Chemosphere 92 (2013) 258–264

Contents lists available at SciVerse ScienceDirect

Chemosphere
journal homepage: www.elsevier.com/locate/chemosphere

Modeling the dioxin emission of a municipal solid waste incinerator

using neural networks
Sond Bunsan a,b, Wei-Yea Chen c, Ho-Wen Chen c,⇑, Yen Hsun Chuang c, Nurak Grisdanurak d
a
International Postgraduate Programs in Environmental Management, Graduate School, Chulalongkorn University, Bangkok 10330, Thailand
b
Center of Excellence for Environmental and Hazardous Waste Management (EHWM), Chulalongkorn University, Bangkok 10330, Thailand
c
Department of Environmental Science and Engineering, Tunghai University, No. 181, Sec. 3, Taichung Port Rd., Xitun Dist., Taichung City 407, Taiwan, ROC
d
Department of Chemical Engineering, Faculty of Engineering, Thammasat University, Phathumthani 12121, Thailand

h i g h l i g h t s

" An ANN-based model is proposed to predict the dioxin emission of an incinerator.

" Multivariate data analysis detects the variables for prediction modeling of building.
" Sensitivity analysis sorts the major input variables for predicting dioxin emission.

a r t i c l e i n f o a b s t r a c t

Article history: Incineration is considered as an efficient approach in dealing with the increasing demand for municipal
Received 17 September 2012 and industrial solid waste treatment, especially in areas without sufficient land resources. Facing the con-
Received in revised form 10 January 2013 cern of health risk, the toxic pollutants emitted from incinerators have attracted much attention from
Accepted 19 January 2013
environmentalists, even though this technology is capable of reducing solid waste volume and demand
Available online 4 April 2013
for landfill areas, together with plenty of energy generation. To reduce the negative impacts of toxic
chemicals emitted from incinerators, various monitoring and control plans are made not only for use
Keywords:
in facilities performance evaluation but also better control of operation for stable effluent quality. How
Incinerators
Dioxin
to screen out the key variables from massive observed and control variables for modeling the dioxin
Waste treatment emission has become an important issue in incinerator operation and pollution prevention. For these rea-
Artificial neural network sons, this study used 4-year monitoring data of an incinerator in Taiwan as a case study, and developed a
Sensitivity analysis prediction model based on an artificial neural network (ANN) to forecast the dioxin emission. By doing
this, a simplified monitoring strategy for incinerators with regarding to dioxin emission control can be
achieved. The result indicated that the prediction model based on a back-propagation neural network
is a promising method to deal with complex and non-linear data with the help of statistics in screening
out the useful variables for modeling. The suitable architecture of an ANN for using in the dioxin predic-
tion consists of 5 input factors, 3 basic layers with 8 hidden nodes. The R2 was found to equal 0.99 in both
the training and testing steps. In addition, sensitivity analysis can identify the most significant variables
for the dioxin emission. From the obtained results, the frequency of activated carbon injection showed as
the factor of highest relative importance for the dioxin emission.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction technology is an efficient method to reduce the large amount of so-

Limited by insufficient land resources, incinerators are consid-
ered in many countries such as Japan, and Germany as the major
technology for a waste management scheme capable of dealing
with the increasing demand for municipal and industrial solid
waste treatment in urban regions. Incineration is one of several op-
tions for waste disposal in countries with limited land area. This
⇑ Corresponding author. Tel.: +886 4 23590121x33636; fax: +886 4 23594276.
E-mail address: hwchen@thu.edu.tw (H.-W. Chen).
lid waste volume and demand for landfill areas. Taiwan is an island
with a total area of 36 008 km2, where about 90% of the population
of 23 million lives on the urbanized and industrialized western
plain of the island. Rapid economic development caused an explo-
sive increase of the municipal solid waste streams in the 1980s and
1990s. Restricted to limited land resources, in the 1990s municipal
solid waste incinerators began to be considered as a strategic op-
tion for waste reduction and disposal. The solid waste incineration
system in Taiwan is exemplary in this regard. Up to the present,
there are 27 large-scale municipal solid waste incinerators in

