Waste Management 48 (2016) 14–23

Contents lists available at ScienceDirect

Waste Management
journal homepage: www.elsevier.com/locate/wasman

Verifying the performance of artificial neural network and multiple

linear regression in predicting the mean seasonal municipal solid waste
generation rate: A case study of Fars province, Iran
Sama Azadi, Ayoub Karimi-Jashni ⇑
Department of Civil and Environmental Engineering, School of Engineering, Shiraz University, Shiraz, Fars 71348-51156, Iran

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

Article history: Predicting the mass of solid waste generation plays an important role in integrated solid waste manage-
Received 10 April 2015 ment plans. In this study, the performance of two predictive models, Artificial Neural Network (ANN) and
Revised 18 July 2015 Multiple Linear Regression (MLR) was verified to predict mean Seasonal Municipal Solid Waste
Accepted 24 September 2015
Generation (SMSWG) rate. The accuracy of the proposed models is illustrated through a case study of
Available online 9 October 2015
20 cities located in Fars Province, Iran. Four performance measures, MAE, MAPE, RMSE and R were used
to evaluate the performance of these models. The MLR, as a conventional model, showed poor prediction
Keywords:
performance. On the other hand, the results indicated that the ANN model, as a non-linear model, has a
Artificial neural network
Seasonal municipal solid waste generation
higher predictive accuracy when it comes to prediction of the mean SMSWG rate. As a result, in order to
Fars Province develop a more cost-effective strategy for waste management in the future, the ANN model could be used
Multiple linear regression to predict the mean SMSWG rate.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction Intharathirat et al., 2015; Mateu et al., 2013; Lebersorger and

Beigl, 2011; Boer et al., 2010; Boldrin and Christensen, 2010).
The Municipal Solid Waste (MSW) – more commonly known as Gomez et al. (2009) studied MSWG in three neighborhoods of
trash or garbage – is one of the most challenging problems in mod- Chihuahua, Mexico with different income levels in three seasons
ern societies, especially in many developing countries. As solid of year. Their results showed that higher income generally leads
waste management is a relatively new subject in developing to the greater amounts of SWG and less waste generated during
countries, in order to develop and implement effective solid waste winter.
management programs it is necessary to have an accurate and An integrated solid waste management system requires the
reliable prediction of the mass of Municipal Solid Waste precise prediction of SWG rate. One common approach in predict-
Generation (MSWG) (Batinic et al., 2011; Kumar et al., 2011; ing SWG is the consideration of the trend of waste generation rate
Cherian and Jacob, 2012; Intharathirat et al., 2015). The quantity variation along with time in a target area. Unfortunately, there is a
of generated and collected solid wastes is of great importance in lack of comprehensive historical records on the SWG rate in many
the solid waste management program (Tchobanoglous et al., developing countries due to insufficient funds, inadequate man-
1993; Fu et al., 2015). agement and limited measurement instruments in many cases
Solid Waste Generation (SWG) rate is affected by a variety of (Rimaityte et al., 2011; Dyson and Chang, 2005). Thus, the predic-
factors including the geographical location, season, use of kitchen tion of the SWG rate by common methods, which consider the
food waste grinders, collection frequency, characteristics of trend of the waste generation rate variation along with its time,
service area, onsite processing, people’s food habits, economic is quite challenging. In order to tackle this problem effectively, it
conditions, recovery and reuse boundaries, existence of solid is suggested that a new approach be developed which does not
waste management laws, local culture and beliefs, population require the trend of waste generation rate variation along with
growth, weather conditions and household size (Tchobanoglous its time to predict the SWG rate.
et al., 1993; Abdoli et al., 2011; Chen, 2010; Keser et al., 2012; Various studies have made use of many forecasting techniques
such as multiple linear and nonlinear regression techniques and
⇑ Corresponding author. expert systems to simulate SWG (Jalili Ghazi Zade and Noori,
E-mail address: akarimi@shirazu.ac.ir (A. Karimi-Jashni). 2008; Kollikkathara et al., 2010; Dyson and Chang, 2005; Lohani

http://dx.doi.org/10.1016/j.wasman.2015.09.034
0956-053X/Ó 2015 Elsevier Ltd. All rights reserved.
 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23 15

