ARTICLE IN PRESS

Journal of Environmental Management 85 (2007) 891899

www.elsevier.com/locate/jenvman

Multiple regression models for the lower heating value of municipal

solid waste in Taiwan
Y.F. Changa, C.J. Linb, J.M. Chyanb,, I.M. Chena, J.E. Changc
a
Department of Environmental Resource Management, Chia Nan University of Pharmacy and Science, Tainan 717, Taiwan, ROC
b
Department of Environmental Engineering and Science, Chia Nan University of Pharmacy and Science, Tainan 717, Taiwan, ROC
c
Department of Environmental Engineering, National Cheng Kung University, Tainan 701, Taiwan, ROC
Received 23 November 2004; received in revised form 8 October 2006; accepted 30 October 2006
Available online 17 January 2007

Abstract

A multiple regression analysis was used to develop two predictive models of lower heating value (LHV) for municipal solid waste
(MSW), using 180 samples gathered from cities and counties in Taiwan during 20012002. These models are referred to as the original
proposed model (OPM) and the simplied model (SM). The coefcients of multiple determinations for the OPM and SM were 0.983 and
0.975, respectively. To verify the feasibility of the models, a demonstration program based on sampling of MSW in Kaohsiung City was
conducted. As a result, the OPM showed superior precision in terms of relative percentage deviation (RPD) and mean absolute
percentage error (MAPE), when compared to the conventional models based on the proximate analysis, physical composition and
ultimate analysis. The SM was derived by neglecting the three minor physical components used in the OPM. The resulting SM was less
precise when compared to the OPM, but it was still acceptable, with a precision level better than the conventional models. It was
concluded that the predictability of empirical models could be improved signicantly through selection of the appropriate physical
components and multiple regression analysis.
r 2006 Elsevier Ltd. All rights reserved.

Keywords: Municipal solid waste; Physical combustible component; Multiple regression analysis; Lower heating value

1. Introduction a laboratory bomb calorimeter, or calculation based on

Municipal solid waste (MSW) is one of the waste
products generated from our daily life and various
activities. It increases with the growth of personal income
and population. In Taiwan, 6.72 million metric tons of
municipal waste were generated in 2002 and this caused a
severe disposal problem. To reduce the demand for landll
space, incinerators have been used to reduce the nal
disposal volume of MSW. According to the Engineering
Project for the Construction of Refuse Incineration Plants
in Taiwan, 21 MSW incinerators will have been con-
structed by the year 2005. For designing and operating a
mass-burn incinerator, the energy content of MSW is a
very important control parameter. The energy content can
be determined by using a full-scale boiler as a calorimeter,
Corresponding author. Tel./fax: +886 6 2660266.
E-mail address: mjmchyan@mail.chna.edu.tw (J.M. Chyan).
empirical models. Because of the difculty of constructing
a full-scale boiler, most of the data are obtained from the
results of other methods. When compared to the measure-
ments of a bomb calorimeter, an empirical model can
determine the energy content of MSW more economically
if the parameters in the model are well-established.
Generally, the lower heating value (LHV), which is
dened as the energy content released from the combustion
of the organic components of MSW in an incinerator, can
be used to represent the energy content of MSW. From
previous research regarding conventional empirical mod-
els, as listed in Table 1, the LHV can be predicted from the
models based on the proximate analysis [Eqs. (1) and (2)],
the physical composition analysis [Eqs. (3)(5)], and the
ultimate analysis [Eqs. (6)(9)]. Among these models, the
prediction models based on the ultimate analysis data, as
shown in Eqs. (6)(8) (Wilson, 1977) and Eq. (9) (Cooper
et al., 1999) are better than the others. Analyses of C, H, O,

0301-4797/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jenvman.2006.10.025
 ARTICLE IN PRESS
892 Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899

Nomenclature Pte textiles (wt%)

