM is a binary classifier, as shown in Fig. 5, especially for the CO, which
whereas the task of predicting gasification product levels is complicates the classification task.
inherently a multi-class classification problem [39]. Fig. 3 In order to further evaluate the performance of the machine
illustrates the effectiveness of both classifiers in predicting learning classifiers in discriminating different levels of
the products of the gasification experiment, where we gasification products, we also utilize confusion matrices
achieved over 89% and 96% classification accuracy values computed for each product. A confusion matrix, also
using the RF and the LS-SVM, respectively. One can also named as error matrix, is a two dimensional layout for
see that the LS-SVM classifier performs better than the RF visual inspection such that one could see the prediction
classifier for all gasification products. Furthermore, the LS- performance of a classifier for each class of data and identi-
SVM classifier is able to predict each product with fy how many instances of the predicted classes were cor-

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-

0.36 0.56 0.83 0.69 0.80

7 SVM SVM SVM SVM SVM
M

80
8

± 0.33 ± 0.32 ± 0.11 ± 0.26 ± 0.17

Actual Class Labels (Concentration %)

9
11 70

0.51 0.75 0.95 0.84 0.84

12
13 60
P
± 0.46 ± 0.34 ± 0.07 ± 0.32 ± 0.29
14
15
D

50

0.36 0.55 0.79 0.68 0.77

16
17

± 0.34 ± 0.33 ± 0.16 ± 0.26 ± 0.24

40
18
19

0.45 0.63 0.86 0.79 0.83

30
TE

20
21
R
± 0.44 ± 0.36 ± 0.24 ± 0.32 ± 0.28
22 20
23
24

0.36 0.55 0.81 0.68 0.77

10
25

± 0.33 ± 0.33 ± 0.13 ± 0.26 ± 0.21

26
0
6 7 8 9 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Predicted Class Labels (Concentration %)
EP

0.45 0.67 0.88 0.81 0.83

(b) F1
± 0.43 ± 0.35 ± 0.18 ± 0.32 ± 0.28
Figure 4. Confusion matrices computed for predicting the H2 levels
using the RF and LS-SVM classifiers.
C

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

200 CO CO2 CH4 H2 HHV

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

where 1, is the predicted class label, 1, is the actual class

for the gasification product estimation. Moreover, the
binary LS-SVM classifier outperforms the multi-class RF
D

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
compared to the RF classifier. However, both classifiers
achieve very low NRMSE values, in particularly for the
CO2, H2 and HHV. On the other hand, previous work uti-
lized stoichiometric and non-stoichiometric models for
EP
score, RMSE and NRMSE values prove that the proposed
techniques are great candidates for automatically
estimating levels of gasification products accurately, in
contrast to stoichiometric and non-stoichiometric models.

With this study, we show that the use of temperature

estimating total syngas composition values using seven
different data sets, where RMSE obtained by the stoichio-
metric and non-stoichiometric models ranges from 1.32 to
5.44 and 5.23 to 9.59, respectively [41]. In addition, the
C
distribution is capable of predicting gasification products
using the proposed AI based approaches and the
temperature distribution has more influence on the syngas
composition compared to the fuel parameters. In addition,