and Hartono, 1985). Lohani and Hartono (1985) formulated several In the second step, ANN and MLR models were developed. In this
equations for determining SWG rates in different domestic and step, the most suitable structure of ANN model for the mean
industrial sources in the city of Bandung in Indonesia using MLR SMSWG rate prediction was specified. In addition, before develop-
model. The authors concluded that the presented formulas were ing MLR model, principal component analysis was conducted on
useful for the prediction of SWG rates in Bandung. Dyson and the explanatory variables to remove their multicollinearity. In
Chang (2005) used a dynamic system modeling technique to pre- the third step, the performances of the applied models were com-
dict the mass of MSWG in San Antonio, Texas. They concluded that pared, and based on statistical tests, the most accurate model for
dynamic modeling can be used when there is a lack of complete the prediction of the mean SMSWG rate was determined. More
historical records on SWG, and traditional statistical least- details on the components of the proposed methodology will be
squares regression methods are not applicable. Jahandideh et al. discussed in the following sections.
(2009) used ANN and MLR models to predict the medical waste
generation rate of 50 hospitals in Fars Province, Iran. The results
showed that ANN could predict SWG rate reasonably well, while 2.2. The study area and predictor variables
the prediction achieved by the MLR model was not adequately
accurate. Jalili Ghazi Zade and Noori (2008) investigated different Fars Province is located in the southwest of Iran from 27°030 and
structures of feed forward ANN model to predict the SWG volume. 31°420 N and 50°300 and 55°360 E (see Fig. 2). Fars Province has an area
They concluded that a network with 16 neurons in the hidden layer of about 122,608 km2, covering approximately 8.1% of Iran
was suitable for the prediction of the weekly volume of waste gen- (Nafarzadegan et al., 2010). According to the latest figures provided
erated in Mashhad, Iran. Batinic et al. (2011) used ANN model to by the Census Bureau of Iran (2011), it has a population of about
predict the amount and composition of SWG in Serbia. They indi- 4.58 million, constituting about 6.15% of the total population of Iran.
cated that ANN model was capable and suitably effective in this About 61.2% and 38.1% of the population of Fars live in urban and
assignment. Abdoli et al. (2011) applied ANN and MLR models to rural areas, respectively, and about 0.7% are nomads. Due to the vast-
a long-term (over 22 years) prediction of SWG mass by analyzing ness of the province and the presence of the Zagros Mountains to the
monthly series of datasets over the first decade of the 21st century north and northwest, the plain of Khuzestan to the west, and the Per-
(2000–2010) in Mashhad, Iran. The results showed that ANN can sian Gulf to the south, the climate and weather conditions are highly
follow trends and fluctuations more accurately than MLR model. variable. The difference in average annual rainfall between the
It is worth mentioning that MLR techniques, which are widely northern and southern parts of the province exceeds 300 mm.
applied in forecasting MSWG, are mainly used to model a linear The changes of the mean SMSWG rate in spring, summer,
relationship between a dependent variable and one or more inde- autumn and winter of 2009–2010 for 20 cities in Fars Province
pendent variables (Xu et al., 2013; Thanh et al., 2010). Thus, non- are shown in Fig. 3. As can be seen, the mean SMSWG rate in all
linear and dynamic modeling techniques such as fuzzy systems, cities approximately follows the same trend within the year and
evolutionary programming, ANN and the combination of these it can be said that the per capita generation of MSW increases in
methods are essential for the development of accurate predictive summer which is due to the dietary habits of the people, the rise
models (Intharathirat et al., 2015; Patel and Meka, 2013). in tourist trips, and the availability of more fruits and vegetables.
Interest in using ANNs for forecasting has increased in research However, the appropriate measures can be taken by predicting
activities during the past decade (Patel and Meka, 2013; Roy et al., the mean SMSWG rate in order to determine the frequency of
2013; Shahabi et al., 2012; Jalili Ghazi Zade and Noori, 2008; waste collection and estimate the number of required trucks to
Batinic et al., 2011). The ANNs are a class of non-linear mathemat- carry the waste in different seasons and finally reduce the costs.
ical models which were developed on the basis of human neural In this study, based on the correlation test and previous related
system. The ANN models are characterized by architectures and literature, four influential variables on SMSWG rate whose data
structures of interconnected computational elements, i.e. neurons. were available were chosen as the explanatory variables of our
These architectures enable ANN to learn from examples and study. The selected variables were required to be able to be fore-
capture functional relationships among the data and can often casted with a relatively high accuracy in a specified forecasting
correctly infer the unseen part of sample data even if the data horizon.
contain some noise (Zhang et al., 1998). The first selected variable was the population of each city
In this study, two predictive models ANN and MLR were applied because the consumption pattern changes as the population grows
to verify the performance of the models for the prediction of mean and therefore the per capita generation of MSW increases (Abdoli
SMSWG rate in 20 urban regions of Fars Province, Iran. Five et al., 2011; Dyson and Chang, 2005). The second variable consid-
influential variables on the mean SMSWG rate, population, solid ered as an explanatory variable was the frequency of waste collec-
waste collection frequency, maximum seasonal temperature and tion, because if the frequency of waste collection in a city increases,
altitude were chosen as the input data of our study based on the collected waste will increase (Tchobanoglous et al., 1993; Keser
correlation test. et al., 2012). The third selected parameter was the maximum sea-
sonal temperature. This parameter which was previously utilized
2. Material and methods as an effective parameter on SWG rate by Abdoli et al. (2011)
was used to show the change of seasons. On the other hand, one
2.1. Model framework of the parameters that affect the solid waste generation rate in dif-
ferent cities is the climate of each city. So, the fourth selected vari-
Fig. 1 shows the main components of the proposed methodol- able was the altitude of each city because this parameter can be a
ogy to predict the mean SMSWG rate. As shown in this figure, sign of changing weather conditions as well as the geographic and
the methodology includes three main parts: (1) specifying climatic conditions of each city.
explanatory variables and data collection, (2) developing ANN Hence, the four explanatory variables population (Pop), solid
and MLR models, training and testing them, (3) statistically analyz- waste collection frequency (Freq), maximum seasonal temperature
ing the results and determining the most accurate predictive (Tmax) and altitude (Alt) were considered for the prediction of the
model. In the first step, effective parameters for the simulation of mean SMSWG rate. Thus, the 2009–2010 data for 20 urban areas
the mean SMSWG rate were investigated, and some measurable of Fars Province which had air station were used in this study.
parameters and those whose data were available were selected. Table 1 indicates the statistical characteristics of the data. Also,
16 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23

Fig. 1. Flowchart of proposed methodology for determining the best model to predict the mean SMSWG rate.

correlation among the four explanatory variables and the mean nodes of the adjacent layer and the hidden layer, and all of the
SMSWG rate are shown later in Table 3. interconnections have a weight which represents the strength of
the connection. In this regard, in the hidden layer a summation
2.3. Artificial neural network of weights is passed through a function called the activation
function which determines the relationship between the inputs
A particular structure of ANNs is Feed-Forward Neural Network and outputs (Zhang et al., 1998; Wang et al., 2015). The output yj
(FFNN) which is popular and widely applied in many applications of any node j is given as Eq. (1) (Kocadağlı, 2015):
including forecasting (Wang et al., 2015; Azadi and Sepaskhah, !
2011). A one-hidden layer FFNN is shown in Fig. 4. X
m
yj ¼ f W ji X i þ bj ð1Þ
The input data are introduced to the network through the input i¼1
layer which acts as a buffer layer (Azadi and Sepaskhah, 2011). The
number of nodes in this layer is equal to the input data or the num- where Xi is the output of the ith connecting node in the previous
ber of independent variables. These nodes are connected to the layer; Wji is the weight of the interconnection between the ith node
 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23 17

Fig. 2. The geographic location of study area.