0045-6535/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.chemosphere.2013.01.083
 S. Bunsan et al. / Chemosphere 92 (2013) 258–264
operation. The density of incinerators has become the highest in
the world. About 90% of municipal solid waste can be smoothly
handled by these incinerators (TWEPA, 1999).
To achieve the target of building at least one incinerator in each
city or county, the government offers various incentive-based
plans for promotion such as the Build–Operation–Transfer (BOT)
or the Build–Operation–Own (BOO) projects to encourage private
sectors to join the venture. As the municipal solid waste incinera-
tors were considered to be the main contributors of toxic air pollu-
tants, such as dioxin, in the ambient air (Jones et al., 1993), there is
great concern about the potential human health hazards emitted
from municipal solid waste incinerators. This misgiving has been
vigorously represented by the phenomenon of the not-in-my-
backyard (NIMBY) syndrome (Hsu, 2006). This triggered various
investigations concerning toxin pollutants from incinerators
(Chang and Lin, 2001; Ma et al., 2002; Chen et al., 2006), combined
with the promulgation of many emission standards delegated to
the Environmental Protection Administration in Taiwan (TWEPA).
Polychlorinated dibenzo-p-dioxins (PCDDs) or dioxins are re-
leased as an undesired by-product from the combustion process
of waste incineration. With the health concern, the formation of
PCDDs in the combustion process has been extensively investi-
gated (Altwicker, 1996; Cunliffe and Williams, 2009). It is believed
that PCDDs can be formed by the de novo synthesis where macro-
molecular carbon exists in the temperatures in the range of 250–
450 °C. In addition, the precursors derived from incomplete com-
bustion of organic waste in the chamber can lead to PCDDs forma-
tion in the post-combustion regions of incineration processes as
well (McKay, 2002; Xhrouet and De Pauw, 2003). It is because
PCDDs can be rapidly accumulated in the fatty tissue of animals
(Hsu et al., 2007) and results in immune system damage of human
beings, thus governments launch various guidelines to regulate the
PCDDs emissions (Matés et al., 2010). Due to the concern of public
health, the Taiwan EPA sets the standard value to be 1.0 ng-
TEQ Nm3 for existing MSW incinerators and 0.1 ng-TEQ Nm3
for latterly built incinerators (Chang-Chien et al., 2001), and
encourages the developments of various control technologies for
dioxin and other toxic pollutant reductions. According to Taiwan
EPA 2003 regulations, incinerators have to be equipped with a
control panel which displays real-time conditions of listed opacity
Fig. 1. The schematic of taichung city refuse incineration plant (information source: <http://www.tccip.com.tw/02-4map.htm>).
259
pollutants, carbon monoxide concentration, and emission oxygen
concentration in order to accomplish the goals set by legislation.
Given this fact, many researches have revealed that combustion
conditions may affect the formation of dioxin emission in an incin-
erator. Several dioxin reduction technologies based on combustion
control have been established so far (Buekens and Huang, 1998;
Milosavlkevic and Pullumbi, 2000; Chang et al., 2009). However,
it is worth noting that effective PCDDs emission control strategies
should be supported by the accurate prediction of the pollutants
(Chang and Chen, 2000). Thus, several research studies have re-
ported the prediction of dioxin emission by using mathematical
modeling approaches. Blumenstock et al. (1999) have conducted
a study on the estimation of dioxin emission in the fuel and stack
gas of a hazardous waste incinerator. This study applied the meth-
od of principal component analysis (PCA) and correlation coeffi-
cients based on linear regression to find the relationships
between the PCDDs emission and potential indicator substances.
The study concluded that the chlorinated benzenes can be identi-
fied as a good indicator factor for the prediction of PCDDs emission.
This finding relied on the result from the regression line and indi-
cated the important factor in the prediction approach. Choi et al.
(2007) applied multi-regression analysis to predict the variations
of polychlorinated dibenzofurans (PCDD/DF), with sampling data
sets collected from seven MSW incinerators in Korea. The result
indicated that the changes of temperature and dust concentration
at the inlets and outlets of wet scrubbers are the important factors
that affected the PCDD/DF in the flue gas. This prediction is based
on empirical data in the study area incinerators. Thus, this discov-
ery might not consistently estimate the levels of PCDD/DF in differ-
ent incineration plants. However, the most significant parameters
were identified by this study.
The ANNs have been identified as an efficient tool to determine
the input–output relationships of a sophisticated nonlinear system.
Over the past century, ANNs has been widely applied to waste
management systems and complex systems. In 2003, Dong et al.
used a feed forward neural network to predict the calorific value
of the municipal solid waste (MSW) in China. The physical compo-
nents of MSW were used as the predominant factor and this re-
search stated that the neural network provides a more accurate
prediction than a traditional model. Hamed et al. (2004) published
260 S. Bunsan et al. / Chemosphere 92 (2013) 258–264

