Prediction of Urban Ozone Concentration with Artificial Neural Networks
Negar Banan, Mohd Talib Latif and LiewJuneng
School of Environmental and Natural Resource Sciences, Faculty of Science and Technology,
Universiti Kebangsaan Malaysia, 43600 Bangi, Selangor, Malaysia
Abstract. Artificial neural network model was used to predict the daily maximum of ozone concentration
and we have applied it to the case of four monitoring stations in the urban namely (Gombak, Shah Alam,
Putrajaya and Petaling) and predict values of meteorological variables. Artificial neural network was
trained using samples of daily maximum data provided by the Malaysian Department of the Environment
(DOE) over a period of nine year (2003-2011). Based on the results, it can be deduced that the relationship
between the parameters and the ozone concentrations are highly complex and non-linear. Evaluation of
the regression based model results between the four stations were analysied using the ANN. Based on the
sample results it was confirmed that Putrajaya has the highest regression result of R= 0.64 in comparison
with other stations. This study shows that one hidden layer more adroit in predicting the daily maximum
ozone concentration.
Keywords: Prediction, Artificial neural network, time series, urban ozone
INTRODUCTION
Ozone (O3) is considered to be secondary pollutant, photochemical oxidant and the main component of
smog. It is also regarded as a crucial air pollutant in the atmosphere because O3 is capable of causing
damage to human health via respiratory disease (Ho et al., 2007; Karakatsani et al., 2010; Neidell and
Kinney, 2010; Mills et al., 2011; Huang et al., 2012). In addition, exposure to O 3 can lead to a decrease
in lung function (Highfill and Costa, 1995). High concentrations of surface ozone also affect vegetation
and forests due to the phytotoxic nature of O3. Ozone concentrations greater than 40 ppbv may be harmful
to the crop yield, biomass production, vitality and stress tolerance of forest trees (Fuhrer et al., 1997).
Excessively, high levels of O3 may be an obstacle to a forests’ capacity to seize carbon should there be
an excess of carbon dioxide in the future (Karnosky et al., 2003; Banan et al., 2013).
Modeling of O3 concentrations is not a trivial procedure due to the intrinsic relationship between
pollutants and meteorological variables (Sousa et al., 2007). Widespread review of the statistical
modeling of surface ozone based mostly on the basis of meteorological data has been presented by
Thompson et al. (2001). Wang et al. (2003) studied statistical characteristics of ozone to enable selection
of appropriate predictors of daily maximum ozone levels and emphasized the inclusion of factors that
affect both photochemical production and atmospheric accumulation of ozone when predicting ozone
levels. ANN models are more suitable amid the data-based models without needing explicit mathematical
representations. ANN performs better in prediction cases rather than statistical regression analysis
(Tasadduq et al., 2005; Yigit and Ertunc, 2006). This paper artificial neural network is used to predict
daily maximum ozone concentration in urban areas using the precursors and meteorological variables.
MATERIALS AND METHODS
1. Data Sets
This experiment is performed on four datasets collected from the air quality monitoring sites by the
Department of the Environment (DOE) in Malaysia and managed by a private company, Alam Sekitar
Sdn Bhd (ASMA). Daily maximum concentration of air pollutants (O3, NO, NO2, NOx, CO, PM10 and
NMHC) and daily maximum meteorological variables (Ambient temperature (AT), Humidity (H), Wind
speed (WS) and wind direction (WD)) were used in this research. The logarithm of the daily maximum
ozone concentrations of four stations, namely Gombak, Shah Alam, Putrajaya and Petaling Jaya stations,
which are located in Kalang Valley in the middle of the Malaysian Peninsula, were extracted and used
for modeling. The concentration data were obtained during the nine-year period from (2003-2011). All
2
the data were used to evaluate the predicting performance of the modeling data. The characteristics of the
datasets used are summarized in Table 1.
TABLE1. Characteristics of the datasets
2.