Pwo wood (wt%)
C carbon content (wt%) R coefcient of multiple correlation
Cl chlorine content (wt%) R2 coefcient of multiple determinations
2
E error of the predictive model R adjusted coefcient of multiple determination
Ep predetermined maximum error RPD relative percentage deviation (%)
H hydrogen content (wt%) S sulfur content (wt%)
k number of independent variables SSE sum of squares of deviations of observed results
LHV lower heating value (kcal/kg) from the tted line
LHVdna lower heating value on a dry, no-ash basis SM simplied model
(Btu/lb) SST sum of squares of observed results deviated
MAPE mean absolute percentage error (%) from the mean
MRA multiple regression analysis V physically combustible component content (%)
MSW municipal solid waste W moisture content (%)
n number of simples X a, X b arbitrary data
O oxygen content (wt%) Xk selected independent variables
OPM original proposed model Xm measured LHV (kcal/kg)
Pfo food waste (wt%) Xp predicted LHV (kcal/kg)
Pga garbage (wt%; textiles, wood, food waste, yi observed data
miscellaneous also included) ymean mean of yi
Pmi miscellaneous component (wt%) Z value of standard normal distribution
Ppa paper and cardboard (wt%) bk regression coefcients
Ppl plastics (wt%) s standard deviation

S, Cl and moisture content in MSW are needed for the
prediction of LHV, however, obtaining such data can be
costly and time-consuming. Therefore, it is desirable to
develop a cost-effective model for the prediction of the
LHV of MSW. Some empirical equations based only on
physical components, such as food waste, paper and
cardboard, plastics, textiles, wood, etc., have been estab-
lished. These are Eqs. (3)(5) listed in Table 1 and were
proposed by the Japan National Municipal Solid Waste
Foundation (1991), Liu (1996), and Lin (2000), respec-
tively. From previous research, it has been found that the
prediction error with this method is lower than that of the
proximate analysis (Wilson, 1977), but higher than that of
the ultimate analysis. Considering operation practices, it is
quicker and cheaper to predict LHV by employing the
model based on physical component analysis. To improve
the precision of the predictive model, the parameters of the
physical components are included and the predictive model
is then established by using multiple regression analysis.
Based on 180 samples gathered from 45 towns and counties
in Taiwan, empirical models for LHV prediction were
obtained and compared with those of other equations
proposed in previous research.
2. Theoretical considerations and data sampling
2.1. Theoretical considerations
2.1.1. Development of LHV empirical prediction models
Multiple regression analysis was used to develop an
empirical equation for the LHV of MSW in Taiwan. The
dependent variable was the lower heating value. The
independent variables usually were the weight percents of
paper and cardboard, plastics, food waste, and garbage. To
improve the precision, the term garbage was further
divided into textiles, wood, rubber and leather, and
miscellaneous materials. The regression analysis was then
performed using SPSS statistical software (Navarro-Esbr
et al., 2002; Mann, 1995), where the stepwise regression
selection procedure was used to determine the best tting
regression model. The governing equation can be described
as follows:
y b0 b1 X 1 b2 X 2    bk X k E, (10)
where b0, b1,y, bk are the regression coefcients; and X1,
X2,y, Xk are the selected independent variables; E is a
term accounting for the error of the predictive model. In
the beginning, the independent variable that was correlated
most highly with the dependent variable was selected to
enter into the model, and subsequently the associated
straight-line regression equation was tted. The coefcient
of multiple determinations (R2) reecting the strength of
the straight-line relationship will be high when the
independent variable is highly correlated. The R2 is dened
as
SSE
R2 1  , (11)
SST
where SST S(yiymean)2 represents the sum of squares of
the difference between observed data (yi) and the mean
(ymean) of yi, and SSE S(yiypredicted)2 denotes the sum
of squares of deviations of observed data (yi) from the
 ARTICLE IN PRESS
Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899 893

Table 1
Summary of equations for predicting LHV of municipal slid waste

Proximate analysis
LHV 45V  6W (1)

LHV 44:75V  5:85W 21:2 (2)

Physical composition analysis  

   100  W
LHV 88:2Ppl 40:5 Pfo Ppa  6W (3)
W

LHV 2229:91 7:90Ppa 28:16Ppl 4:87Pga  37:28W (4)

