Multivariate methods for ground-level ozone modeling

Atmospheric Research xxx (2011) xxx–xxx

ATMOS-02455; No of Pages 9

Contents lists available at ScienceDirect

Atmospheric Research

j ourna l homepage: www.e lsev ie r.com/ locate /atmos

Multivariate methods for ground-level ozone modeling

Bilge Özbay a,⁎, Gülşen Aydın Keskin b, Şenay Çetin Doğruparmak a, Savaş Ayberk a

a Department of Environmental Engineering, Kocaeli University, 41380 Kocaeli, Turkeyb Department of Industrial Engineering, Kocaeli University, 41380 Kocaeli, Turkey

a r t i c l e i n f o

⁎ Corresponding author. Tel.: +90 262 3033199; faE-mail address: [email protected] (B. Özbay)

0169-8095/$ – see front matter © 2011 Elsevier B.V.doi:10.1016/j.atmosres.2011.06.005

Please cite this article as: Özbay, B., etdoi:10.1016/j.atmosres.2011.06.005

a b s t r a c t

Article history:Received 21 July 2010Received in revised form 3 June 2011Accepted 3 June 2011Available online xxxx

The aim of this study is to apply multivariate statistical methods in predicting ozone (O3)concentrations at the ground level of the troposphere as the function of pollution andmeteorological parameters. PM10, SO2, NO, NO2, CO, O3, CH4, NMHC, temperature, rainfall,humidity, pressure, wind direction, wind speed and solar radiation were measured hourly forone year period in order to predict O3 concentrations of 1 h later. In the study, relationshipsbetween O3 data and other variables were investigated by bivariate correlation analysis. CH4,NMHC, NO2 exhibited considerable negative correlations with O3 described with the Pearsoncorrelation coefficients of −0.67, −0.55, −0.51, respectively whereas highest positivecorrelation was noted for temperature with correlation coefficient of 0.60. Multiple regressionanalysis (MLR) was used for modeling annual and seasonal O3 concentrations. Adjusted R2

values were determined as 0.90, 0.85 and 0.92 respectively for annual period, cooling andwarming seasons. In order to decrease the number of input variables principle componentanalysis (PCA) was applied by using annual data. MLR analysis was repeated using fourprinciple components and new adjusted R2 was calculated as 0.63.

© 2011 Elsevier B.V. All rights reserved.

Keywords:OzoneStatistical analysisMultiple regression analysisPrinciple component analysisKocaeli

1. Introduction

Tropospheric ozone (O3), the major component associatedwith photochemical smog, is produced when the primarypollutants, nitrogen oxides (NOX) and volatile organic com-pounds (VOCs) interact under the action of sunlight (Abdul-Wahab, 2001; Brulfert et al., 2007). NOX, VOCs (especially non-methane hydrocarbons, NMHC) and carbonmonoxide (CO) areamong the most important O3 precursors (Jun et al., 2007). O3

concentrations monitored at the tropospheric site are alsostrongly influenced by meteorological conditions like temper-ature, humidity, wind direction, wind speed, cloud cover,pressure, solar radiation, rainfall etc. (Kovač-Andrić et al., 2009).

The seasonal and diurnal variations of surface O3, itsprecursors and the related meteorology have been extensivelystudied around the world due to well-known harmful impactsof high O3 levels on biosphere, human health, animal popula-tions, agriculture productivity and forestry (Lippmann, 1989;

x: +90 262 3033003..

All rights reserved.

al., Multivariate metho

Schenone and Lorenzini, 1992; Brauer and Brook, 1997;Chatterton et al., 2000; Lehman et al., 2004; Paoletti, 2006;Pulikesi et al., 2006; Shan et al., 2008; Debaje and Kakadeb,2009). Several methods such as statistical regression, graphicalanalysis, fuzzy logic based methods have been used formodeling tropospheric O3 levels (Clark and Karl, 1982; Coxand Chu, 1991; Buhr et al., 1995; Blankinship, 1996; Lavecchiaet al., 1996; Peton et al., 2000). Among thesemethods,MultipleLinear Regression (MLR) has provided successful results in O3

modeling studies (Abdul-Wahab et al., 2005 and Sousa et al.,2007).