800
2010). Detailed theoretical discussions about FFNNs can be found
Abadeh in ASCE Task Committee (2000).
700
Arsenjan In this study, four independent variables population, solid
Estahban
Eghlid
waste collection frequency, maximum seasonal temperature and
altitude for 20 urban areas of Fars Province were used. The mean
SMSWG rate (gr/capita.day)

600 Izad khast

Bavanat SMSWG rate in these areas, as the target parameter, was intro-
500 Jahrom duced to ANN. Hence, 80 datasets of one year were employed in
Darab
Zarghan
modeling. Experimental results and theoretical works have shown
400
Sepidan that a single hidden layer is enough to approximate the most likely
Sad Dorodzan problems and any complex nonlinear functions with any desired
300 Farashband
accuracy (Wang et al., 2015; Zhang et al., 1998). For this reason,
Fasa
Firuz Abad
one-hidden layer FFNN was used in this study for the prediction
200
Ghir va Karzin of the mean SMSWG rate. The number of nodes in each hidden
Lar layer is dependent on the complexity of the process which is mod-
100 Lamerd
eled and usually determined by trial and error (Wang et al., 2015;
Noor Abad
0 Neyriz
Azadi and sepaskhah, 2011). In the case of one hidden layer net-
Spring Summer Autumn Winter work, some useful guidelines such as ‘‘2n + 1”, ‘‘2n”, ‘‘n”, ‘‘n/2” have
been presented for choosing the number of neurons in the hidden
Fig. 3. The seasonal variation of the mean MSW generation rate in 2009–2010 for
layer where ‘‘n” is the number of independent variables (Zhang
20 cities in Fars Province.
et al., 1998). However, none of these guidelines are applicable for
all problems. So, based on these guidelines and considering a con-
fidence interval, this number varied from 1 to 15 and then the
in the previous layer, and the jth node; and bj is the node threshold. results of prediction error were compared to find the optimum
Afterward, the outputs of the hidden layer are fed into the output number of hidden neurons.
layer and then multiplied by their associated weights and trans- In theory, any differentiable function can be used as an
formed through a linear activation function in the output layer. activation function. Yet in practice, only a small number of
The number of the output nodes is directly related to the problem well-behaved activation functions, including log-sigmoid (logsig),
in question (Zhang et al., 1998; Wang et al., 2015; Beale et al., hyperbolic tangent sigmoid (tansig), sine or cosine and linear

Table 1
Statistical characteristics of the data.

Variable Unit Mean SD Minimum Maximum

Pop Person 44723.05 31963.96 1627.00 112054.00
Freq Times per week 3.25 1.07 2 6
Tmax °C 35.37 8.24 18.60 49
Alte m 1491.00 660.85 425.00 3119.00
The mean SMSWG rate g/capta.day 498.41 88.84 229 677
18 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23

Table 2
Some important advantages and disadvantages of the four accuracy measures.

Accuracy Advantages Disadvantages

measures
MAPE  Provides an easy and intuitive method of judging the importance of errors  Lacks a statistical theory on which to base itself
 Being relative, it allows us to average it across forecasting horizons and series  Equal errors when P iobs is larger than P isim and gives smal-
 Ability to make comparisons involving more than one method since the MAPE of each ler percentage errors than when P iobs is smaller than P isim
tells us about the average relative size of their errors
MAE  Its meaning is more intuitive because it is a linear measure  The most salient disadvantage of MAE is that it is an
 The average size of forecasting errors when negative signs are ignored absolute measure
 Provides a linear weighting between large and small
errors
RMSE  Provides a quadratic loss function  Emphasizes larger errors than smaller ones
 It is a measure of the uncertainty in forecasting  More sensitive to the presence of erroneous data when
compared to MAE
NRMSE  Its meaning is more intuitive, since it is a non-dimensional parameter  Drawbacks are similar to RMSE disadvantages
 Provides a quadratic loss function
 It is a measure of the uncertainty in forecasting
R  It is a measure of the strength and direction of the linear relationship between pre-  A correlative finding does not reveal which variable
dicted and observed values influences the other
 Provides an easy and intuitive way of judging the correlation between two variables  If the relationship is not linear then the result is
inaccurate

Table 3 trapped into the local minima. Thus, in this study, in order to
Correlation coefficients between explanatory variables and dependent variable. alleviate the problem of local minima, for each structure of the
Pop Freq Tmax Alt neural network, the network was trained 40 times and each time
SMSWG 0.61 0.42 0.48 0.33
the initial random weights were different. Then, according to their
performances on the test dataset, the most desirable solutions of
the weights were selected competitively from all the other
solutions.
(purelin) function are applied (Zhang et al., 1998; Wang et al., Several training algorithms for training network and updating
2015). In this research, the logsig and tansig function in the hidden weights exist and selecting the most appropriate one for a partic-
layer and a pure linear function in the output layer were used. ular ANN model in advance is difficult. In this study, five common
Then prediction errors of each state were compared to determine training functions with special characteristics Levenberg–
which activation function in the hidden layer is more suitable for Marquardt (LM), Bayesian Regulation (BR), Gradient Descent with
predicting the mean SMSWG rate in Fars Province. Momentum (GDM) (with learning rate and momentum equal to
In the neural network training process, the weights of a net- 0.01 and 0.9, respectively), Resilient back Propagation (RP) (with
work are initialized by random numbers in the interval of 1 to learning rate equal to 0.01) and Scaled Conjugate Gradient (SCG)
+1 and iteratively modified to minimize the total squared error were examined. Then prediction errors of each network were
between the predicted and actual output values using a suitable compared to find the training function that is more suitable for
algorithm. Since all optimization algorithms in practice inevitably predicting the mean SMSWG rate in Fars Province.
suffer from local optima problems, there is no algorithm currently Data normalization is often fulfilled before training process
available to guarantee the global optimal solution for a nonlinear starts, to meet algorithm requirement and to facilitate network
problem (Zhang et al., 1998; Kocadağlı, 2015). Hence, the most that learning (Wang et al., 2015). There are different methods for input
can be done is to use the proposed techniques which provide the normalization such as along channel, across channel, mixed chan-
‘‘best” local optima if the global minima are not available. One of nel and external normalization. For causal forecasting in which
these techniques is starting with appropriate weights. Learning independent variables are applied to predict target variable, along
algorithms such as back-propagation, which is used in this study, or across channel normalization should be used (Zhang et al.,
are quite sensitive to initial weights (Fukuoka et al., 1998). So, 1998). Here, for data normalization, across channel normalization
starting with inappropriate weights is one reason for getting was used as follows (Wang et al., 2015):

Fig. 4. A typical FFNN.