 
100  W
LHV 38:52Ppa 92:09Ppl 49:24Pte 38:34Pwo 37:55Pfo 64:07Pmi  6W (5)
W

Ultimate analysis 
O
LHV 81C 342:5 H  22:5S  6W 9H (6)
8
   
3 3 O
LHV 81 C   O 57  O 345 H  25S  6W 9H (7)
8 8 16
 
3O 3O
LHV 81 C  342:5H 22:5S 57   6W 9H (8)
4 4
 
O Cl
LHVdna 7:91 170:50C 320:30 H   45:91S (9)
8 35:5

Remark:
C: carbon content (wt%) Pmi: miscellaneous component (wt%)
Cl: chlorine content (wt%) Ppa: paper & cardboard (wt%)
H: hydrogen content (wt%) Ppl: plastics (wt%)
LHV: lower heating value (kcal/kg) Pte: textiles (wt%)
LHVdna: lower heating value on a dry, no-ash basis (Btu/lb) Pwo: wood (wt%)
O: oxygen content (wt%) S: sulfur content (wt%)
Pfo: food waste (wt%) V: physically combustible component content (%)
Pga: garbage (wt%; textiles, wood, food waste, miscellaneous also included) W: moisture content (%)
Equation references:
Eq. (1): JNMSWF (1991) Eq. (6): Wilson (1977)
Eq. (2): JNMSWF (1991) Eq. (7): Wilson (1977)
Eq. (3): JNMSWF (1991) Eq. (8): Wilson (1977)
Eq. (4): Liu et al. (1996) Eq. (9):Cooper et al. (1999)
Eq. (5): Lin (2000)

tted regression line. The value of R2 increases as more
explanatory variables are added to the regression model. It
cannot represent the true explanatory power of the
regression model. To eliminate this shortcoming, an
2 
adjusted coefcient of multiple determination, R , was
introduced and dened as follows.
2 SSE=n  k  1
R 1 , (12)
SST=n  1
where n represents the sample size and k is the number of
independent variables in the model.
In the present study, the independent variables selected
were weight percents of paper and cardboard, plastics,
textiles, wood, food waste, rubber and leather, miscella-
neous materials, and moisture content, and the dependent
variable was the lower heating value. Based on Eq. (10), the
prediction model can be rewritten as
LHV b1 Ppa b2 Ppl b3 Pte b4 Pwo b5 Pf0

100  W
b6 Pru b7 Pmi  6W , 13
100
where Pru represents the percentage of rubber and leather in
MSW and the denitions of the rest of the parameters are
given in Table 1. The coefcients of b1 , b2 , b3 , b4 , b5 , b6 ,
and b7 represent the regression coefcients corresponding
to the above physical components, respectively. During
20012002, 180 samples were collected from 45 towns
and counties around Taiwan. The coefcients of the
prediction equation could be determined from the analytical
 ARTICLE IN PRESS
894 Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899

measurements of those samples by using multiple regression

analysis. The details of the sampling and analytical procedure
of the MSW will be described later. The coefcients of
Eq. (13), bk, were further examined using the t-test to ensure
the correlation between the dependent variable and bk. The
procedure followed the test statistic process by Mann (1995).

2.1.2. Evaluation of the prediction model

To evaluate the prediction model developed in the
present study, two methods were employed to estimate
the relative error between the predicted LHV (Xp) and the
measured LHV (Xm). The results were compared with
those from the conventional prediction equations. As
dened in the following, the relative percentage deviation
(RPD) was employed to evaluate the relative differences
between the two arbitrary data values; Xa and Xb.
jX a  X b j
RPD% 1  100%. (14)
2 X a X b

In addition to Xp and Xm, the RPD also served to

evaluate the uniformity of the MSW samples during the
sampling and analysis processes. The lower the RPD, the
less the difference between Xa and Xb will be.
In comparison, the mean absolute percentage error
(MAPE) was also used and dened as
n 
 