a paper in which they investigated ANN modeling for prediction calculation is based on non-diluted dry emission volume at tem-
of biochemical oxygen demand and suspended solids in the perature of 273 K, pressure of 1 atm and 10% emission oxygen con-
wastewater treatment plant. Moreover, this issue also showed centration as the reference standard regulated by Taiwan
that the high performance of ANNs was obtained in a situation environmental protection administration (1997). To build the
with limited historical data. Besides, Rene et al. (2011) applied ANN-based prediction model, 4-year monitoring data sets were
ANN modeling in air pollution control in 2011. The ANN was collected, and a systematic flowchart, including data collection,
proposed in order to predict the removal efficiency of gas-phase variable reduction, and neural network construction is established
styrene in biological systems and the results revealed that the as shown in Fig. 2. Following the standard procedure made by Tai-
ANN model gave high prediction accuracy as well. It is interest- wan EPA, continuous sampling for 8 h is carried out to take each
ing to note that ANN systems perform well and provide satisfac- dioxin sample. To make sure that the input and output variables
tory prediction results. are within the same time frame, the mean of input variables during
So far the ANN method has been applied to the prediction ap- the sampling period is calculated and used to stand for the opera-
proach. To enhance the performance of ANN, the number of in- tion characteristics of selected incinerator.
put variables is the important key to achieving an acceptable
error. Basically, the patterns between input and output will be
recognized through the learning step. Thus, selecting the suitable
input variables will enable the learning algorithm to work faster
and more effectively (Saxén and Pettersson, 2006) with the help
of statistical analyses. Principal component analysis (PCA) is one
of the common multivariate statistical techniques that are used
to achieve great efficiency of data compression from the original
data as well as to indicate natural associations between samples
and/or variables (Wenning and Erickson, 1994; Astel et al., 2007)
by gaining some information useful in the interpretation of
environmental system. PCA consists of diagonalization of the
covariance or correlation matrix transforming the original mea-
surements into linear combinations of these measurements,
and then the explained variance of each principal component
can be maximized. It has been widely used to reveal the rela-
tionships among variables as well as to classify them into differ-
ent grouped variable sets, so that some special features inherent
in the measured system can be characterized (Lautre and Ferná-
ndez, 2004; Macciotta, et al., 2006; Lucas et al., 2008). To build
an accurate prediction model for dioxin emission control, this
study develops a series of procedures for planning a long-term
effectiveness policy in an incinerator, in which an integrated sta-
tistical approach composed of correlation analysis and principal
component analysis was used to screen out the key variables
from massive observed and control variables, and then a feed-
forward ANN is established after training and testing steps. Fi-
nally, sensitivity analysis is utilized to find the key contributor
to PCDDs emission control.

2. Methodology

An incineration plant located in central Taiwan was selected as

the case study of this research. This plant has been in operation at
Taichung city since 1995 with an average daily design capacity
equal to 900 tons of municipal solid waste. As shown in Fig. 1,
the incinerator consists of typical components: the furnace for
combustion, the heat recovery equipment, and the flue gas treat-
ment section. Various monitoring sites and dioxin sampling points
are schematically identified in Fig. 1 as well. The sampling method
was in compliance with the standard sampling method of PCDDs in
flue exhausted, NIEA A807.74C, which was issued by Environmen-
tal Analysis laboratory EPA, Executive Yuan (ROC). The particles
and gaseous are collected by XAD-2 absorbing tube for 8 h and
washed by solvents (acetone, methylene dichloride and toluene).
In order to determine the concentration of PCDDs, the sampling
from MSW incineration will be analyzed by NIEA A808.73B meth-
od. The collected sample will be diluted by 13-C isotope dilution
following with gas chromatography and high-resolution mass
spectrometer. The concentration of dioxin emission standard
in discharge flue gas is usually reported as toxicity weight equiva- Fig. 2. Flowchart of the artificial neural network approach for predicting the dioxin
lent quantity of 2,3,7,8-tetrachlorinated dibenzo-p-dioxin. The emission.
 S. Bunsan et al. / Chemosphere 92 (2013) 258–264 261

2.1. Data collection, preparation and variable screening
For building the prediction model, the operation data sets, con-
sisting of a total of 23 observed and control variables in selected
incinerator plants were gathered at various time points during
the years 2005–2008. These variables in this research are summa-
rized Table 1 and corresponding monitoring sites are listed in
Fig. 1. The PCDDs was investigated once a week from the treated
exhaust gases with a concentration unit of ng-Toxic Equivalency.
Finally, a total of 63 data sets were obtained after screening out
some missing or unusable data. To avoid the noise derived from
different unit scales, Eq. (1) is used to normalize the raw data sets.
xi  Min
x0i ¼
Max  Min
where x0i is the normalized data, xi is the raw data, Min is the min-
imum of the raw data, and Max is the maximum of the raw data.
Based on the normalized data, the following statistical analyses
can be made.
In order to build a simplified and accurate prediction model, an
integrated multivariate analysis consisting of PCA and correlation
analysis is proposed to select the key variables from a massive
number of monitoring variables in an incinerator. Firstly, Pearson’s
correlation matrix was calculated to present the linear relationship
between variables. It is believed that higher similarity will occur
between two variables if their correlation coefficient is higher.
Thus, from the results of Pearson’s correlation, we can exclude
some variables with high correlation coefficients from the candi-
dates for building the prediction model. The mathematical formula
for computing the Pearson correlation coefficient (r) in this paper
is:
Pn 0 The connection between the layers is linked with the neural
 xÞðyi  y
i¼2 ðxi
Þ
r¼
ðN  1Þðsx Þðsy Þ
 N, and s are sample mean, sample size, and sample stan-
where ðxÞ,
dard deviation, respectively.
As shown earlier the relationship between the inputs and the
output (PCDDs emission) can be identified by the correlation anal-
ysis and several factors can be deleted from the system. For the
purpose of variables reduction, principal component analysis was
applied in this research. The basis of PCA has been well-explained
(Jolliffe, 2002). Briefly, PCA is used for characterizing patterns
within large sets of data by re-expressing to a rotated coordinate
system in which as much variance as possible is explained by the
first few dimensions. In this coordinate system the eigenvectors
of the variance covariance matrix are calculated, so that the princi-
pal component score, i.e. the weight of the eigenvector, can be ob-
tained. The scores of the original variables, also called principal
component loadings (PC loadings), can be used to indicate the rela-
tionship between the variable and the principal component. In
other words, the variable and the principal component will be
more strongly related if the PC loading is larger. By doing this,
the raw data matrix can be reduced to two or three principal com-
ð1Þ ponent loadings that account for the majority of the variance. Thus,
these factors can be used to account for the required information
approximately as well as the original observations do. Thus, this
paper uses the software package SPSSÒ (Version 19.0) for Windows
to determine the structure in the relationships between parame-
ters and identify the most important factors contributing to this
structure based on the eigen analysis of the correlation matrix.
2.2. Modeling the dioxin emission by ANNs
The artificial neural network is one of several data-driven ap-
proaches and is widely used to determine the nonlinear relation-
ships among input and output variables. It imitates the behaviors
of the nervous system in the human brain, being capable of recog-
nizing the patterns of a system after experiencing learning from a
set of training data. The basic structure of an artificial neural net-
work can be described as consisting of three independent layers:
input, hidden, and output layers (Fig. 2).
ð2Þ synaptic weight coefficients (wij). Commonly, when inputs (xi) en-
ter the system, they are multiplied by synaptic weights and
P
summed up at each node ( ðwij Þðxi Þ). The summed values can be
passed throughout the network and activate all hidden nodes until
they reach the output layer when these values are greater than the
threshold values. After the input is processed for all layers, the er-
rors between the computed and actual output are presented as
well.