The complexity of O3 formation mechanisms in the tropo-sphere (Seinfeld and Pandis, 2006), the complexity of meteo-rological conditions in urban areas and the uncertainty in themeasurements of all the parameters involved, make the fastand accuratemodeling of O3 very difficult. In order to avoid thisproblem, usage of the principal component analysis (PCA), abasic method in the framework of multivariate analysistechniques, has been suggested (Sousa et al., 2007). Due to itssimplicity and efficiency in processing huge amount of processdata, it is recognized as a powerful tool of statistical process

ds for ground-level ozone modeling, Atmos. Res. (2011),

2 B. Özbay et al. / Atmospheric Research xxx (2011) xxx–xxx

monitoring. It has been widely used in numerous areas in-cludingdata compression, feature extraction, image processing,pattern recognition, signal analysis, and process monitoring(Ding et al., 2010). PCA is also successfully employed inenvironmental investigations for separating interrelationshipsbetween statistically independent basic components (Tianet al., 1989; Shi and Harrison, 1997; Vaidya et al., 2000;Abdul-Wahab et al., 2005).

In this study firstly, bivariate correlation analyses wereachieved by using allmeasured parameters in order to evaluateappropriateness of data for modeling studies. Consideringcomplex photochemical reactions resulting O3, using bothprecursor concentrationsandmeteorological variables improvemodeling efficiencies (Argiriou, 2007). For this reason, thesegroups have been evaluated together in statistical analyses.MLR analyses were performed for annual and seasonal periodsas O3 concentrations exhibit evident seasonal variations. Fur-thermore, PCA was applied by using annual data in order toevaluate the relative influence of precursor concentrations andmeteorological variables on O3 formation. Finally, MLR wasrepeated by using components obtained from PCA analysis anda regression equation was formed.

2. Materials and methods

2.1. Investigation area and data collection

In this study, the Dilovasi region was selected for datacollection. Dilovasi, located in north-west of Kocaeli, is a well-known industrial region with many factories working indifferent sectors. There are five industrial zones installed inthe region. Transport facilities have been improved with twomotorways, a railway and many seaports. Due to economicaldevelopment, population of the region has been increasing

Turkey

Station of Dilovası IndustriStation of Kocaeli EnvironmLocation of Dilovası on Tu

Fig. 1. Map of the s

Please cite this article as: Özbay, B., et al., Multivariate methodoi:10.1016/j.atmosres.2011.06.005

continuously and residential areas are enlarging. The popu-lation was over 50,000 in 2010 (TUIK, 2010). The studiedregion is shown in Fig. 1.

As a result of intensive industrialization natural ecologicalstructure has been destroyed markedly in Dilovasi. Geo-graphical structure of the region also enhances air pollution inDilovasi as it is located in a gully with 10 m altitude. Airquality of the region has been monitored by continuousmeasurements carried out in two different stations (Fig. 1).Station of Kocaeli Environmental and Forestry Departmentwas installed in 25.07.2006. It lies between 29.54 eastmeridian and 40.78 north parallel at 44 m height. Otherstation belonging to Dilovasi Industrial Zone Office wasconstructed in 15.08.2007. It is located on 29.52 east meridianand 40.77 north parallel at 23 m height. Particulate material(PM10), sulphur dioxide (SO2), nitrogen monoxide (NO),nitrogen dioxide (NO2), carbon monoxide (CO) and ozone(O3) parameters have been measured in Station of KocaeliEnvironmental and Forestry Department whereas PM10,methane (CH4), nonmethane hydrocarbon (NMHC), NO,NO2, SO2, temperature (T), humidity (H), pressure (P), winddirection (WD) and speed (WS) have been measured inStation of Dilovasi Industrial Zone Office. Additionally, rainfall(R) in the region has been monitored by Kocaeli Meteoro-logical Office.

2.2. Method of measurements

Data used in the study was collected in period betweenSeptember 2008 and August 2009. With this aim O3, NO, NO2,SO2, CO, PM10, CH4, NMHC, T, H, P, WD, WS, SD and R weredetermined with hourly measurements. Most of the pollutionparameters (PM10, SO2, NO, NO2, CO andO3)weremeasured inthemain station, Station of Kocaeli Environmental and Forestry

Marmara Sea

al Zone Officeental and Forestry Department

rkey map

tudied area.

ds for ground-level ozone modeling, Atmos. Res. (2011),

3B. Özbay et al. / Atmospheric Research xxx (2011) xxx–xxx

Department whereas CH4 and NMHC concentrations weremeasured in Station of Dilovasi Industrial Zone Office. For themeasurement of meteorological data, opportunities of KocaeliEnvironmental and Forestry Ministry, Kocaeli MeteorologicalOffice and Kocaeli University were used. Environmental andForestry Ministry has measured T, H, P, WD and WS by usingDelta OHM model device while SR has been measured withVentage PRO2 model analyzer belonging to Kocaeli University.Tipping-bucket pluviometer has been used by Kocaeli Meteo-rological Office in order to measure rainfall.

O3 was measured by using Thermo Environmental modelphotometric O3 analyzer. The U.V. photometer determined O3

concentrationbymeasuring theattenuationof light due toO3 inthe absorption cell, at a wavelength of 254 nm. The concen-tration of O3 was related to the magnitude of the attenuation.Precision of the measurements was ±1.0 ppb.