1X X pi  X mi 
MAPE  100 (15)
n i1 X mi
where n is the data number. The results of the MAPE
examinations can be classied into four types, which are
excellent (MAPEo10), good (MAPE 1020), acceptable
(MAPE 2050), and unacceptable (MAPE450).
2.2. Data sampling and analysis
2.2.1. Sampling
The main purpose of the present study was to establish a
quicker and cheaper model for predicting the LHV of MSW.
To conduct multiple regression analysis, 180 samples were
collected quarterly from 45 towns and counties in Taiwan
from April 2001 to March 2002. With regard to sampling
representation, there were 45 sampling locations and 180
samples, including cities, towns and counties all over
Taiwan, as shown in Fig. 1. To ensure fair representation
of the waste stream, the minimum number of samples
should be determined before sampling. According to Mann
(1995), the sample size can be estimated using Eq. (16) if the
population distribution is a standard normal distribution;
Z 2 s2
n , (16)
E 2p
where n is the number of samples. Z and Ep denote the
values of the standard distribution and the predetermined
maximum error, respectively. s is the standard deviation of
the population. Mann (1995) suggested that when s cannot
be estimated, it should be 0.5 for a conservative estimate.
Fig. 1. Sampling locations around Taiwan.
For a 90% condence level and a 5% maximum error, the Z
value was found to be 1.282 and n was 164. This shows that,
in the present study, 180 samples was sufcient to represent
the MSW of Taiwan.
2.2.2. Sample analysis
The characteristics of the collected samples were
examined according to the standard sampling and analysis
procedures (ASTM, 1997ac; Taiwan Environmental
Protection Agency, 1994) by EPA, Taiwan. The analyses
included physical component contents, proximate analysis,
ultimate analysis, and lower heating value. The procedure
used for sampling and analyzing is shown in Fig. 2.
In the individual sampling, if the refuse generation rate
was less than 300 ton/day, 5 ton of MSW were selected
randomly. The amount of MSW selected would increase by
2.5 ton for every increment of 100 ton/day. Before sam-
pling, the plastic bags were torn by hand throughout.
A shovel was used to mix the sample by moving the waste
from one place to another place at least four times. Then,
the sample was quartered two or three times. To examine
the uniformity of the MSW samples, bulk density tests
were carried out twice for every sampling site. If the RPD
of the two bulk densities was less than 10%, it implied that
the sample was well mixed and acceptable for sequential
analysis. 2040 kg of MSW was collected for the laboratory
analysis. The collected samples were dewatered and
 ARTICLE IN PRESS
Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899 895