Table 1
Basic Statistical characteristic of variables

Variables Unit Max Min Mean Standard Deviation

Waste loading (X1) ton 16.3 3.51 9.13 4.25
Activated carbon injection frequency (X2) Hz 26.2 14.5 20 2.66
Activated carbon injected (X3) Kg/hr 5.66 3.94 4.63 0.44
Concentration of hydrogen chloride in the flue gas at the stack emission (X4) ppm 34.18 11.11 22.41 5.42
Concentration of carbon monoxide in the flue gas at the stack emission (X5) ppm 21.68 0.87 3.53 4.51
Concentration of nitrogen oxide in the flue gas at the stack emission (X6) ppm 140.46 56.15 98.22 19.40
Concentration of sulfur dioxide in the flue gas at the stack (X7) ppm 8.19 0.22 3.13 2.08
Carbon dioxide at stack emission (X8) % 8.26 4.74 6.86 0.75
Water at stack emission (X9) % 23.44 15.19 20.26 1.84
Oxygen at stack emission (X10) % 12.81 8.87 10.64 0.90
Opacity at stack emission (X11) % 5.45 0.48 2.2 1.31
Dust at stack emission (X12) mg/Nm3 4.84 0.48 1.91 1.18
Flue gas flow rate at the chimney final exit (X13) Nm3/hr 75856.36 42936.28 55811.14 5037.72
Temperature at the 1st furnace room (X14) °C 184.98 146.54 167.38 8.97
Temperature at SAH export (X15) °C 201 89 155 30.23
Temperature at the 2nd furnace room (X16) °C 1069 721 926 65.14
Temperature at burning heart furnace room (X17) °C 1176 745 944 109.57
Temperature at vice flue (X18) °C 1151 859 1010 60.58
Temperature at rotary kiln exit (X19) °C 1136 788 882 61.16
Temperature at mixing chamber (X20) °C 1099 916 981 47.52
Temperature at boiler convection zone (X21) °C 724 607 667 29.66
Temperature at chimney entrance (X22) °C 243 149 167 21.81
Temperature at final fuel gas emission (X23) °C 65.28 47.86 55.48 3.93
Dioxin Emission (Y) ng-TEQ 0.59 0.003 0.15 0.118
262 S. Bunsan et al. / Chemosphere 92 (2013) 258–264

Notice that synaptic weights and threshold values are uncertain

x23
numbers. However, the network can adjust these values by apply-

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
ing a particular example case into the learning process. The overall

b
0.361
learning process is consists of calculating in a circle until it equals

x22

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
the local error. The hidden nodes are activated by the sigmoid acti-

0.186
0.162
vation function. When the networks reach the output layer, the

x21

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
computed errors are sent backward to the first step again for learn-

0.195
0.098
0.075
ing the mistake error and weight adjustment. The network will rec-

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
x20
ognize the optimal pattern in order to get the target or output with

–
the expected results. This process of training data is called a back-

0.259a
a
0.021
0.059

0.305
propagation algorithm. The algorithm works as follows (Chen et al.,

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
x19

–
2002; Tsai et al., 2002; Yetilmezsoy et al., 2011);

a
0.217

0.171
0.317
0.084
0.009
(1) Define the input data set for training in neural networks

x18

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
–
(2) Calculate the actual outputs at each hidden layer (j) by the

0.308a
0.739

0.067

0.038
0.009

0.006
following mathematical formula, where n is the number of

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
x17
inputs of neuron j, and h is the threshold value.

–
!