CO concentrations were monitored with Thermo Environ-mental CO analyzer. Device operated according to the gas filtercorrelation technology. CO absorbed infrared radiation at awavelength of 4.6 μm and instrument electronics transformedthe basic analyzer signal into a linear output. Linearity of themeasurement results was ±1.0%.

BAM-1020PM Monitoring System was used for PM10measurements. During measurements the amount of massdeposited on a filter tape was determined by measuring theamount of beta attenuationbefore andafter a sampling interval.During sampling, the flow rate was precisely controlled. Theaccuracy of the measurements was ±2.0 μg/m3 for 24 h.

Concentrations of CH4 and NMHC species were deter-mined by using The Synspec ALPHA M/TNMHC analyzer. Theanalyzer operated as a gas chromatograph. It contained acompact oven with a column that separated CH4 from totalNMHCs. The gas sample passes through the column with aspecial layered packing (Carbograph). The CH4 passesthrough and is first injected into the detector. After then,the column is backflushed and all other hydrocarbons pass tothe detector. Finally two peaks are generated by the FID: amethane and a TNMHC-peak. Linearity of the measurementswas determined as ±1.0%.

The Thermo Scientific Sulfur Dioxide Analyzer utilizedpulsed fluorescence technology to measure the amount ofSO2 in the air up to 100 ppm. The pulsing of the U.V. sourcelamp served to increase the optical intensity whereby agreater U.V. energy throughput and lower detectable SO2

concentration were realized. Precision of the measurementswas 1.0% of reading or ±1.0 ppb.

NO and NO2 measurements were carried out by usingThermo NO-NO2-NOx Analyzer, Model 42i. During measure-ments light produced by the gas-phase titration of nitricoxide and O3 was used for NO/NOx gas analysis. Precision ofthe measurement results was ±1.0 ppb in full scale.

All instruments were serviced and calibrated once a yearin the responsibility of Kocaeli Environmental Ministry.Furthermore quality assessment included zero- and span-checking were also made monthly by collaboration of KocaeliEnvironmental Ministry and relevant commercial firms.

3. Theory and calculations

The following section is consisted of two differentstatistical techniques. Firstly, Multiple Linear Regression

Please cite this article as: Özbay, B., et al., Multivariate methodoi:10.1016/j.atmosres.2011.06.005

(MLR) was defined that used for modeling O3 concentrations.The model evaluates O3 concentrations effected by itsprecursors and meteorological factors. Secondly, PrincipalComponent Analysis (PCA) was explained which was used toreduce the number of predictive variables and transformthem into new variables. SPSS 17 statistical programwas usedfor MLR and PCA applications whereas measurement resultswere normalized by MATLAB 6.5.

3.1. Multiple linear regression analysis (MLR)

Regression-based methodologies are commonly used inO3 prediction studies. MLR is one of the most widely usedmethods for modeling O3 concentrations (dependent orresponse variable) in dependence of meteorological param-eters and different atmospheric pollutants (independent orpredictor variables). It can be expressed according to thefollowing equation (Kovač-Andrić et al., 2009):

y = b0+b1x1+b2x2+…+ bkxk+ε ð1Þ

where, bi are the regression coefficients, xi are the explana-tory variables and ε is stochastic error associated with theregression.

All data were standardized before application of MLRprocedure. Normalized data were calculated according tofollowing equation (Keskin et al., 2010):

NIi;j =I i; jð Þ−min jð Þmax jð Þ−min jð Þ ð2Þ

where, I is the input value, NI is the standardized value, i is thenumberofmeasurements, j is themeasuredvalueof thevariable.

3.2. Principal component analysis (PCA)

Principle component analysis was first proposed byPearson (1901) to reduce dimensionality of a data setconsisting of a large number of interrelated variables, whileretaining as much as possible of the variation present in thedata set. This is achieved by transforming to a new set ofvariables, the principal components (PCs), which are uncor-related and ordered (Lam et al., 2010). By this way, the firstseveral components explain most of the variation present inall of the original variables (Lu et al., 2006). Those few PCswillbe the index to explain the summarization of parameters(Hsieha and Yang, 2008).

PCA uses the eigenvalues of the covariance matrix and itonly finds the independent axes of the data under the Gaussianassumption. Eigen values determined from PCA are special setof scalars associated with a linear system of equations (i.e., amatrix equation) that are sometimes also known as character-istic values. The eigenvalues of the standardized matrix arecalculated according to the following equation:

jC−λI j = 0 ð3Þ

where C is the correlation matrix of the standardized data, λ isthe eigenvalues and I is the identity matrix. The weights of thevariables are determined by using Eq. (4):

jC−λI jW = 0 ð4Þ

ds for ground-level ozone modeling, Atmos. Res. (2011),

Table 1Monthly average values for the measured (a) pollution (b) meteorological parameters.