Table 2
Physical components contents of municipal solid waste in Taiwan

Physical components Percentage by weight

Range Average

Physical combustible components 86.8

Paper & cardboard 1354 28.4
Plastic 741 19.7
Textiles 021 5.8
Wood & weed 110 4.1
Food waste 845 22.7
Rubber & leather 09 0.4
Miscellaneous material 022 5.7
Physical noncombustible components 13.1
Metal 017 4.1
Glass 021 5.7
Fig. 2. Procedure of sampling and analyzing of MSW. Ceramics 012 1.0
Stone & sand 020 1.7
manually separated to determine the contents of water and Others 018 0.6
physical components. The sample size was reduced to
around 12 mm by a shredder. To further improve the
From the results of proximate analysis, as shown in
uniformity of the MSW samples, a ball grinder was used to
Table 3, it was found that the moisture content was 55.6%,
reduce the particle size to around 1 mm for the examination
higher than the combustible composition (32.1%) and ash
of energy content and ultimate analysis. A calorimeter
(12.3%). Usually, the moisture content of MSW in Taiwan
(IKA C4000, Germany) was used for the determination of
is much higher than that in European and American
LHV with an individual sample weight of 1 g. The organic
countries, which have only about 2030%. The differences
elements such as carbon, hydrogen, oxygen, nitrogen,
in climate and life style cause the higher moisture content
sulfur were analyzed by an Elemental Analyzer (Vario EL
in Taiwans MSW. Due to the higher moisture content, the
III, Germany). To ensure the data quality, a duplicate
LHV was rather low and was found to be about 1434 kcal/
analysis was conducted for every sample. When the RPD
kg. Table 3 also presents the results of ultimate analysis
of the data from the duplicate analyses exceeded 15%, the
(C, H, O, N, S and Cl). It reveals that carbon, oxygen and
results would be examined carefully, and the sample
hydrogen were the major organic fractions, while the minor
analyzed again if necessary. The experimental results were
fractions were nitrogen, sulfur, and chlorine. The contents
then employed to conduct the multiple regression analysis.
of the elements C, O, H, N, S and Cl, in decreasing order,
were 17.3%, 11.5%, 2.6%, 0.5%, 0.1% and 0.1%,
3. Results and discussion respectively. The contents of some of the elements
discussed above could be introduced into the LHV
3.1. Characteristics of MSW prediction model in Table 1. The chlorine content was
low and of little signicance to energy contribution.
Table 2 shows the physical composition of the 180 MSW However, chlorine played an important role in dioxins
samples used in the present study. The average physical formation and corrosion in the incineration exhausts.
combustible and non-combustible fractions of the MSW
were 86.8% and 13.1%, respectively, whereas the former
was only 82.94% in 1992. The variation was mainly caused 3.2. LHV prediction models
by the increase in paper and food waste. Among the
combustible components, the most important fractions To develop the LHV prediction model based on the
were paper and cardboard (28.4%), food waste (22.7%), results of physical and chemical analysis, the physical
and plastics (19.7%). Since the sum of these physical component contents of paper and cardboard, plastics,
compositions was 70.8%, it implied that energy was mainly textiles, wood, food waste, other miscellaneous materials
contributed by these components. Therefore, it was and moisture content were used as independent variables,
reasonable to assume that there was a relationship between with LHV as the dependent variable. By means of stepwise
these three components and the LHV of MSW in Taiwan. regression, two types of LHV empirical prediction models
It can be seen in Table 2 that the percentages of textiles, have been developed and established. The original pro-
wood & weed, and miscellaneous material ranged from 4 to posed model (OPM) includes all of the analyzed physical
6%. The minimum was for leather and rubber, which was components. However, from Table 2, 70.8% of the
only 0.4%. In the non-combustible fraction, the largest one physical content consists of paper and cardboard, food
was glass (5.7%), the second was metal (4.1%), and the rest waste, and plastics. The simplied model (SM) based on
were relatively low in the MSW. these components will be also discussed.
 ARTICLE IN PRESS
896 Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899

2
3.2.1. Original proposed model in R . The reason for doing so was that when more
The results from the calorimeter measurements and variables are used for the model regression, R2 increases
proposed regression models are presented in Fig. 3. The (Mann, 1995). To eliminate the difference due to the
predicted values are quite consistent with the experimental variables used in the OPM and SM, the adjusted coefcient
2
measurements. As shown in Table 4, the coefcient of of multiple determination, R , was employed to evaluate
multiple correlation (R) is 0.992, within the 95% con- the OPM and SM. In the present study, the adjusted R2 is
dence interval. This indicates that the variations in the 0.9827 and it implies that the OPM developed is suitable
dependent variable (LHV) can be well-explained by the for the LHV prediction. In the analysis of physical
independent variables in the OPM. In the work of Liu components, the content of leather and rubber was less
(1996), the R of Eq. (4) based on 34 samples was reported than 1% and the result of the t-test ratio was only 0.314,
to be 0.983. Further breaking the physical component which was smaller than t(0.95; 172) 1.65. This indicates
(garbage) into textiles, wood, food waste, and miscella- that the parameter of leather and rubber could be removed
neous materials, as employed in Lius model, the precision from the OPM without inuencing the precision signi-
of the prediction model is improved signicantly. Cer- cantly. After the multiple regression, the coefcients of the
tainly, more samples improved the correlation coefcient in LHV empirical prediction model in Eq. (13) could be
the present study. determined and rewritten as
Furthermore, taking the degrees of freedom into
consideration, an adjustment was made to R2, resulting LHV 35:19Ppa 71:17Ppl 36:24Pte 48:06Pw0
 