0.457b
b
0.745

0.197
0.034
0.057
X

0.003
0.003
n
yi ¼ f wij x0i  hj ð3Þ

–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
x16

–
i¼1

b
0.403b

0.307a
0.359

0.188
0.191

0.117
0.054

0.091
(3) Transfer the output values (yj) in the hidden layer by using

x15

–
–
–
–
–
–
–
–
–
–
–
–
–
–
1
–
the sigmoid function as shown;

0.641b
b
0.249a
0.245

0.573
0.072
0.087
0.058

0.220
0.000
1
f ðxÞ ¼   ð4Þ

–
–
–
–
–
–
–
–
–
–
–
–
–
1
x14

–
1  expðyj Þ

0.348b
0.401b

0.198
0.206

0.027
0.206
0.015

0.082

0.042
0.22
(4) Calculate the actual outputs at the output layer (k), where m
x13

–
–
–
–
–
–
–
–
–
–
–
–
1
–

is the number of inputs of neuron k, and transfer yk by apply-

ing the sigmoid function

0.341b
0.191

0.214
0.012
0.075
0.063

0.054
0.077
0.031

0.202
0.002
!
–
–
–
–
–
–
–
–
–
–
–
1
x12

X
m
–

yk ¼ f wjk xjk  hjk ð5Þ

0.345b
b

0.199

0.215
0.077
0.057

0.013
0.042
0.086
0.031
0.001

0.200
1.000
j¼1
–
–
–
–
–
–
–
–
–
–
1
x11

(5) Calculate the error at the output layer (ek ), where yd;k is the

0.355b

0.555b
b
0.520b
0.290a
0.224

0.148

0.139

0.169

0.434
0.220

0.027

0.201

0.074
desired output of neuron k at the output layer
–
–
–
–
–
–
–
–
–
1
x10

e ¼ yk ð1  yk Þðyd;k  yk Þ ð6Þ
0.323b
b

0.251a
0.324

0.151
0.131
0.229

0.139

0.129

0.232
0.230

0.067

0.029

0.016

0.006
(6) Change the weights at the output layer (wij ), where g is a
–
–
–
–
–
–
–
–
1
x9

constant called the learning rate

0.615b

0.487b
0.479b
0.386b

0.332b
0.171
0.114

0.186

0.144
0.120

0.021

0.032

0.044
0.100
0.003
wij ¼ wij þ ðgÞðyj ðek ÞÞ ð7Þ
–
–
–
–
–
–
–
1
x8

(7) Calculate (back-propagate) the error at the hidden layer (ej ),

0.341b

0.552b

0.425b

0.396b
b
0.259a

0.257a
0.194
0.189
0.164

0.112

0.344
0.190

0.011

0.056

0.028
0.003

where l is the number of outputs of neuron k at the hidden

–
–
–
–
–
–
1
x7

layer
Selected variables for building artificial neural network by correlation analysis.

0.344b
b

a
0.135
0.117

0.189
0.141

0.479

0.268
0.075
0.032

0.180

0.078
0.044
0.201

0.203
0.240
0.150
0.052

X
l
ej ¼ ðyj Þð1  yj Þ ek wjk
–
–
–
–
–
1

ð8Þ
x6

k¼1
0.292a

0.250a
0.154
0.168
0.243

0.199

0.197
0.144
0.162

0.239
0.205

0.150
0.063
0.026

0.106
0.200

0.020

0.08

(8) Change the weights at the hidden neurons (wij )

–
–
–
–
1
x5

wij ¼ wij þ gðxi Þðej Þ ð9Þ

Correlation is significant at the 0.05 level (2-tailed).
Correlation is significant at the 0.01 level (2-tailed).
0.512b

0.392b

0.268a
0.239

0.223
0.162
0.168
0.128

0.129

0.121
0.116

0.188
0.110
0.190

0.103

0.035

0.088

0.097
0.010
–
–
–
1

(9) Repeat calculating back through the algorithm and allow the
x4

system to learn how the error is related to the inputs, out-

0.343b

0.438b
0.420b
0.135

0.177

0.114
0.213

0.125
0.527
0.162

0.239
0.263
0.193
0.347

0.919
0.420

0.202
0.026

0.440
0.020

puts, and weights. Finally, the best fit condition among the

–
–
1
x3

parameters will be identified.

0.847b

0.425b

0.456b
0.520b

0.380b
b
0.296a

0.311a
a

0.280a
0.116

0.121
0.172

0.313
0.118

0.156

0.213
0.147
0.212
0.093
0.061

0.031

0.804

3. Results and discussion

–
1
x2

b
0.258a
0.231

0.231
0.215

0.177

0.174
0.173
0.068
0.037

0.091
0.092
0.013

0.025
0.049

0.063

0.071
0.104
0.370
0.040

0.100

0.100

3.1. Basic statistical analysis and variable selection

0.17
1
x1

There are 23 variables considered as the candidates for building

b

0.435b
0.313a

0.296a
0.354

0.181

0.217
0.182

0.147
0.053

0.209

0.103
0.099

0.064

0.110
0.094
0.086
0.061

0.054
0.011
0.084
0.100

0.009

0.100

the ANN-based prediction model. For excluding some redundant

1
y

variables, correlation analysis and PCA were carried out. Table 2

Variable

presents the Pearson’s coefficients among the 23 variables in this

Table 2

study, where ten variables including x1, x6, x8, x11, x12, x13, x15,
x10

x20
x1
x2
x3
x4
x5
x6
x7
x8
x9

x11
x12
x13
x14
x15
x16
x17
x18
x19

x21
x22
x23
y

b
a

x16, x17 and x18 are eliminated owing to their low Pearson’s
 S. Bunsan et al. / Chemosphere 92 (2013) 258–264 263