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
LS-SVM classifiers perform better than model-oriented
approaches in predicting gasification products. Moreover,
equilibrium models have an inherent tendency to underes-
timate yields of methane and overestimate yields of hydro-
gen [41]. On the other hand, machine learning based tech-
niques do not have such a tendency since they do not in-
clude any parameter related to the chemicals or chemical
reactions. Moreover, both the stoichiometric and non-
stoichiometric equilibrium models suppose that chemical
system is on equilibrium. But for low ER values and gasifi-
cation temperatures, gasification products leave the gasifier
without reaching the equilibrium condition. Therefore, the
heaters and fuel/air injectors. Therefore, this study offers
more flexibility in controlling and optimizing composition
of the syngas from woody biomass gasification.
The future work will focus on machine learning regression
techniques for the prediction and control of syngas
composition obtained by downdraft biomass gasification.
We believe that machine learning algorithms could allow
researchers and professionals to optimize syngas
composition without using models that are prone to the
equilibrium condition.
ACKNOWLEDGEMENTS
 ACCEPTED MANUSCRIPT
This work was supported in part by the Scientific Research
Projects Coordinatorship of Izmir Katip Celebi University
by grant 2018-GAP-MÜMF-0009.
AUTHOR INFORMATION
Corresponding author
*Email: aliyener.mutlu@ikc.edu.tr
REFERENCES
[1] G. Cau, V. Tola, and A. Pettinau, “A steady state model for
predicting performance of small-scale up-draft coal
gasifiers,” Fuel, vol. 152, pp. 3–12, 2015.
[2] E. Azzone, M. Morini, and M. Pinelli, “Development of an
equilibrium model for the simulation of thermochemical
gasification and application to agricultural residues,” Renew.
Energy, vol. 46, pp. 248–254, 2012.
[3] O. Yucel and M. A. Hastaoglu, “Kinetic modeling and
simulation of throated downdraft gasifier,” Fuel Process.
Technol., vol. 144, pp. 145–154, 2016.
[4] C. Di Blasi and C. Branca, “Modeling a stratified downdraft
wood gasifier with primary and secondary air entry,” Fuel,
vol. 104, pp. 847–860, 2013.
[5] B. Rahmanian, M. R. Safaei, S. N. Kazi, G. Ahmadi, H. F. Oztop,
U
and K. Vafai, “Investigation of pollutant reduction by
simulation of turbulent non-premixed pulverized coal
combustion,” Appl. Therm. Eng., vol. 73, no. 1, pp. 1222–
AN
1235, 2014.
[6] M. R. Safaie, B. Rahmanian, and M. Goodarzi, “Investigation of
the coal diameter effect on pulverized coal combustion for
pollutant reduction,” J. Math. Comput. Sci., vol. 12, pp. 143–
151, 2014.
[7] M. Oevermann, S. Gerber, and F. Behrendt, “Euler-
M
Lagrange/DEM simulation of wood gasification in a bubbling
fluidized bed reactor,” Particuology, vol. 7, no. 4, pp. 307–
316, 2009.
[8] T. M. Ismail, M. Abd El-Salam, E. Monteiro, and A. Rouboa,
D
“Eulerian - Eulerian CFD model on fluidized bed gasifier
using coffee husks as fuel,” Appl. Therm. Eng., vol. 106, pp.
1391–1402, 2016.
TE
[9] A. Z. Mendiburu, J. A. Carvalho, R. Zanzi, C. R. Coronado, and J.
L. Silveira, “Thermochemical equilibrium modeling of a
biomass downdraft gasifier: Constrained and unconstrained
non-stoichiometric models,” Energy, vol. 71, pp. 624–637,
2014.
[10] A. Gambarotta, M. Morini, and A. Zubani, “A non-
EP
stoichiometric equilibrium model for the simulation of the
biomass gasification process,” Appl. Energy, 2017.
[11] J. Han et al., “Modeling downdraft biomass gasification
process by restricting chemical reaction equilibrium with
Aspen Plus,” Energy Convers. Manag., vol. 153, pp. 641–648,
C
2017.
[12] E. S. Aydin, O. Yucel, and H. Sadikoglu, “Development of a
semi-empirical equilibrium model for downdraft gasification
AC
systems,” Energy, vol. 130, pp. 86–98, 2017.
[13] M. Fani, M. Haddadzadeh Niri, and F. Joda, “A Simplified
Dynamic Thermokinetic-Based Model of Wood Gasification
Process,” Process Integr. Optim. Sustain., 2018.
[14] E. Shayan, V. Zare, and I. Mirzaee, “Hydrogen production
from biomass gasification; a theoretical comparison of using
different gasification agents,” Energy Convers. Manag., vol.