100  W
42:21Pf0 44Pmi  6W . 17
100
Table 3
Chemical characteristics and energy contents of the MSW in Taiwan

Range Average
3.2.2. Simplified model
Proximate analysis (%) In order to develop a quicker and cheaper method for
Moisture 3277 55.6 the determination of the LHV of MSW, the physical
Ash 234 12.3
Combustible composition 1758 32.1
components of the OPM were reduced to form a SM.
According to the dry-basis analysis of physical combustible
Ultimate analysis (%) fractions shown in Table 2, the total content of paper and
C 6.532.8 17.3
O 4.7631.1 11.5
cardboard, food waste and plastics was 70.8%, i.e., most of
H 1.24.9 2.6 the LHV should be contributed by these three components.
N 0.141.55 0.5 The energy content from wood & weeds, textiles, and
S 0.010.51 0.1 miscellaneous materials is of little importance and can be
Cl 0.040.28 0.1 ignored in constructing the SM. Therefore, the physical
Energy content (kcal/kg) components of waste paper & cardboard, plastic and food
LHV 3903035 1434 waste, and the moisture content were used for the
development of the SM for LHV prediction by MRA.

a b
4000 4000
Predicted LHV by original proposed model (kcal/kg)

Predicted LHV by simplified model (kcal/kg)

Original proposed model Simplified Model

3500 3500
Line of best fit Line of best fit
3000 3000

2500 2500

2000 2000

1500 1500

1000 1000

500 500

0 0
0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000
LHV by calorimeter (kcal/kg) LHV by calorimeter (kcal/kg)

Fig. 3. Comparisons of LHVs from calorimeter and prediction models.

 ARTICLE IN PRESS
Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899
Table 4
Statistical results and coefcients of LHV empirical model
Model type Physical component Estimated
coefcient
OPM Plastics 71.17
Food waste 42.21
Paper & cardboard 35.19
Miscellaneous material 44.00
Wood & weed 48.06
Textiles 36.24
SM Plastics 101.47
Food waste 38.47
Paper & cardboard 39.04
When neglecting some of those parameters, Eq. (13) could
be rearranged to obtain the SM as follows:
LHV 39:04Ppa 101:47Ppl 38:47Pf0
 
100  W
  6W .
100
From Table 4, it was found that the adjusted R2 of the SM
was only slightly smaller than that of the OPM. Obviously,
neglecting wood & weeds, textiles, and miscellaneous
materials did not severely inuence the correlations
between the experimental results and the predicted data,
as shown in Fig. 3(b). The predicted results showed a
positive correlation with the experimental measurements
with a slight underestimation.
3.3. Evaluation of prediction models
3.3.1. Relative percentage deviation
To evaluate the precisions of the prediction models, a
demonstration program was conducted with 12 samples
taken from Kaohsiung City, Taiwan. The procedures for
sampling and data analysis were the same as those in the
present study. As shown in Fig. 4, the LHV estimated by
the previous prediction models was compared with the
calorimetric results. Generally, the trend of predictions was
consistent with the measurements. However, the predic-
tions from Eqs. (1) to (5) underestimated the LHV to
different degrees. Obviously, this situation can be improved
by using the OPM and SM, as shown in Fig. 5. The RPDs
between LHVs determined by the calorimeter and the
prediction models can be used to quantitatively examine
the applicability of these models. The results are listed in
Table 5. Among the models based on proximate analysis,
the RPD for Eq. (2) was less than 22% and showed better
precision than that of Eq. (1). This implies that the
empirical model offers more satisfactory precision. In
Eqs. (3)(5), based on physical combustible components,
the RPD of Eq. (5) was relatively low and showed precision
similar to that of Eq. (2). It is concluded that the
breakdown of physical composition would improve the
precision of prediction models. In Eqs. (6)(9), based on
897
t-test ratio Coefcient of multiple Coefcient of multiple Adjusted R2
correlation, R determination, R2
11.002 0.9923 0.9831 0.9827
12.237
9.442
6.481
4.785
3.693
14.855 0.9874 0.9753 0.9738
10.010
8.698
the content of elements, it was found that the model
proposed by Cooper et al. (1999) and Dulongs formulas,
i.e. Eqs. (9) and (6) showed almost similar ranges of
precision. When compared to the physical models, Eq. (2)
showed a good estimation when compared to those
18
equations based on the ultimate analysis. In general cases,
the RPD of equations based on ultimate analysis was
smaller than that of the other equations. Theoretically, the
LHV of MSW comes from the combustion of organic
elements, and the organic elements certainly provided a
better prediction of the LHVs. However, the physical
compositions of the MSW were determined according to
the appearance, rather than the chemical compositions of
the MSW. This implied that the LHV is not related to the
proportions of the physical components. For instance, the
same percentage of plastics might exhibit different LHVs
because of different chemical compositions. The present
study came to the same conclusion, except that Eqs. (2) and
(4) have similar precision.
Furthermore, Table 5 also shows the RPD of the OPM
and the SM proposed in the present study. These models
were based on physical composition and showed much
better predictions than the results of Eqs. (3) and (5). In
Eq. (4) proposed by Liu, multiple linear regression was also
employed in the determination of the coefcients. It is
obvious that only the three physical compositions included
in the model would inuence the precision. The intercept,
which was not assigned manually to be zero, might also be
an important factor in improving the precision. The
intercept term and the coefcient of moisture content (W)
in Lius model were rather high. In the multiple linear
regression, a minimum SEE was desired and this resulted in
coefcients in Lius model that were different from those of
other equations. As a consequence, the RPD and MAPE of
Lius model were also higher than the other equations.
In the present study, multiple regression analysis was
employed to determine the coefcients for the same
physical compositions of Eq. (5). As depicted in Table 4,
the RPD of the OPM was further decreased and the
precision was even better than that from the ultimate
analysis. In the SM, the RPD increased from 5.6 to 11.4
which indicated that neglecting minor physical compositions
 ARTICLE IN PRESS
898 Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899