Table 3 Normally, activated carbon has been applied as the adsorbent

Selected variables for building artificial neural network by principal component for removal of unwanted gases in the waste incinerator gas stream.
analysis.
Moreover, there are several kinds of gases that might be adsorbed
Variable Component in the activated carbon such as HCl, SOx, NOx, CO2 and/or volatile
1 2 3 4 5 organic compounds. In this study, the results also accord with the
x3 0.945 0.125 – – –
earlier studies, which stated that activated carbon injection is
x2 0.923 – – – 0.148 widely used in waste incinerators as the major air pollution control
x23 0.913 0.116 – – 0.214 device for control of the PCDDs emission (Chi et al., 2006 and His
x4 – 0.851 – – 0.139 et al., 2007). Therefore, the frequency of activated carbon injection
x7 0.223 0.790 – 0.295 0.243
is closely related to the PCDDs emission factors in incineration
x14 0.485 0.539 0.410 0.120 0.120
x9 0.301 – 0.764 0.126 – plants. From the previous studies about the PCDDs emission from
x22 0.408 0.275 0.652 – 0.135 incinerators, the mechanism of PCDDs formation can occur with
x5 – 0.349 0.598 0.525 0.147 the presence of hydrogen chlorine during the burning process
x10 0.490 0.205 0.589 – 0.216
(Ruokojarvi et al., 2004). The possible explanation for this is that
x20 0.147 0.229 – 0.881 –
x19 0.256 0.144 0.142 0.156 0.747
hydrogen chloride can react with oxygen that provides chlorine
x21 0.422 0.238 0.105 0.354 0.647 (Deacon reaction). Therefore, chlorine is a key component for the
% of Variance 32.995 12.651 11.011 10.471 8.651 formation of PCDDs precursors when chlorine reacts with organic
Cumulative 32.995 45.646 56.657 67.128 75.779 compounds (Kilgroe, 1996 and Thomas and McCreight, 2008).

coefficient values that may indicate a weak relationship with the
dioxin emission. And then, the other 13 candidates survive into
the PCA process. The result obtained from the PCA analysis can
be summarized in Table 3.
As shown in Table 3, a five-factor model was determined by
PCA, in which their percentage of variance and cumulative percent-
age of variance explained 75.78% of the total variance in the data
set. For each factor the PC loadings were 33.00%, 12.65%, 11.01%,
10.47% and 8.91% of the variance, and the result shows that there
are five variables which were chosen with loading values over
0.8. These are considered to be the input variables for building
the dioxin emission prediction model. The input variables consist
of activated carbon injection (x2 and x3), concentration of hydrogen
chloride in the flue gas at the stack emission (x4), temperature at
the mixing chamber (x20), and temperature of final fuel gas emis-
sion (x23).
A strong relationship between PCDDs emission and operation
temperature has been reported in the literature. In a typical incin-
eration process, MSW are successfully burned in a furnace with the
series of grate rods while the PCDDs can be formed in the temper-
ature range of 400–800 °C (Altarawneh et al., 2009). The reduction
of PCDDs emission will occur with temperature greater than 900 °C
(Mckay, 2002). The finding from the correlation analysis in this re-
search is similar to the earlier work that can be stated as ‘‘when the
temperature increases the PCDDs emission should decrease.’’ Thus,
the temperature at the mixing chamber (x20) can be noted as the
key factor for the dioxin emission as well.
3.2. Accuracy and reliability analysis of ANNs
After a series of testing, an optimized frame of ANNs is obtained,
where one input layer, one hidden layer with 8 hidden nodes, and
one output layer are constructed. Raw data sets were separated
into training and testing by 70% and 30% of the data set, respec-
tively. The back-propagation algorithm was used as the supervised
training for weight adjustment by minimizing the root mean
square error between predicted and observed values of PCDDs.
Figs. 3 and 4 present the comparisons observed and predicted val-
ues in training and testing steps, respectively. It is apparent from
these figures that the ANN model with the back-propagation learn-
ing algorithm provides high performance in the prediction of
PCDDs emission. The low variance shown in Figs. 3 and 4 indicates
that the total random error is almost equal to zero. However, a sys-
tematic error happened in the training and testing steps. In the
testing step, the predicted values are systematically larger than ob-
served data. Fortunately, such errors can be revised by curve
translation.
3.3. Sensitivity analysis
Sensitivity analysis can be applied to investigate the relative
importance among the various input factors on PCDDs emission
by means of the weights method proposed by Gevrey et al.
(2003). To assess the relative importance approach, the connection
weights between input layer, hidden layer, and output layer were