Months O3 (μg/m) 3) SO2 (μg/m3) NO (μg/m3) NO2 (μg/m3) CO (μg/m3) PM10 (μg/m3) CH4 (μg/m3) NMHC (μg/m3)

Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev.

(a)September 28.346 8.612 34.076 5.983 5.973 3.992 24.197 4.892 1939.843 36.321 62.080 14.280 1540.928 76.123 392.335 72.398October 24.142 7.226 19.668 3.330 7.728 5.681 24.803 6.440 1907.703 76.738 56.970 10.445 1690.711 76.660 444.347 49.427November 22.387 3.889 10.313 5.770 8.025 6.436 23.539 7.751 2059.918 101.835 66.859 7.368 1926.094 61.256 993.621 82.468December 12.125 1.899 11.160 6.647 19.660 5.649 41.997 3.945 3086.304 357.701 64.611 26.019 1985.001 55.409 699.492 49.434January 12.159 2.824 16.932 3.010 18.873 3.992 32.499 5.052 3260.455 323.088 82.377 23.032 2075.740 56.236 801.830 69.660February 5.600 0.884 25.140 7.719 15.850 4.784 31.189 6.633 3111.333 292.730 67.020 18.511 1980.527 75.364 733.001 39.429March 5.072 0.898 26.820 7.945 17.487 3.560 35.104 5.905 3055.525 216.407 83.434 16.142 1980.303 68.708 756.264 106.075April 6.201 1.606 13.607 9.913 11.224 5.583 30.816 5.390 3100.737 80.798 61.854 8.580 1868.667 66.530 760.486 164.686May 8.840 3.271 9.729 9.087 8.542 7.555 21.861 8.518 2899.090 184.072 51.500 10.759 1852.549 70.454 1009.369 158.015June 13.240 4.019 11.316 11.092 9.379 4.872 23.935 4.107 3068.874 83.897 57.518 9.171 1852.992 19.964 550.871 35.684July 35.153 8.892 5.486 6.094 5.549 3.139 14.657 4.210 3103.180 111.773 57.970 10.382 1604.878 49.632 73.802 33.350August 43.108 10.050 5.664 8.618 4.957 2.791 12.335 5.149 3085.683 83.409 50.939 7.941 1347.746 93.416 112.176 19.361

Months T (°C) H (%) P (mb) R (mm) SR (kW/m2) WS (m/s) WD (degrees)

Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev. Avg. Std. dev.

(b)September 20.631 2.123 64.737 9.181 1008.225 0.962 0.010 0.013 141.459 177.521 2.214 0.890 164.416 53.678October 17.226 1.965 68.397 7.541 1014.999 1.085 0.000 0.000 108.678 152.563 1.560 0.452 181.630 34.865November 14.348 1.832 73.410 6.202 1018.473 0.418 0.000 0.000 70.668 106.587 2.159 0.655 186.757 36.110December 7.860 1.797 73.870 5.734 1019.136 1.089 0.072 0.071 40.192 61.659 1.631 0.259 245.406 50.249January 7.074 1.535 69.468 6.008 1012.755 1.006 0.107 0.042 40.515 61.837 1.578 0.189 224.502 33.370February 8.169 0.738 73.323 2.813 1005.058 0.588 0.076 0.057 40.146 59.509 1.738 0.447 217.644 24.924March 8.836 1.451 67.227 5.498 1008.529 0.638 0.052 0.019 100.043 132.234 1.401 0.143 224.844 31.600April 11.634 2.392 65.157 9.036 1010.672 0.599 0.022 0.035 151.009 182.547 2.006 0.721 202.440 45.432May 19.278 3.384 55.833 8.962 1011.820 1.896 0.000 0.000 269.340 327.108 2.035 0.900 144.507 108.307June 22.539 2.933 55.581 10.374 1006.612 0.642 0.025 0.042 274.127 296.208 1.584 0.600 198.399 63.470July 24.675 2.524 58.963 10.370 1006.043 0.600 0.022 0.024 252.857 273.996 2.169 0.962 167.995 47.491August 23.670 2.571 58.945 12.712 1007.960 0.595 0.000 0.000 227.998 264.623 2.683 1.337 109.003 52.218

4B.Ö

zbayet

al./Atm

osphericResearch

xxx(2011)

xxx–xxx

Pleasecite

thisarticle

as:Özbay,

B.,et

al.,Multivariate

methods

forground-level

ozonemodeling,

Atm

os.Res.