Datasets
No. of
attributes
Training
set 70%
Testing
set 20%
Validation
set 10%
Gombak
10
1019
291
146
Shah Alam
10
1732
495
247
Putrajaya
10
1468
419
210
Petaling Jaya
10
1988
568
284
ANN Model for Air Quality Prediction
ANN is modeled as the neurons connected or functionally-related to each other; imitate the behavior of
human biological neurons. Neural networks have found extensive application in diverse areas such as;
pattern recognition and classification, time series prediction and modeling (Behrang et al., 2010).
Neurons which are the basic components of the neural network are interconnected through different layers
such as input, hidden and output layers. The degree of interconnection is expressed by the weight which
means the impact of a neuron on a neuron. The representation of neural network is shown in Figure 1.
NO
NO2
NOx
CO
PM10
O3
NMHC
Output layer
AT
H
WS
Hidden layer
WD
Input layer
FIGURE 1. Schematic of a neural network
The topology of ANN plays a critical role in its performance evaluation. A multilayer feed forward
network model with one hidden layer has the inherent capability to forecast any complex nonlinear
function. In as much as there are adequate numbers of hidden layer neurons in the model. Hence, in this
literature, multilayer feed forward networks models consisting of one hidden layer were used. The choices
of optimum number of hidden layers are very crucial to ensure optimal performance accuracy of ANN.
In the meantime, there has not been any acceptable theory to determine the adequate number of hidden
layer neuron need to be used for a particular problem. A rule of thumb to determine the optimal numbers
of neurons is to start with a few numbers of neurons and increase in a stepwise direction. In the course of
the experiment, the performance of the each hidden neuron was monitored against a chosen performance
criteria. Lastly, in order to reduce the training difficulty and balance the importance of each parameter
during the training process, the data inputs were normalized. It is recommended that the data be
3
normalized between slightly offset values such as 0.1 and 0.9. One way to scale the input and output
variables in the interval [0.1, 0.9] as Pn= 0.1+ (0.9-0.8) (P − Pmin) / (Pmax − Pmin) Xu et al. ( 2007).
The data set is randomized into three distinct sets called training, testing and validation sets. The training
set constitutes the largest data inputs of about 70% and is used by the neural network to acquire the
patterns present in the data. Next is the testing set making up of about 20% of the data set and is used to
evaluate the generalization ability of supposedly trained network. While the remaining 10% of the data
is for checking on the performance of the trained network.
The model of employed ANN is shown in Figure 1. It consists of three layers of neurons
including input, hidden, output layers. The input layer is the first layer of neurons. Each input neuron
represents a separate attribute in the train/test datasets station (for example from NO to WD in Figure 1).
The number of the inputs is equal to the number of attributes in the dataset. The number of nodes for
other hidden layers is equal to the half of the number of nodes in the previous layer.
Experimental Results
Artificial neural network have been applied in this study. The input and target values were
normalized into the range of [0.1and 0.9] in the pre-processing phase. The weights and biases were
adjusted based on gradient-descent method in the training phase. The correlation coefficient (R) was
chosen as the statistical criteria for measuring the network performance. The results are value was shown
in Table 2.
TABLE2. Correlation coefficient
Datasets
Correlation coefficient (R)
Gombak
0.62945
Shah Alam
0.63819
Putrajaya
0.64199
Petaling Jaya
0.57657
Figure 2 represents the network performance versus the number of epochs. The network performance
starts by a large value at the first epochs and due to training, the weights are adjusted to minimize this
function which makes it decreasing. Moreover, a black dashed line is plotted representing the best
validation performance of the network. The training stops when the green line which represents the
validation training set (network performance) intersects with the black line.