 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23 19

ðx0max  x0min Þðx  xmin Þ represents the contribution degree of each explanatory variable in
xn ¼ þ x0min ð2Þ
ðxmax  xmin Þ a specific PC. Detailed theoretical discussions about PCA can be
found in Abdi and Williams (2010).
in which x is the observed data; xn is the normalized observed data; To develop the MLR model, it is necessary to determine regres-
xmin and xmax are the minimum and maximum observed data, sion coefficients in such a way that they be statistically significant
respectively; and x0min and x0max are the minimum and maximum (Pires et al., 2008). Some methods such as enter, stepwise, back-
considered amount for normalized data, respectively. ward elimination and forward methods can be applied to valid
One of the major concerns about ANN is the overfitting problem regression coefficients. In this study, stepwise method was applied
of the network, meaning that the network seems to improve con- to determine the significant regression coefficients. The application
siderably during the training set but becomes worse during the of these methods leads to the exclusion of explanatory indepen-
unseen examples. To overcome this, early stopping technique is dent variables which have less correlation with the output variable
applied as a form of regularization, which is widely used to solve (Pires et al., 2008).
overfitting problem (Kocadağlı, 2015). In this study, four explanatory independent variables popula-
To perform this technique, the total available data are divided tion (Pop), solid waste collection frequency (Freq), maximum sea-
into a training set, a validation set and a test set. The training set sonal temperature (Tmax) and altitude (Alt) for 20 urban areas of
is applied for ANN model development, the validation set is applied Fars Province were used. The mean SMSWG rate in these areas
for determination of the stopping point of the training process to was considered as dependent variable. Hence, 80 datasets were
avoid the overfitting problem, and the test set is utilized to evalu- employed in modeling. At first, the collinearities of explanatory
ate the forecasting ability of the model (Zhang et al., 1998; Kocadağlı, variables were removed through the application of PCA method.
2015). Here, a maximum of 1000 iterations was considered to Then, to avoid any possible bias in the selection of test set individ-
train the network and the training stopped when the validation uals, a 5-fold cross-validation was fulfilled. So, using 80% of total
error increased for six iterations. The ratio of data division into datasets (64 datasets), constant coefficients of regression equation
the training, validation and test set is a concern and no solution have been computed in such a way that the summation of square
exists for this problem. Most authors choose them randomly based errors is minimized. Then, the prediction was performed on
on the rule of 60% vs. 20% vs. 20%, 80% vs. 10% vs. 10% and 70% vs. remaining 20% of datasets (16 datasets). This process was repeated
15% vs. 15%, etc. In this study, in order to divide 80 datasets into until all cases within the datasets were tested. Finally, the average
three sets, the rule of 60% (training) vs. 20% (validation) vs. 20% results of five different stages were reported.
(test) was applied. To avoid any possible bias in selecting the test
set individuals, a 5-fold cross-validation was conducted. So, 20% 2.5. Performance evaluation
of dataset (as test set) was removed from the dataset and the net-
work was trained and validated using the remaining dataset. Then, Based on training data, prediction accuracy which is the most
the test set was examined by the trained model. This process was important measure of performance can be estimated (Zhang
repeated until all cases within the dataset were tested. Finally, the et al., 1998). The difference between the actual (target) and the
average results of five different simulations were reported. predicted value is its forecasting error which represents the accu-
racy measure. In order to evaluate the performance of the applied
2.4. Multiple linear regression (MLR) models in this study, four accuracy measures have been calculated
as follows:
In MLR method, which was first introduced by sir Francis Galton
in the late 19th century, the relationship between two or more (1) Mean Absolute Percentage Error (MAPE): The closer MAPE is
explanatory or independent variables and a response or dependent to zero, the better the performance of the model is.
variable are modeled by fitting a linear equation to the learning data  
n  i i 
as shown in Eq. (3) (Jahandideh et al., 2009; Shu and Lam, 2011): 1X Psim  Pobs 
MAPE ¼    100 ð4Þ
n i¼1  Piobs 
X
k
y¼c
b0 þ bbi xi þ
ð3Þ (2) Mean Absolute Error (MAE): if MAE approaches zero, it is an
i¼1
indication of the model’s high accuracy.
in which y is the dependent variable, xi (i = 1, 2, . . ., k) are the Pn i
i¼1 jP sim  Piobs j
explanatory independent variables, bb (i = 0, 1, . . ., k) are the regres- MAE ¼ ð5Þ
i n
sion coefficients, and
is the residual error.
(3) Root Mean Square Error (RMSE): RMSE values close to zero
MLR is developed based on two assumptions: (i) the explanatory
also denote good performance on the part of the model.
variables must be independent, and (ii) the dependent variable
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
must be normally distributed with zero mean and constant Pn i i 2
i¼1 ðP sim  P obs Þ
variance. RMSE ¼ ð6Þ
In regression equations the collinearity between the explana- n
tory variables can lead to some drawbacks due to the fact that high
correlations between predictor variables could cause some diffi- (4) Normalized Root Mean Square Error (NRMSE): NRMSE value
culty for a correct analysis (Sousa et al., 2007). To solve this prob- close to 1 indicates a poor model performance, whereas
lem, Principal Component Analysis (PCA) can be applied to remove value close to 0 shows a good model performance
the dependency of the explanatory variables. (Mentaschi et al., 2013).
PCA is a multivariate statistical method widely applied to create vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uPn 2
new variables called Principal Components (PCs) that are uncorre- u ðPisim  Piobs Þ
lated. These independent variables are linear combinations of the NRMSE ¼ t i¼1 Pn 2
ð7Þ
i
original variables and contain the information of the original vari- i¼1 ðP obs Þ

ables with minimum losses (Noori et al., 2010). To evaluate the

influence of each explanatory variable in the PCs, varimax rotation (5) Correlation Coefficient (R): coefficients of 1 or almost 1
matrix can provide useful information. Each element of this matrix indicate that the model can yield reliable results.
20 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23