(2011),doi:10.1016/j.atm

osres.2011.06.005

5B. Özbay et al. / Atmospheric Research xxx (2011) xxx–xxx

whereW is thematrix of theweights. Varimax rotation is oftenused in investigations to see how groupings of items measurethe same concept (Zhang et al., 2008). Factor loadings obtainedfrom varimax rotation represent the contribution of eachvariable in a specific principal component. The PCs wereobtained through multiplication of the standardized datamatrix by the previously calculated weights (W).The applica-bility of the PCA to the data sets used in this study is providedfrom Bartlett's sphericity test. It is used to examine thehypothesis that the variables are uncorrelated in the popula-tion. It is expressed by the following equation:

x2k =

"n−k−2 p−kð Þ+7+2= p−kð Þ

6

+ ∑k

j=1

λλj−λ

!2#x − ln ∏

p

J=k+1λj+ p−kð Þ lnλ

" #ð5Þ

In the equation p is the number of components, λj

represents the eigenvalue for the kjth component, n is thenumber of observations. Following equation can be used inorder to calculate λ (Sousa et al., 2007):

λ = ∑p

J=k+1

λj

p−kð6Þ

4. Results and discussion

4.1. Evaluation of the measured data

Monthly average values of variables were calculated byusing hourly measured data in order to investigate generalpollution and meteorological characteristics of the region.Table 1(a) and (b) summarizes the obtained results.

Investigating the concerned literature, variations werenoted in air pollution levelsof citiesdue todifferentbackgroundpollution levels, specific emission conditions, general meteo-rological conditions and location of monitoring station (Mayer,1999). In the study that aimed to investigate ambient air qualityin Eskisehir, Turkey annual SO2, PM, NO2 and O3 averages were

Table 2Correlations between input variables.

O3 SO2 NO NO2 CO PM10 CH4 NM

O3 (μg/m3) 1 −0.165 −0.358 −0.511 −0.161 −0.190 −0.675 −SO2 (μg/m3) 1 0.281 0.379 −0.093 0.227 0.095NO (μg/m3) 1 0.716 0.239 0.389 0.352NO2 (μg/m3) 1 0.124 0.387 0.347CO (μg/m3) 1 0.175 0.203PM10(μg/m3)

1 0.277

CH4 (μg/m3) 1NMHC(μg/m3)

T (0C)H (%)P (mb)R (mm)SR (kW/m2)WS (m/s)WD(degrees)

Please cite this article as: Özbay, B., et al., Multivariate methodoi:10.1016/j.atmosres.2011.06.005

found as 47.13, 30.45, 22.57 and 46.08 μg/m3 (Ozden et al.,2008), whereaswe have determined average concentrations ofthese pollutants as 15.83, 63.59, 26.41 and 18.03 μg/m3,respectively in this paper. In another study that aimed topredict O3 levels for Khaldiya region of Kuwait, averageconcentrations of CH4, CO, NO, NO2, SO2 and O3 weredetermined as 1137.97, 2410.48, 84.232, 23.64, 4.69, 11.33 μg/m3, respectively in June (Abdul-Wahab and Al-Alawi, 2002). Asseen from Table 1(a) we have observed average O3 concentra-tion as13.24 μg/m3 in JuneandhighestO3 levelwas recorded inAugust with 43.10 μg/m3.

4.1.1. Bivariate correlation analysisAs modeling studies aimed to investigate prediction of

hourly O3 levels, bivariate correlation analyses were alsoperformed by using hourly measured data of one year period.In Table 2, cross correlations among all variables werepresented. As seen from the table, O3 concentrations werenegatively correlated with SO2, NO, NO2, CO, PM10, CH4,

NMHC, H, R and WD. Among the negatively correlatedparameters CH4, NMHC, NO2 were the most efficient pollut-ants. This result is congruent with the literature as thesepollution parameters have been known as O3 precursors(Abdul-Wahab and Al-Alawi, 2002; San Jose et al., 2005; Duanet al., 2008). Correspondingly Abdul-Wahab et al. (2005),have determined negative correlations between O3 and CH4,NMHC, CO, CO2, NO, NO2 and SO2.