Gombak
Shah Alam
4
Putrajaya
Petaling Jaya
FIGURE 2.Performance function of the network during training
Network structure is, 10-5-1
Regression analysis was performed to investigate the correlation between the actual and
predicted results based on the value of correlation coefficient (R). The perfect fit between the training
data and the produced results was indicated by the value of R which is equal to 1. Figure 3 shows the
regression analysis plots of the network structure. In a regression plot, the perfect fit which shows the
perfect correlation between the predicted and targets is indicated by the solid line. The dashed line
indicates the best fit produced by the algorithm.
Gombak
Putrajaya
Shah Alam
Petaling Jaya
FIGURE 3.Regression plot analysis
Network structure, 10-5-1
From Table 2, it can be seen that model with network structure 10-5-1 is the best models for the
air quality prediction as it yields the highest values of correlation coefficient (R). As shown in Figure3, a
good agreement between predicted and measured values based on the value of correlation coefficient (R).
5
In order to get a sense of the solutions distribution, we examined the minimum, maximum,
median, lower quartile and upper quartile using a box and whisker plots as presented in Figure 4,a ( one
hidden layer for Putrajaya , Petaling Jaya, Shah Alam, Gombak stations) and Figure 4,b (two hidden layer
for Putrajaya , Petaling Jaya, Shah Alam, Gombak stations). From Figure4,a based on one hidden layer
on ANN algorithm, it can be seen that most of the observations are concentrated in the low end of the
scale. This shows that the solution’s distribution is skewed to the lower end. However, there is a case
where the solution’s distribution is symmetric, where the observation is evenly split at the median (i.e.,
Shah Alam at left side). However, based on the two hidden layer on ANN algorithm (Figure4,b), there is
a case that the observation is concentrated on the upper end of the scale which shows that the solution’s
distribution is skewed to the upper end (i.e., Gombak at the left side and Putrajaya at the right side) and
there is a case that the observation is concentrated in the lower end of the scale which shows that the
solution’s distribution is skewed to the lower end (i.e., Putrajaya at the left side and
Petaling Jaya at
the right side). Based on the patterns presented for the distribution of the result, generally we can say that
the one hidden layer on ANN algorithm is better than two hidden layer on ANN algorithm when tested
on the Putrajaya, Petaling Jaya, Shah Alam, Gombak stations, but of course a statistical test should be
carried out to support this claim.
4
0.675
0.625
0.650
0.600
5
Correlation coefficient
Correlation coefficient
0.625
3
0.600
0.575
1
0.550
0.575
7
7
10
0.550
0.525
8
Gombak
Shah Alam
Putrajaya
petaling Jaya
Stations
Gombak
Shah Alam
Putrajaya
Petaling Jaya
Stations
1 hidden layer (a)
2 hidden layers (b)
FIGURE 5(a) and 5(b). Box-and-whisker plots on the four datasets
Conclusion
Artificial neural network employed to predict ozone concentrations as a perform of
meteorological conditions and precursor concentrations. The network was trained utilizing daily
maximum data that provided from the Malaysian department of the Environment throughout a nine year
(2003-2011) into the cities of Gombak, Shah Alam, Putrajaya and Petaling Jaya in Malaysia. In this
research, urban ozone concentrations are estimated utilizing surface meteorological variable as predictors
by an artificial neural network for four areas in Malaysia. The experimental results using four stations
have exposed that the proposed algorithms effectively to predict ozone concentrations. Further studies
will be devoted to validate the artificial neural network for the function of predict ozone concentrations
as a function of meteorological conditions and precursor concentrations for monthly scales.
ACKNOWLEDGEMENT
The authors would like to thank Universiti Kebangsaan Malaysia for Research University Grant (DIP2012-020 and LRGS/TD/2011/UKM/PG01) and the Malaysian Department of the Environment (DOE)
for providing all the necessary investigative information and air quality data in the process of conducting
research.
REFFERENCES
6
Banan, N. Latif, M.T. Juneng, L. & Ahmad, F. 2013. Characteristics of surface ozone concentrations at
stations with different backgrounds in the Malaysian Peninsula. Aerosol Air Qual. Res.13: 10901106.