!
E½ðPsim  Psim ÞðPobs  Pobs Þ correlation coefficient shown in Table 3, it can be concluded that
R¼
rsim robs population and altitude have the most and least correlation with
Pn the mean SMSWG rate, respectively.
ðP i  Psim ÞðPiobs  Pobs Þ After confirming the correlation between the four aforemen-
¼ Pn i¼1i sim Pn i
ð8Þ
i¼1 ðP sim  P sim Þ i¼1 ðP obs  P obs Þ
tioned explanatory independent variables and the mean SMSWG
rate, two predictive models, i.e., ANN and MLR were developed
where P sim is the simulated value of mean SMSWG rate; Pobs and the results of their performance for the prediction of the mean
is the actual value of the mean SMSWG rate; P sim is the mean SMSWG rate are discussed in Sections 3.2 and 3.3, respectively.
amount of total simulated values of the mean SMSWG rate;
Pobs is the mean amount of total actual values of the mean
SMSWG rate; n is the number of observation; rsim is the 3.2. Verifying the performance of ANN in predicting the mean SMSWG
standard deviation of simulated values of the mean SMSWG rate
rate; and robs is the standard deviation of actual values of the
mean SMSWG rate. In order to introduce the best ANN structure for predicting the
mean SMSWG rate, various structures of FFNN with three layers
Applying a few accuracy measures alongside each other (input, hidden and output layer) and different numbers of neurons
increases the accuracy of the study and how the problem is dealt in hidden layer, different training algorithms and two different
with, as each accuracy measure has its own advantages and limita- transfer functions were surveyed. Finally, based on five accuracy
tions and there is no single measure that is universally applicable indices R, MAPE, MAE, NRMSE and RMSE, optimum networks were
under all conditions (Zhang et al., 1998). For example, a brief selected. The best results of training and testing of ANN for each
description of the advantages and disadvantages of the aforemen- transfer and training function are presented in Table 4. According
tioned accuracy measures is provided in Table 2 (Makridakis & to Table 4 the optimum network has LM training algorithm and
Hibon, 1995; Tina et al., 2012; Zhao et al., 2011). tansig transfer function with 9 neurons in hidden layer (4–9–1).
As previously described, early stopping method was used to
3. Results and discussion overcome the overfitting problem in all of the above networks. In
this method, the available data were divided into three subsets,
3.1. Investigation of the correlation coefficients between explanatory training, validation and test set. Training set was used for updating
variables and dependent variable the network weights and biases during different iterations. The
second subset was the validation set. The error of the validation
In this study, four explanatory independent variables popula- set was monitored during the training process for specifying the
tion (Pop), solid waste collection frequency (Freq), maximum sea- process stopping point. Fig. 5 shows the variation of the Mean
sonal temperature (Tmax), and altitude (Alt) were considered; for Square Error (MSE) index (RMSE to the power of two) for the differ-
the first step, the correlation coefficients between the explanatory ent iterations of the optimum ANN model. As can be observed from
independent variables and the dependent variable were analyzed Fig. 5, the validation and training errors decrease to the sixth iter-
to evaluate the influence of each variable on the mean SMSWG ation during the initial phase of the training process. Afterward,
rate. The values of each correlation coefficient are presented in when the network begins to overfit the data, the error in the vali-
Table 1. These coefficients provide a measure of the linear relation dation set begins to rise, while the error in training set is still
between two variables. But checking the statistical significance of decreasing. When the validation error increases for six epochs,
these coefficients is more important than just computing them. the training is stopped, and the weights and biases at the minimum
For checking the statistical significance of the correlation coeffi- of the validation error are returned and reported.
cients, critical correlation coefficient Rcrit must be calculated. Rcrit In order to illustrate the relationship and degree of correlation
was calculated by Eq. (8): between two variables, scatter plot is a useful tool (Flott, 2012).
In scatter plot if data resemble a straight line the correlation of
t crit two variables will be high. Furthermore, assuming that the correla-
Rcrit ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð8Þ
df þ t 2crit tion is good, the considerable importance of the relationship is
indicated if the slope of the data becomes one. Fig. 6 shows the
where df = n  k is the degree of freedom; n is the number of data- scatter plot of observed versus predicted values of the optimum
sets and k is the number of explanatory variables. The tcrit value is ANN model for total data. As can be seen, the correlation between
the cutoff between retaining and rejecting the null hypothesis and the observed and predicted values of the mean SMSWG rate is rel-
must be found from existing tables. The correlation between two atively high. In this figure, a linear regression model has been fitted
variables is statistically valid if Rcrit is lower than the correlation on data. This linear regression between the observed and predicted
coefficient. In this study Rcrit, using a significance level of 0.05 values has been compared with identity line (i.e. 1:1 line), and the
(two-tailed test), is equal to 0.22 and since it is lower than the abso- difference between fitted line and 1:1 line at 1% level of signifi-
lute value of correlation coefficients in Table 3, it is obvious that cance has been verified. The results showed that the difference
there is a significant linear relation between each explanatory inde- between the regression line slope and intercept with 1.0 and 0.0,
pendent variable and mean SMSWG rate. respectively in this level of confidence was not statistically signif-
As observed in Table 3 the correlation coefficient among each of icant. Thus, based on the results, the high capability of ANN to pre-
the parameters of pop, maximum seasonal temperature, solid dict the mean SMSWG rate for a large number of urban regions is
waste collection frequency and the mean SMSWG rate is positive. confirmed.
This means that if any of the parameters increases in its amount, Fig. 7 illustrates the comparison between the observed and pre-
the mean SMSWG rate increases as well. Furthermore, the negative dicted values of the mean SMSWG rate in the test set. As can be
correlation coefficient between the altitude and the mean SMSWG observed, there is a meaningful agreement between these values.
rate shows the inverse relation between them. This means that the Therefore, if there are some data on solid waste generation in some
higher the altitude, the colder the weather becomes and it is more areas, it is not necessary to spend time and money on gathering
likely that the dietary habits of the people change in a way that time series data and predicting solid waste generation by consider-
leads to a lesser generation of solid waste. According to the ing the generation trend in each area.
 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23 21

Table 4
The specifications and performance measures of the best ANN of training and test stages in each case.