Investigating positively correlated variables, T and WSshowed relatively higher correlations, with coefficients of0.60 and 0.39, respectively. The highest positive correlationobtained for Twas expected as the increase of O3 concentrationwas influenced by temperature because of temperaturedependence of the numerous reactions. As known fromliterature especially PAN chemistry is mostly responsible forthe dependence of O3 formation on temperature. O3 concen-trations have a tendency of increase under hot, sunny con-ditions favorable for photochemical O3 production (Tarasovaand Karpetchko, 2003; Chelani, 2009; Shan et al., 2009).Furthermore lower humidity usually corresponds to highertemperatures, higher solar radiation and higher O3 formationrates as seen from Table 2. It was obtained that temperature

HC T H P R SR WS WD

0.555 0.608 −0.363 0.006 −0.064 0.233 0.394 −0.3540.205 −0.068 −0.109 −0.100 −0.038 0.141 −0.118 0.0340.280 −0.267 0.085 0.007 0.010 0.063 −0.212 0.1120.388 −0.301 0.104 −0.029 0.000 −0.037 −0.398 0.2520.011 −0.152 −0.002 −0.098 0.035 −0.018 −0.058 0.1000.255 −0.086 −0.068 −0.046 −0.034 −0.112 −0.128 0.127

0.537 −0.633 0.292 0.209 0.097 −0.218 −0.300 0.3071 −0.489 0.074 0.139 0.060 −0.224 −0.133 0.202

1 −0.605 −0.414 −0.135 0.403 0.210 −0.2801 0.230 0.112 −0.330 −0.299 0.273

1 0.019 −0.130 0.063 −0.0591 −0.086 0.020 0.018

1 0.074 −0.1291 −0.470

1

ds for ground-level ozone modeling, Atmos. Res. (2011),

5.7-8.8

3.8-5.7

2.1-3.8

0.5-2.1

Resultant vector

9 deg

NORTH

EASTWEST

SOUTH

15%

12%

9%

6%

3%Wind Speed

Fig. 2. Wind rose for the region.

a

0

0,2

0,4

0,6

0,8

1

1,2

0 0,2 0,4 0,6 0,8 1 1,2

measured values

pred

icte

d va

lues

6 B. Özbay et al. / Atmospheric Research xxx (2011) xxx–xxx

was very effective onO3 formationwhereas solar radiationwasnot effective as expected. As seen from the table correlationcoefficient for the relation betweenO3 and SRwas found as 0.23which is considerably lower than that of T. WS was anotherparameter exhibiting positive correlation with O3. Correlationbetween WS and O3 can be expressed with the correlationcoefficient of 0.39.

In the study, WDwas measured in terms of degrees (in thescale of 0–360°). As known correlation analysis can also beapplied in cases involving variables definedwithdifferentunits.In this context, relationship betweenWDandO3 levels could bedetermined with correlation analysis (Abdul-Wahab et al.,2005). We have calculated Pearson correlation coefficient forthis relation as −0.35 in the present study. Seasonal correla-tions between O3 levels andWDwere calculated as−0.34 and−0.24 respectively forwarmingand coolingperiods.Wind rosegiven in Fig. 2 summarizes WS and WD for Dilovasi. As seenfrom the figure, winds of the region are not strong. Dependingon this,WDandWSwerenot found effective onother variables.

0

0,2

0,4

0,6

0,8

1

1,2

0 0,2 0,4 0,6 0,8 1 1,2measured values

pred

icte

d va

lues

Fig. 3. MLR analysis of annual hourly ozone concentrations.

b

1

0 0,2 0,4 0,6 0,8 1 1,2

measured values

0

0,2

0,4

0,6

0,8

1,2

pred

icte

d va

lues

Fig. 4. MLR analysis of seasonal ((a) warming; (b) cooling periods) hourlyozone concentrations.

Please cite this article as: Özbay, B., et al., Multivariate methods for ground-level ozone modeling, Atmos. Res. (2011),doi:10.1016/j.atmosres.2011.06.005

4.2. Multiple regression analysis

MLR modeling of O3 concentrations were carried out inorder to find predictive equations for the concentration of the

Table 4Rotated component matrix.

Component

1 2 3 4

NO2 0.786 0.056 0.307 −0.090NO 0.786 0.026 0.093 0.095PM10 0.639 −0.047 −0.022 0.185SO2 0.561 −0.134 0.017 −0.503NMHC 0.544 0.503 0.026 −0.108O3 −0.494 −0.478 −0.455 −0.065T −0.257 −0.838 −0.223 −0.082P −0.068 0.644 −0.295 −0.171H −0.151 0.607 0.466 0.036CH4 0.480 0.585 0.245 0.164SR 0.089 −0.555 −0.117 −0.200WS −0.185 −0.026 −0.810 0.089WD 0.087 0.109 0.766 0.054CO 0.261 −0.075 0.065 0.826R −0.027 0.217 −0.056 0.269

Table 3The explained variance of principal components.