Predicted LHV by Equation (2) (kcal/kg)

Predicted LHV by Equation (1) (kcal/kg)

Predicted LHV by Equation (3) (kcal/kg)

1900 1900 1900
Equation (1) Equation (2) Equation (3)
1700 Line of best fit 1700 1700
Line of best fit Line of best fit
1500 1500 1500
1300 1300 1300
1100 1100 1100

900 900 900

700 700 700

500 500 500

500 700 900 1100 1300 1500 1700 1900 500 700 900 1100 1300 1500 1700 1900 500 700 900 1100 1300 1500 1700 1900
LHV by calorimeter (kcal/kg) LHV by calorimeter (kcal/kg) LHV by calorimeter (kcal/kg)
Predicted LHV by Equation (4) (kcal/kg)

Predicted LHV by Equation (5) (kcal/kg)

Predicted LHV by Equation (6) (kcal/kg)

1900 1900 1900
Equation (4) Equation (5) Equation (6)
1700 Line of best fit 1700 Line of best fit 1700 Line of best fit
1500 1500 1500

1300 1300 1300

1100 1100 1100

900 900 900

700 700 700

500 500 500

500 700 900 1100 1300 1500 1700 1900 500 700 900 1100 1300 1500 1700 1900 500 700 900 1100 1300 1500 1700 1900
LHV by calorimeter (kcal/kg) LHV by calorimeter (kcal/kg) LHV by calorimeter (kcal/kg)

Predicted LHV by Equation (9) (kcal/kg)

Predicted LHV by Equation (7) (kcal/kg)

Predicted LHV byEquation (8) (kcal/kg)

1900 1900 1900 Equation (9)

Equation (7) Equation (8)
1700 Line of best fit 1700 Line of best fit 1700 Line of best fit
1500 1500 1500

1300 1300 1300

1100 1100 1100

900 900 900

700 700 700

500 500 500

500 700 900 1100 1300 1500 1700 1900 500 700 900 1100 1300 1500 1700 1900 500 700 900 1100 1300 1500 1700 1900
LHV by calorimeter (kcal/kg) LHV by calorimeter (kcal/kg) LHV by calorimeter (kcal/kg)

Fig. 4. Comparisons of LHVs by calorimeter and predictions for demonstration program.

a b
Predicted LHV by simplified model (kcal/kg)
Predicted LHV by original proposed model

1900 Original proposed model 1900 Simplified Model

Line of best fit Line of best fit
1700 1700

1500 1500
(kcal/kg)

1300 1300

1100 1100

900 900

700 700

500 500
500 700 900 1100 1300 1500 1700 1900 500 700 900 1100 1300 1500 1700 1900
LHV by calorimeter (kcal/kg) LHVby calorimeter (kcal/kg)

Fig. 5. LHV predictability of OPM and SM for demonstration program.