Fig. 3. The dioxin emission predicted value versus target value (a) training step and (b) testing step.
264
25
20
Relative Importance (%)
15
10
5 municipal solid waste incinerator. J Hazard Mater. 161, 1436–1443.
0 polychlorinated dibenzo-p-dioxins and dibenzofurans in the bottom ash of
Activated Carbon Activated Carbon Hydrogen chloride
Injection (Kg/h) Injection (kg) (ppm)
Fig. 4. The result of sensitivity analysis from weights method.
carried out to identify which input factor plays the most important
role in the dioxin emission. The five input factors considered were:
the frequency of activated carbon injection, the amount of acti-
vated carbon injection, the concentration of hydrogen chloride,
the temperature at the mixing chamber, and the temperature at
the flue excretion. Fig. 4 indicates the resulting relative importance
between the factors. The result can be summarized as showing that
the frequency of activated carbon injection (kg h1) is the most sig-
nificant factor for the PCDDs emission, followed by the concentra-
tion of hydrogen chloride and the temperature at the mixing
chamber.
4. Conclusion
Dioxin emission from Incinerators attracts much attention from
the public owing to their high health risk. Building an accurate pre-
diction model can help the operators to plan control strategies for
pollution reduction. However, how to screen out the key variables
from the massive number of observed and control variables for
modeling the dioxin emission is the key issue in ensuring the suc-
cess of such tasks. This study proposed a series of procedures to
solve the problems that happen in modeling. The results indicate
that the ANN-based prediction model can successfully simulate
the behaviors of PCDDs emission in an incinerator even though a
systematic error exists. The results from sensitivity analysis, tally-
ing with the findings of in the literature, also provide advantageous
information for practical operation.
Acknowledgement
The authors would like to thank the National Science Council of
the Republic of China for financially supporting this research under
Contract No. NSC 97-2221-E-029-029-MY3.
References
Altarawneh, M., Dlugogorski, B.Z., Kennedy, E.M., Mackie, J.C., 2009. Mechanisms for
formation, chlorination, dechlorination and destruction of polychlorinated
dibenzo-p-dioxins and dibenzofurans (PCDD/Fs). Prog. Energy Combust. 35,
245–274.
Altwicker, E.R., 1996. Relative rates of formation of polychlorinated dioxins and
furans from precursor and de novo reactions. Chemosphere 33, 1897–1904.
Astel, A., Tsakouski, S., Barbieri, P., Simeonov, V., 2007. Comparison of self-
organizing maps classification approach with cluster and principal components
analysis for large environmental data sets. Water Res. 41, 4566–4578.
Blumenstock, M., Zimmermann, R., Schramm, K.W., Kaune, A., Nikolai, U., Lenoir, D.,
Kettrup, A., 1999. Estimation of the dioxin emission (PCDD/FI-TEQ) from the
concentration of low chlorinated aromatic compounds in the flue and stack gas
of a hazardous waste incinerator. J. Anal. Appl. Pyrol. 49, 179–190.
S. Bunsan et al. / Chemosphere 92 (2013) 258–264
Buekens, A., Huang, H., 1998. Comparative evaluation of techniques for controlling
the formation and emission of chlorinated dioxins/furans in municipal waste
incineration. J Hazard Mater. 62, 1–33.
Chang, N.B., Chen, W.C., 2000. Prediction of PCDDs/PCDFs emissions from municipal
incinerators by genetic programming and neural network modeling. Waste
Manage. Res. 18, 341–351.
Chang-Chien, G.P., Lee, W.S., Tsai, J.L., Jeng, S.H., 2001. Liquid crystalline
polysiloxane polymer as stationary phase in gas chromatography capillary
column for the separation of dioxin/furan compounds. J Chromatogr A. 932, 97–
105.
Chang, M.B., Lin, J.J., 2001. Memory effect on the dioxin emissions from municipal
waste incinerator in Taiwan. Chemosphere 45, 1151–11570.
Chang, Y.M., Hung, C.Y., Chen, J.H., Chang, C.T., Chen, C.H., 2009. Minimum feeding
rate of activated carbon to control dioxin emissions from a large-scale
Chen, W.C., Chang, N.B., Chen, J.C., 2002. GA-based fuzzy neural controller design for
municipal incinerators. Fuzzy Set Syst. 129, 343–369.
Chen, C.K., Lin, C.E., Wang, L.C., Chang-Chien, G.P., 2006. The size distribution of
Mixing chamber Flue excretion
(°C) (°C) municipal solid waste incinerators. Chemosphere. 65, 514–520.