Component Initial eigenvalues Extraction sums of squared loadings Rotation sums of squared loadings

Total % of Variance Cumulative % Total % of Variance Cumulative % Total % of Variance Cumulative %

O3 4.283 28.553 28.553 4.283 28.553 28.553 2.942 19.616 19.616SO2 2.057 13.715 42.268 2.057 13.715 42.268 2.706 18.042 37.658NO 1.336 8.905 51.173 1.336 8.905 51.173 1.989 13.261 50.919NO2 1.151 7.675 58.848 1.151 7.675 58.848 1.189 7.929 58.848CO 0.982 6.547 65.396PM10 0.950 6.335 71.731CH4 0.847 5.644 77.375NMHC 0.724 4.825 82.201T 0.633 4.222 86.423H 0.525 3.501 89.923P 0.504 3.362 93.285R 0.361 2.406 95.692SR 0.285 1.901 97.592WS 0.183 1.217 98.810WD 0.179 1.190 100.000

7B. Özbay et al. / Atmospheric Research xxx (2011) xxx–xxx

pollutants and meteorological factors. MLR analysis wasperformed for annual and seasonal periods as O3 levels showseasonal variations.Monthly averageO3and temperature levelswere presented in Table 1(a) and (b). As seen from the table O3

levels show increasing tendency in warmer months whereasdecrease was seen in cooler months. This can be explainedwith the high correlation between O3 concentrations andtemperature (Table 2). In the study April, May, June, July andAugust months were evaluated in warming period whileNovember, December, January, February and March wereevaluated as cooling period.

Fig. 5. Principle components

Please cite this article as: Özbay, B., et al., Multivariate methodoi:10.1016/j.atmosres.2011.06.005

Data of warming and cooling periods were evaluatedseparately byMLR analyses in order to testmodelling efficiencyfor different conditions.

In the study, Model 1 evaluated the usability of MLR forannual O3 prediction. Relationship between annually mea-sured and predicted O3 concentrations was presented inFig. 3. As seen from the graphic in Fig. 3, there was a strongcorrelation between measured and predicted values. Adjust-ed R2 was calculated as 0.90 for Model 1.

Although accuracy of measurements was satisfactory asmentioned before, effects of measurement errors on modelingstudieswerealso investigated.With this aim,MLRanalyseswereapplied consideringmeasurement error of eachparameter. Littlevariance (≈0.01) was obtained in standard error of theestimates whereas no variance was noted in adjusted R2 values.

Data measured during warming period was used forModel 2. Fig. 4(a) shows the predicted O3 values versus bymeasured ones for this period. Adjusted R2 for Model 2 wasdetermined as 0.92. Similarly efficiency of MLR analysis forcooling period was investigated in Model 3. Calculated R2 ofModel 2 was 0.85 (Fig. 4(b)).

Model parameters and equations obtained from MLRanalysis were given in Table 5.

4.3. Principle component analysis

PCA was applied by using annual data in order to decreasethe number of components explaining O3 formation. In PCAapplications varimax rotation was used for maximizing thesum of the variances of the squared loadings. Table 3 and 4

for ozone prediction.

ds for ground-level ozone modeling, Atmos. Res. (2011),

variable x y z

O3 -0.317 -0.298 -0.780

SO2 0.395 -0.049 0.080

NO 0.837 0.058 0.093

NO2 0.812 0.019 0.313

CO 0.183 0.085 0.183

PM10 0.444 -0.062 0.136

CH4 0.283 0.417 0.557

NMHC 0.300 0.296 0.447

T -0.216 -0.911 -0.343

H -0.001 0.572 0.243

P 0.004 0.519 -0.174

R -0.011 0.153 0.050

SR 0.082 -0.382 -0.211

WS -0.247 -0.002 -0.454

WD 0.108 0.120 0.427

Fig. 6. 3D plot of rotated loadings for truncated data.

8 B. Özbay et al. / Atmospheric Research xxx (2011) xxx–xxx

summarized the results of the varimax rotation on the fifteenPCs together with the amount of variance explained by eachcomponent. PCs with an eigenvalue greater than or equal to 1,are usually considered as being of statistical significance (theKaiser criterion). From Table 3, it can be seen that the firstfourth PCs accounted for 59% of the total variation.

PCs obtained from PCA analysis of 15 different inputvariables were determined by considering coefficients pre-sented inTable 4. Thesenewcomponentswere clearly shown inFig. 5. Variables effecting O3 formation were decreased to fourprinciple groups after PCA application as seen from Fig. 5.

As mentioned in the above paragraph variables constitut-ing the O3 in the appearance of the three dimensional space isshown in Fig. 6.