Activation function Training function Structure Train set Test set

R MAPE (%) MAE NRMSE RMSE R MAPE (%) MAE NRMSE RMSE
Tansig LM 4–9–1 0.92 6 37.67 0.08 53.87 0.86 8 48.37 0.10 68.32
BR 4–14–1 0.76 12 58.15 0.13 84.27 0.65 15 71.92 0.16 124.58
GDM 4–13–1 0.80 12 57.71 0.12 83.68 0.63 17 78.09 0.17 126.99
RP 4–10–1 0.86 9 43.96 0.10 75.56 0.82 9 40.07 0.11 78.30
SCG 4–8–1 0.84 10 45.49 0.11 78.67 0.73 10 54.14 0.12 85.78
Logsig LM 4–10–1 0.86 8 43.16 0.10 75.76 0.81 9 39.04 0.11 76.77
BR 4–14–1 0.88 9 41.19 0.10 71.44 0.74 10 47.36 0.13 83.44
GDM 4–11–1 0.79 12 58.29 0.13 84.99 0.67 16 77.30 0.17 123.03
RP 4–9–1 0.85 9 45.21 0.11 76.23 0.77 15 66.88 0.15 97.63
SCG 4–11–1 0.78 12 59.17 0.13 83.62 0.71 11 49.44 0.12 89.54

Note: bold values are corresponding to the performance measures of the optimum FFNN for prediction of the mean SMSWG rate in Fars Province.

10 also shows an Asymp. Sig. (2-tailed) value of 0.802. This parameter

Mean Squared Error (MSE)

train is also known as p-value. If the p-value is less than 0.05 the nor-
1 test mality hypothesis has to be rejected, and if the p-value is greater
validation than 0.05, the normality hypothesis is acceptable. Since the p-value
is 0.802 and greater than 0.05, there is no reason to doubt the
0.1
distribution of the mean SMSWG rate is normal with a significance
level of 0.05. Thus, the first condition is satisfied.
0.01 The second condition is to remove multicollinearity of indepen-
dent variables. A measure of the linear relationship between two
0.001 different variables is the correlation coefficient. The correlation
0 1 2 3 4 5 6 7 8 9 10 11 12 matrix between the explanatory variables is shown in Table 6.
12 iterations
For the validation of these values, they have to be compared with
Fig. 5. The variation of MSE index for different epochs of the optimum ANN model the value known as critical correlation coefficient (Sousa et al.,
(4–9–1, Tansig, LM). 2007). Correlation coefficient is valid if the absolute value is higher
than the critical value. In this study, the critical value was calcu-
lated (i.e. 0.22) with a significance level of 0.05. In Table 6, statis-
As shown previously, using ANN and considering a number of tically valid correlation coefficients are shown in bold.
features that affect the generation of solid waste may help us to To remove multicollinearity, PCA method was utilized. Applying
have adequate precision in predicting the mean SMSWG rate in this method, the variables were transformed into an equal number
20 cities of Fars Province. of PCs, preserving all information of original data.
For conducting PCA, in order to maximize the loadings of
3.3. Verifying the performance of MLR in predicting the mean SMSWG explanatory independent variables on each PC, varimax rotation
rate was applied. Table 7 shows the results of the varimax rotation,
which demonstrates the contribution of each explanatory indepen-
As previously mentioned, two conditions have to be considered dent variable in the PCs calculations. The values in bold correspond
in applying MLR model. First, the normality of dependent variable to the explanatory variable that has the most influence on each
distribution should be checked. Kolmogorov–Smirnov test can be principal component.
used to determine whether the mean SMSWG rate, as dependent The first principal component (PC1) has an important influence
variable, is normally distributed. Therefore, this test was con- on maximum seasonal temperature (Tmax). PC2 is heavily loaded on
ducted and the results are presented in Table 5. According to this solid waste collection frequency (Freq). PC3 and PC4 have impor-
table, the number of datasets (N) is 80 and the mean and standard tant contributions of population (Pop) and Altitude (Alt), respectively.
deviation of these datasets are 498.41 and 88.84, respectively. It After generating four principal components without any multi-
collinearity, a cross validation PCA-MLR model was developed and
1200 t-test (significance level of 0.05) was used to calculate the statisti-
cally valid coefficients. Finally, Eq. (9) was used to predict the mean
1000 y = 0.720x + 203.2 SMSWG rate:
Predicted mean SMSWG rate

R = 0.86
SMSWG ¼ 45:18  PC1 þ 50:70  PC2 þ 76:27  PC3
800
 29:06  PC4 þ 675:87 ð9Þ
(gr/capita.day)

600
In order to estimate the MLR model’s accuracy in predicting the
mean SMSWG rate, statistical indices were used, and the results
400
are presented in Table 8.
In Fig. 8, with linear regression model, the relationship between
200
the predicted mean SMSWG rate by MLR model and the observed
value in each city has been plotted. This plotted line has been com-
0
pared with 1:1 line, and at 1% level of significance the difference
0 200 400 600 800 1000 1200
Observed mean SMSWG rate (gr/capita.day) between these two lines has been surveyed. Based on the results,
the difference between the obtained slope and intercept with 1.0
Fig. 6. Scatter plot of observed versus predicted mean SMSWG rate by ANN. and 0.0, respectively in this level of significant was statistically
22 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23

1200
Observed

SMSWG rate (gr/capita.day)

1000 Predicted
800

600

400

200

0
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79
Test set

Fig. 7. Predicted and observed mean SMSWG rate in test set.

Table 5 1200
The results of Kolmogorov–Smirnov test.

Predicted mean SMSWG rate

1000
y = 0.535x + 314.0
Kolmogorov–Smirnov test Solid waste R = 0.70

(gar/capita.day)
No. of datasets 80 800
Normal parameters Mean 498.41
Std. deviation 88.84
600
Asymp. Sig. (2-tailed) 0.802

400

200
Table 6
Correlation matrix of the explanatory variables.
0
Pop Alt Tmax Freq 0 200 400 600 800 1000 1200

Pop 1 0.306 0.140 0.225 Observed mean SMSWG rate (gr/capita.day)

Alt 1 0.320 0.145
Tmax 1 0.011 Fig. 8. The scatter plot of observed versus predicted mean SMSWG rate by MLR.
Freq 1

Note: bold values are corresponding to statistically valid correlation coefficients.