159, pp. 30–41, 2018.
[15] D. Castello and L. Fiori, “Supercritical water gasification of
biomass: A stoichiometric thermodynamic model,” Int. J.
Hydrogen Energy, vol. 40, no. 21, pp. 6771–6781, 2015.
[16] A. A. A. A. Alrashed, M. S. Gharibdousti, M. Goodarzi, L. R. de
Oliveira, M. R. Safaei, and E. P. Bandarra Filho, “Effects on
thermophysical properties of carbon based nanofluids:
Experimental data, modelling using regression, ANFIS and
ANN,” Int. J. Heat Mass Transf., vol. 125, 2018.
[17] T. K. Patra and P. N. Sheth, “Biomass gasification models for
downdraft gasifier: A state-of-the-art review,” Renew.
Sustain. Energy Rev., vol. 50, pp. 583–593, 2015.
[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.
[19] R. Mikulandrić, D. Böhning, R. Böhme, L. Helsen, M.
Beckmann, and D. Lončar, “Dynamic modelling of biomass
gasification in a co-current fixed bed gasifier,” Energy
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
downdraft gasifiers,” Biomass and Bioenergy, vol. 98, 2017.
PT
[21] D. S. Pandey, I. Pan, S. Das, J. J. Leahy, and W. Kwapinski,
“Multi-gene genetic programming based predictive models
for municipal solid waste gasification in a fluidized bed
gasifier,” Bioresour. Technol., vol. 179, 2015.
[22] H. Cao, Y. Xin, and Q. Yuan, “Prediction of biochar yield from
RI
cattle manure pyrolysis via least squares support vector
machine intelligent approach,” Bioresour. Technol., vol. 202,
pp. 158–164, 2016.
[23] J. George, P. Arun, and C. Muraleedharan, “Assessment of
SC
producer gas composition in air gasification of biomass
using artificial neural network model,” Int. J. Hydrogen
Energy, vol. 43, no. 20, pp. 9558–9568, 2018.
[24] Y. Li, L. Yan, B. Yang, W. Gao, and M. R. Farahani, “Simulation
of biomass gasification in a fluidized bed by artificial neural
network (ANN),” Energy Sources, Part A Recover. Util.
Environ. Eff., vol. 40, no. 5, pp. 544–548, 2018.
[25] S. Sunphorka, B. Chalermsinsuwan, and P. Piumsomboon,
“Artificial neural network model for the prediction of kinetic
parameters of biomass pyrolysis from its constituents,” Fuel,
vol. 193, pp. 142–158, 2017.
[26] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification.
2000.
[27] a Liaw and M. Wiener, “Classification and Regression by
randomForest,” R news, vol. 2, no. December, pp. 18–22,
2002.
[28] L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp.
5–32, 2001.
[29] J. S. Ham, Y. Chen, M. M. Crawford, and J. Ghosh,
“Investigation of the random forest framework for
classification of hyperspectral data,” IEEE Trans. Geosci.
Remote Sens., vol. 43, no. 3, pp. 492–501, 2005.
[30] P. Xu and F. Jelinek, “Random forests and the data
sparseness problem in language modeling,” Comput. Speech
Lang., vol. 21, no. 1, pp. 105–152, 2007.
[31] J. Ramírez et al., “SPECT image classification using random
forests.,” Electron. Lett., vol. 45, no. 12, pp. 604–605, 2009.
[32] V. Vapnik and R. Izmailov, “Knowledge transfer in SVM and
neural networks,” Ann. Math. Artif. Intell., vol. 81, no. 1–2, pp.
3–19, 2017.
[33] R. Langone, C. Alzate, B. De Ketelaere, J. Vlasselaer, W. Meert,
and J. A. K. Suykens, “LS-SVM based spectral clustering and
regression for predicting maintenance of industrial
machines,” Eng. Appl. Artif. Intell., vol. 37, pp. 268–278, 2015.
[34] U. Thissen, B. Ustun, W. J. Melssen, and L. Buydens,
“Multivariate calibration with least-squares support vector
machines,” Anal. Chem., vol. 76, no. 11, pp. 3099–3105, 2004.
[35] V. Bajaj and R. Pachori, “Classification of Seizure and Non-
seizure EEG Signals using Empirical Mode Decomposition,”
IEEE Trans. Inf. Technol. Biomed., vol. 16, no. 6, pp. 1135–
1142, 2011.
[36] J. A. K. Suykens and J. Vandewalle, “Least Squares Support
Vector Machine Classifiers,” Neural Comput., pp. 1–10, 1999.
[37] D. Glotsos et al., “A multi-classifier system for the
characterization of normal, infectious, and cancerous
prostate tissues employing transrectal ultrasound images,”
Comput. Methods Programs Biomed., vol. 97, no. 1, pp. 53–61,
2010.
[38] C. Ambroise and G. J. McLachlan, “Selection bias in gene
extraction on the basis of microarray gene-expression data,”
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