Behrang, M. A. Assareh, E. Ghanbarzadeh, A. & Noghrehabadi. A. R. 2010. The potential of different
artificial neural network (ANN) techniques in daily global solar radiation modeling based on
meteorological data. Solar Energy. 84: 1468-1480.
Fuhrer, J., Skarby, L. & Ashmore, M.R. 1997. Critical Levels for Ozone Effects on Vegetation in
Europe. Environ. Pollut. 97: 91-106.
Highfill, J.W. & Costa, D.L. 1995. Statistical response models for ozone Exposure: Their generality when
applied to human spirometric and animal permeability functions of the Lung. J. Air Waste
Manage. Assoc. 45: 95-102.
Ho, W.C., Hartley, W.R., Myers, L., Lin, M.H., Lin, Y.S., Lien, C.H. & Lin, R.S. 2007. Air Pollution,
Weather, and Associated Risk Factors Related to Asthma Prevalence and Attack Rate. Environ.
Res. 104: 402-409.
Huang, H.L., Lee, M.G. & Tai, J.H. 2012.Controlling Indoor Bioaerosols Using a Hybrid System of
Ozone and Catalysts. Aerosol Air Qual. Res. 12: 73-82.
Karakatsani, A., Kapitsimadis, F., Pipikou, M., Chalbot, M.C., Kavouras, I.G., Orphanidou, S. D.,
Papiris, S. &Katsouyanni, K. 2010. Ambient Air Pollution and Respiratory Health in Mail
Carriers.Environ. Res. 110: 278-285.
Karnosky, D.F., Zak, D.R. & Pregitzer, K.S. 2003. Tropospheric O3 Moderates Responses of Temperate
Hardwood Forests to Elevated CO2: A Synthesis of Molecular to Ecosystem Results from the
Aspen FACE Project. Funct. Ecol. 17: 289-304.
Mills, G., Hayes, F., Simpson, D., Emberson, L., Norris, D. Harmens, H. &Buker, P. 2011. Evidence of
Widespread Effects of Ozone on Crops and (Semi-) Natural Vegetation in Europe (1990-2006)
in Relation to AOT40-and Flux-Based Risk Maps.Global Change Biol. 17: 592-613.
Neidell, M. & Kinney, P.L. 2010. Estimates of the Association between Ozone and Asthma
Hospitalizations that Account for Behavioral Responses to Air Quality Information. Environ. Sci.
Policy 13: 97-103.
Sousa, S.I.V., Mrtins, F.G., Alvim-Ferraz, M.C.M. & Pereira, M.C. 2007. Multiple linear regression and
artificial neural networks based on principle components to predict ozone concentrations.
Environmental Modeling & Software. 22: 97-103.
Tasadduq, I., Rehman, S. & K. Bubshait. 2005. Application of neural networks for the prediction of
hourly mean surface temperatures in Saudi Arabia. Renewable Energy. 25: 545-554.
Thompson, M.L., Reynolds, J., Cox, L.H, Guttorp. P. & Sampson.P.D. 2001. A review of statistical
methods for the meteorological adjustment of tropospheric ozone. Atmos. Environ. 35: 617-630.
Wang, W., Lu, W., Wang, X. & Leung, A.Y.T. 2003. Prediction of maximum daily ozone level using
combined neural network and statistical characteristics. Environmental International. 29: 555562.
Xu, L., Jiandong, X., Shizhong, W., Yongzhen, Z. &Rui, L. 2007. Optimization of heat treatment
technique of high-vanadium high-speed steel based on back-propagation neural networks.
Materials & Design. 28: 1425-1432.
Yigit, K. S. & Ertunc, H. M. 2006. Prediction of the air temperature and humidity at the outlet of a cooling
coil using neural networks. International Communications in Heat and Mass Transfer. 33: 898907.
7
8
9