Table 9
Statistical indices of MLR method and optimum ANN test stages.

Table 7 Statistical parameters MLR ANN

Principal components varimax rotating loadings.
R 0.70 0.86
PC1 PC2 PC3 PC4 MAPE 13% 8%
MAE 76.38 48.37
Pop 0.060 0.123 0.978 0.156
NRMSE 0.14 0.10
Alt 0.163 0.086 0.159 0.970
RMSE 95.13 68.32
Tmax 0.986 0.003 0.058 0.155
Freq 0.003 0.990 0.118 0.081
% cumulative variance 25.06 50.11 75.11 100

Note: Values in bold specify the variables that most influence each principal
network (4–9–1, Tansig, LM) and MLR model are shown in Table 9.
component. According to this Table, optimum ANN was able to forecast the
mean SMSWG rate of the 20 cities of Fars Province included in this
study with R and RMSE values equal to 0.86 and 68.32, respec-
Table 8 tively. By comparison, in the MLR model these statistical parame-
Statistical indices of MLR model in training and test stages. ters are 0.70 and 95.13. Moreover, forecasting error (MAPE) in
Model Performance measures Training set Test set
ANN is less than two-thirds of the forecasting error in MLR model.
Although each parameter indicates a more accurate performance
PCA-MLR R 0.78 0.70
of ANN compared to MLR model, the statistical significance of this
MAPE 11% 13%
MAE 63.24 76.38 improvement must be checked. Thus, an independent sample t-
NRMSE 0.12 0.14 test, with 5% level of significance was conducted on the error val-
RMSE 81.88 95.13 ues of the prediction for all cities by ANN and MLR. Based on the
obtained results, the accuracy improvement of the prediction by
ANN compared to MLR in this level of significance is statistically
significant. Therefore, it is clear that the performance of MLR significant. Therefore, the superiority of the performance of ANN
model to predict the mean SMSWG rate is not reasonable. Hence, over MLR model with respect to the prediction of the mean
the method suggested in this article may not be suitable to predict SMSWG rate is confirmed.
the mean SMSWG rate by regression technique.
4. Conclusions
3.4. Comparison between the performance of ANN and MLR
This study showed that, in the lack of historical records of SWG,
For better comparison between the ANN and MLR performance, mean SMSWG rate can be predicted by considering some effective
the results of the mean SMSWG rate prediction by optimum neural and measurable parameters. ANN and MLR were used to predict
 S. Azadi, A. Karimi-Jashni / Waste Management 48 (2016) 14–23 23