 ARTICLE IN PRESS
Y.F. Chang et al. / Journal of Environmental Management 85 (2007) 891899 899

Table 5 with a MAPE value of 5.56. In the simplied version,

Relative percentage deviations of LHV for the demonstration program the MAPE is 10.70, which is on the same level as the others.
Model RPD
This implies that estimating the LHV of MSW by the
SM is a good alternative in view of economics and
Range (%) Average (%) efciency.
Proximate analysis
Eq. (1) 13.738.5 27.3 4. Conclusions
Eq. (2) 7.421.2 12.7

Physical composition analysis

To establish precise and economical prediction models
Eq. (3) 0.723.5 15.1 for the LHV of MSW by MRA, empirical models, based
Eq. (4) 30.749.1 40.0 on 180 samples gathered from different cities and counties
Eq. (5) 3.719.1 11.3 in Taiwan during 20012002, were proposed in the present
Ultimate analysis study. From the demonstration program, the OPM showed
Eq. (6) 1.027.9 9.9 superior precision in view of the RPD and MAPE, when
Eq. (7) 1.127.2 11.1 compared to the models based on the proximate analysis,
Eq. (8) 5.833.4 18.6 physical composition and ultimate analysis. When neglect-
Eq. (9) 0.623.4 7.6
ing the three minor physical compositions, the SM would
Proposed model lower the precision. However, it still maintained an
OPM 0.310.5 5.6 acceptable precision level, with results better than the
SM 2.319.7 11.4
other models. It can be concluded that the predictability of
LHV empirical models can be improved signicantly
through correct selection of the major physical composi-
Table 6 tions and multiple regression analysis.
MAPE of LHV prediction models
References
Prediction equations

(1) (2) (3) (4) (5) (6) (7) (8) (9) OPM SM ASTM, 1997a. Isothermal jacket bomb calorimeter, http://www.astm.org.
ASTM, 1997b. D5468-95 standard test method for gross caloric and ash
MAPE 23.87 12.11 13.91 33.18 10.67 9.93 11.34 20.97 7.62 5.56 10.70 value of waste materials, http://www.astm.org.
P
n ASTM, 1997c. D711-87 standard test method for gross caloric value of
Remark: MAPE n 1 jX pi X mi j  100, Xp, predicted LHV, Xm, measured
X mi refuse-derived fuel by bomb calorimeter, http://www.astm.org.
i1
LHV. Cooper, C.D., Kim, B., MacDonald, J., 1999. Estimating the lower
heating values of hazardous and solid wastes. Journal of the Air &
lowered the precision; however, the RPD of the SM was Waste Management Association 49, 471476.
Japan National Municipal Solid Waste Foundation, 1991. Design Guide
still maintained satisfactorily. This suggests that the
for the Facility of Solid Waste Disposal, Tokyo.
proposed SM could be used to estimate the energy content Lin, C.H., 2000. Model to evaluate lower heat value of solid
of MSW in a quicker and cheaper way without lowering waste in Taiwan, Tung Nan Institute of Technology, Taipei, Taiwan,
the precision signicantly when compared to the OPM and ROC.
the historical models. Liu, J.I., Paode, R.D., Holsen, T.M., 1996. Modeling the energy content
of multiple regression analysis. Journal of the Air & Waste Manage-
ment Association 46, 650656.
3.3.2. MAPE of the models Mann, P.S., 1995. Statistics for Business and Economics. Wiley, New York.
The MAPE, as shown in Table 6, can be used as an index Navarro-Esbr , J., Diamadopoulos, E., Ginestar, D., 2002. Time
to evaluate the equations integrally. From the comparisons series analysis and forecasting techniques for municipal solid
of MAPEs, it shows variations similar to those of RPDs waste management. Resources, Conservation and Recycling 35 (3),
in the prediction models based on different analyses. 201214.
Taiwan Environmental Protection Agency, 1994. Handbook of Solid
Better MAPEs for the proximate, physical composition Waste Assay. Taiwan Government Printing Ofce, Taipei.
and ultimate analyses ranged from 7.62 to 12.11. The Wilson, D.G., 1977. Handbook of Solid Waste Management. Van
OPM proposed shows a precision superior to the others, Norstrand Reinhold, New York.