Chi, K.H., Chang, S.H., Huang, C.H., Huang, H.C., Chang, M.B., 2006. Partitioning and
removal of dioxin-like congeners in flue gases treated with activated carbon
adsorption. Chemosphere 64, 1489–1498.
Choi, K.I., Lee, D.H., Osako, M., Kim, S.C., 2007. The prediction of PCDD/DF levels in
wet scrubbers associated with waste incinerators. Chemosphere 66, 1131–
1137.
Cunliffe, A.M., Williams, P.T., 2009. De-novo formation of dioxins and furans and the
memory effect in waste incineration flue gases. Waste Manage. 29, 739–748.
Dong, C.Q., Jin, B.S., Li, D.J., 2003. Predicting the heating value of MSW with a feed
forward neural network. Waste Manage. 23, 103–106.
Gevrey, M., Dimopoulos, L., Lek, S., 2003. Review and comparison of methods to
study the contribution of variables in artificial neural network models. Ecol.
Model. 160, 249–264.
Hamed, M.M., Khalafallah, M.G., Hassanien, E.A., 2004. Prediction of wastewater
treatment plant performance using artificial neural networks. Environ. Modell.
Softw. 19, 919–928.
Hsi, H.C., Wang, L.C., Yu, T.H., 2007. Effects of injected activated carbon and
solidification treatment on the leachability of polychlorinated dibenzo-p-
dioxins and dibenzofurans from air pollution control residues of municipal
waste incineration. Chemosphere 67, 1394–1402.
Hsu, M.S., Hsu, K.Y., Wang, S.M., Chou, U., Chen, S.Y., Huang, N.C., Liao, C.Y., Yu, T.P.,
Ling, Y.C., 2007. A total diet study to estimate PCDD/Fs and dioxin-like PCBs
intake from food in Taiwan. Chemosphere 67, S65–S70.
Hsu, S.H., 2006. NIMBY opposition and solid waste incinerator siting in
democratizing Taiwan. Soc Sci J. 43, 453–459.
Jolliffe, I.T., 2002. Principal Component Analysis. Springer Series in Statistics.
Jones, P.H., Degerlache, J., Marti, E., Mischer, G., Scherrer, M.C., Bontinck, W.J.,
Niessen, H.J., 1993. The Global Exposure of Man to Dioxins - a Perspective on
Industrial-Waste Incineration. Chemosphere. 26, 1491–1497.
Kilgroe, J.D., 1996. Control of dioxin, furan, and mercury emissions from municipal
waste combustors. J. Hazard. Mater. 47, 163–194.
Lautre, I.G., Fernandez, E.A., 2004. A methodology for measuring latent variables
based on multiple factor analysis. Comput Stat Data An. 45, 505–517.
Lucas, R., Bunting, P., Paterson, M., Chisholm, L., 2008. Classification of Australian
forest communities using aerial photography, CASI and HyMap data. Remote
Sens Environ. 112, 2088–2103.
Ma, H.W., Lai, Y.L., Chan, C.C., 2002. Transfer of dioxin risk between nine major
municipal waste incinerators in Taiwan. Environ Int. 28, 103–110.
Macciotta, N.P.P., Dimauro, C., Catillo, G., Coletta, A., Cappio-Borlino, A., 2006.
Factors affecting individual lactation curve shape in Italian river buffaloes.
Livest Sci. 104, 33–37.
Matés, J.M., Segura, J.A., Alonso, F.J., Marquez, J., 2010. Roles of dioxins and heavy
metals in cancer and neurological diseases using ROS-mediated mechanisms.
Free Radical Bio Med. 49, 1328–1341.
McKay, G., 2002. Dioxin characterisation, formation and minimisation during
municipal solid waste (MSW) incineration: review. Chem. Eng. J. 86, 343–368.
Milosavlkevic, I., Pullumbi, P. 2000. Chapter 2 - Reduction of dioxins and furans in
incineration. Reaction Engineering for Pollution, Prevention. 7–17.
Rene, E.R., Estefania, L.M., Veiga, M.C., 2011. Neural networks models for biological
waste-gas treatment systems. New Biotechnol. 29, 56–73.
Ruokojarvi, P.H., Asikainen, A.H., Tuppurainen, K.A., Ruuskanen, J., 2004. Chemical
inhibition of PCDD/F formation in incineration processes. Sci. Total Environ.
325, 83–94.
Saxén, H., Pettersson, F., 2006. Method for the selection of inputs and structure of
feedforward neural networks. Comput Chem Eng. 30, 1038–1045.
Thomas, V.M., McCreight, C.M., 2008. Relation of chlorine, copper and sulphur to
dioxin emission factors. J. Hazard. Mater. 151, 164–170.
Tsai, C.P., Lin, C., Shen, J.N., 2002. Neural network for wave forecasting among multi-
stations. Ocean Eng. 29, 1683–1695.
Wenning, R.J., Erickson, G.A., 1994. Interpretation and Analysis of Complex
Environmental Data Using Chemometric Methods. Trac-Trend Anal Chem. 13,
446–457.
Xhrouet, C., De Pauw, E., 2003. Prevention of dioxins de novo formation by
ethanolamines. Environ. Chem. Lett. 1, 51–56.
Yetilmezsoy, K., Fingas, M., Fieldhouse, B., 2011. An adaptive neuro-fuzzy approach
for modeling of water-in-oil emulsion formation. Colloid Surf. A. 389, 50–62.