the mean SMSWG rate using population, solid waste collection fre- Jalili Ghazi Zade, M., Noori, R., 2008. Prediction of municipal solid waste generation
by use of artificial neural network: a case study of Mashhad. Int. J. Environ. Res.
quency, maximum seasonal temperature and altitude predictors.
2, 13–22.
These predictors were selected based on the correlation test. The Keser, S., Duzgun, S., Aksoy, A., 2012. Application of spatial and non-spatial data
results showed that the use of ANN model to predict mean SMSWG analysis in determination of the factors that impact municipal solid waste
rate led to accurate results and the difference between the pre- generation rates in Turkey. Waste Manage. 32, 359–371.
Kocadağlı, O., 2015. A novel hybrid learning algorithm for full Bayesian approach of
dicted and observed values was not statistically significant. But artificial neural networks. Appl. Soft Comput. 35, 52–65.
MLR model did not have reasonable performance, and the differ- Kollikkathara, N., Feng, H., Yu, D., 2010. A system dynamic modeling approach for
ence between the predicted and observed values was statistically evaluating municipal solid waste generation, landfill capacity and related cost
management issues. J. Waste Manage. 30, 2194–2203.
significant. Thus, the ANN model clearly demonstrated a higher Kumar, J.S., Subbaiah, K.V., Rao, P.V.V.P., 2011. Prediction of municipal solid waste
accuracy compared to MLR model on account of non-linearities. with RBF net work – a case study of Eluru, A.P., India. Int. J. Innovat. Manage.
Based on the results, it is better to use neural network with Technol. 2, 238–243.
Lebersorger, S., Beigl, P., 2011. Municipal solid waste generation in municipalities:
structure of 4–9–1 and LM training algorithm and tansig activation quantifying impacts of household structure, commercial waste and domestic
function for prediction of the mean SMSWG rate in Fars Province. fuel. Waste Manage. 31, 1907–1915.
Thus, if there is little information available on SWG in some Lohani, B.N., Hartono, D.M., 1985. Estimation of solid waste generation rates in the
city of Bandung, Indonesia. J. Waste Manage. Res. 3, 103–117.
areas, it is not necessary to spend a great deal of time and money Makridakis, S., Hibon, M., 1995. Evaluating accuracy (or error) measures. Printed at
on gathering time series data and predicting SWG by considering INSEAD, Fontainebleau, France.
the generation trend in each area. Using the suggested method in Mateu, J., Ricci, I., Villalonga, E., Cabeza, E., 2013. The impact of tourism on
municipal solid waste generation: the case of Menorca Island (Spain). Waste
this study presents a cost effective strategy for SWG prediction.
Manage. 33, 2589–2593.
Mentaschi, L., Besio, G., Cassola, F., Mazzino, A., 2013. Problems in RMSE-based
References wave model validations. Ocean Model. 72, 53–58.
Noori, R., Karbassi, A., Sabahi, M.S., 2010. Evaluation of PCA and Gamma test
Abdi, H., Williams, L.J., 2010. Principal component analysis. Wiley Interdiscip. Rev.: techniques on ANN operation for weekly solid waste prediction. J. Environ.
Comput. Stat. 2, 433–459. Manage. 91, 767–771.
Abdoli, M.A., FalahNezhad, M., SalehiSede, R., Behboudian, S., 2011. Longterm Nafarzadegan, A.R., RezaeianZadeh, M., Kherad, M., Ahani, H., Gharehkhani, A.,
forecasting of solid waste generation by the artificial neural network. J. Environ. Karampoor, M.A., Kousari, M.R., 2010. Drought area monitoring during the past
Prog. Sust. Energy, 1–9. three decades in Fars Province, Iran. J. Quat. Int.
ASCE Task Committee, 2000. Artificial neural network in hydrology. II: Hydrologic Patel, V., Meka, S., 2013. Forecasting of municipal solid waste generation for
applications. J. Hydrol. Eng. 5, 124–137. medium scale town located in the state of Gujarat, India. Int. J. Innovat. Res. Sci.
Azadi, S., Sepaskhah, A.R., 2011. Annual precipitation forecast for west, southwest, Eng. Technol. 2, 4707–4716.
and south provinces of Iran using artificial neural networks. Theor. Appl. Pires, J.C.M., Martins, F.G., Sousa, S.I.V., Alvim-Ferraz, M.C.M., Pereira, M.C., 2008.
Climatol. Selection and validation of parameters in multiple linear and principal
Batinic, B., Vukmirovic, S., Stanisavljevic, N., Ubavin, D., Vukmirovic, G., 2011. Using component regressions. Environ. Model. Softw. 23, 50–55.
ANN model to determine future waste management targets-case study of Rimaityte, I., Ruzgas, T., Denafas, G., Racys, V., Martuzevicius, D., 2011. Application
Serbia. J. Sci. Ind. Res. 70, 513–518. and evaluation of forecasting methods for municipal solid waste generation in
Beale, M.H., Hagan, M.T., Demuth, M.H., 2010. Neural Network ToolboxTM 7 User´s an eastern-European city. Waste Manage. Res. 30, 89–98.
Guide. MathWorks Inc, Natick, MA. Roy, S., Rafizul, I.M., Didarul, M., Asma, U.H., Shohel, M.R., Hasibul, M.H., 2013.
Boer, E.D., Je˛drczak, A., Kowalski, Z., Kulczycka, J., Szpadt, R., 2010. A review of Prediction of municipal solid waste generation of Khulna city using artificial
municipal solid waste composition and quantities in Poland. Waste Manage. 30, neural network: a case study. Int. J. Eng. Res.-Online 1, 13–18.
369–377. Shahabi, H., Khezri, S., Ahmad, B.B., Zabihi, H., 2012. Application of artificial neural
Boldrin, A., Christensen, T.H., 2010. Seasonal generation and composition of garden network in prediction of municipal solid waste generation (case study: Saqqez
waste in Aarhus (Denmark). Waste Manage. 30, 551–557. City in Kurdistan Province). World Appl. Sci. J. 20, 336–343.
Chen, C.C., 2010. Spatial inequality in municipal solid waste disposal across regions Shu, Y., Lam, N.S.N., 2011. Spatial disaggregation of carbon dioxide emissions from
in developing countries. Int. J. Environ. Sci. Technol. 7, 447–456. road traffic based on multiple linear regression model. Atmos. Environ. 45, 634–
Cherian, J., Jacob, J., 2012. Management models of municipal solid waste: a review 640.
focusing on socio economic factors. Int. J. Econ. Financ. 4, 131–139. Sousa, S.I.V., Martins, F.G., Alvim-Ferraz, M.C.M., Pereira, M.C., 2007. Multiple linear
Dyson, B., Chang, N.B., 2005. Forecasting municipal solid waste generation in a fast- regression and artificial neural networks based on principal components to
growing urban region with system dynamics modeling. Waste Manage. 25, predict ozone concentrations. Environ. Model. Softw. 22, 97–103.
669–679. Tchobanoglous, G., Theisen, H., Vigil, S., 1993. Integrated Solid Waste Management.
Flott, L.W., 2012. Using the scatter diagram tool to compare data, show relationship. McGraw-Hill, Inc.
Met. Finish. 110, 33–35. Thanh, N.P., Matsui, Y., Fujiwara, T., 2010. Household solid waste generation and
Fu, H.Z., Li, Z.S., Wang, R.H., 2015. Estimating municipal solid waste generation by characteristic in a Mekong Delta city, Vietnam. J. Environ. Manage. 91, 2307–
different activities and various resident groups in five provinces of China. Waste 2321.
Manage. 41, 3–11. Tina, G.M., De Fiore, S., Ventura, C., 2012. Analysis of forecast errors for irradiance on
Fukuoka, Y., Matsuki, H., Minamitani, H., Ishida, A., 1998. A modified back- the horizontal plane. Energy Convers. Manage. 64, 533–540.
propagation method to avoid false local minima. Neural Netw. 11, 1059–1072. Wang, Y., Li, J., Gu, J., Zhou, Z., Wang, Z., 2015. Artificial neural networks for
Gomez, G., Meneses, M., Ballinas, L., Castells, F., 2009. Seasonal characterization of infectious diarrhea prediction using meteorological factors in Shanghai (China).
municipal solid waste (MSW) in the city of Chihuahua, Mexico. J. Waste Appl. Soft Comput. 35, 280–290.
Manage. 29, 2018–2024. Xu, L., Gao, P., Cui, S., Liu, C., 2013. A hybrid procedure for MSW generation
Intharathirat, R., Abdul Salam, P., Kumar, S., Untong, A., 2015. Forecasting of forecasting at multiple time scales in Xiamen City, China. Waste Manage. 33,
municipal solid waste quantity in a developing country using multivariate grey 1324–1331.
models. Waste Manage. 39, 3–14. Zhang, G., Patuwo, B.E., Hu, M.Y., 1998. Forecasting with artificial neural networks:
Jahandideh, S., Jahandideh, S., Asadabadi, E., Askarian, M., Movahedi, M.M., Hosseini, the state of the art. Int. J. Forecast. 14, 35–62.
S., Jahandideh, M., 2009. The use of artificial neural networks and multiple Zhao, X., Wang, sh., Li, T., 2011. Review of evaluation criteria and main methods of
linear regression to predict rate of medical waste generation. J. Waste Manage. wind power forecasting. Energy Proc. 12, 761–769.
29, 2874–2879.