Multiple regression analysis was repeated by using PCsobtained fromPCA (Model 4). PCswith eigenvalues higher than1were used in themodel so formation of O3 could be explainedwith adjustedR2 of 0.63.Asmentionedbeforemaingoals of PCAare extracting the most important information from the datatable and simplifying the description of the data set (Abdi andWilliams, 2010). Therefore usage of PCA provides processfacilities for the models working with numerous variables(Abdul-Wahab et al., 2005; Sousa et al., 2007; Kovač-Andrićet al., 2009). Table 5 summarizes the results of all MLR applica-tions. Standard errors of the estimate given in the table enabledetermining distribution of actual values around the regressionlayer. Compared to standard errors of the estimates forModel 1

Table 5Results of the different MLR analyses.

Modelnumber

Number ofsamples (N)

Adjusted R2 Standard error ofthe estimate

Equation of the mo

1 5432 0.90 5.45 O3(t+1)=−74.80+0.002 NMHC(t)+0.0

2 2655 0.92 5.50 O3(t+1)=−63.833+0.004 CH4(t)−0.0020.481 WS(t)+0.001

3 1935 0.85 3.57 O3(t+1)=−67.753+0.091 T(t)+0.007 H

4 5432 0.63 0.06 O3(t+1)=0.123−0.

Please cite this article as: Özbay, B., et al., Multivariate methodoi:10.1016/j.atmosres.2011.06.005

and Model 4, it was seen that Model 4 exhibited a much lowererror value. As known lower value of the standard error is ameasure of the accuracy of predictions. This clearly showssuccess of PCA for prediction of ozone levels.

5. Conclusions

Mainobjectiveof thisworkwas toevaluate theperformanceof multivariate methods to predict O3 concentrations of 1 hlater using air pollutant concentrations (PM10, SO2, NO, NO2,CO, O3, CH4, NMHC) and meteorological parameters (T, R, H, P,WD, WS and SR) as predictors. Annual and seasonal periodswere evaluated individually in MLR models and calculated R2

values were found as 0.90, 0.92, 0.85 respectively for Models 1,2, 3. Best modeling efficiencywas obtained for warming periodwith R2=0.92. This result is valuable as O3 formation increasesin summer seasons universally. Furthermore, previous re-searches have also proved the efficiency of MLR models inprediction of O3 levels for time periods of greater than 1 h. Itwas successfully used for one-day ahead predictions of O3

concentrations (Sousa et al., 2006; Chaloulakou et al., 1999).In order to reduce the number of input variables, PCA was

applied by using hourlymeasured data for one year period. Bythis way fifteen initial variables were decreased to four PCs.Determined PCs explained O3 formation with adjusted R2 of0.63 in repeated MLR analysis (Model 4). Efficiency of PCAwas found successful considering lower standard error of the

del

0.89O3(t)−0.005 SO2(t)+0.025NO(t)+0.043NO2 (t)−0.002CH4(t)−83T(t) +0.033H(t)+0.075P(t)+0.908R(t)+0.006 SR (t)+0.33 WS(t)0.888 O3(t)−0.027 SO2(t)+0.025NO(t)+0.045NO2 (t)+0.009 PM(t)−NMHC(t)+0.138 T(t)+0.044 H(t)+0.064P(t)+0.584 R(t)+0.004 SR (t)+WD(t)

0.884 O3(t)−0.011 SO2(t)+0.022NO2 (t)−0.003 PM(t)+0.001 CH4(t)+(t)+0.066 P(t)+0.877 R(t)+0.001 SR (t)+0.093 WS(t)048 PC1(t)−0.053PC2(t)−0.047 PC3(t)−0.009PC4(t)

ds for ground-level ozone modeling, Atmos. Res. (2011),

9B. Özbay et al. / Atmospheric Research xxx (2011) xxx–xxx

estimation in Model 4. Briefly, this method is promising forthis kind of studies as it provides process facilities andsuccessful prediction efficiencies.

Results of the paper also provide an important clue aboutthe surface O3 levels in the region. Highest and lowest averageO3 levels were determined in August and March as 43.10 and5.07 μg/m3, respectively. Furthermore average O3 levels werefound as 21.30 and 11.46, respectively forwarming and coolingperiods. These results demonstrated that O3 levels wereaffected by meteorological factors evidently in the studiedregion. This was also confirmed with the results of bivariatecorrelation analysis as O3 exhibited high positive correlationwith temperature. Furthermore remarkable negative correla-tions were seen for the relations between O3 and its mainprecursors such as CH4, NMHC, NO2.

Acknowledgement

This studywas funded by theUniversity of Kocaeli ResearchFund under Project No. 2009–013.

References

Abdi, H., Williams, L.J., 2010. Principal component analysis. WIREs Compu-tational Statistics. 2, 433–459.

Abdul-Wahab, S.A., 2001. IER photochemical smog evaluation and forecast-ing of short-term ozone pollution levels with artificial neural networks.Trans IChemE. 79, 117–128.

Abdul-Wahab, S.A., Al-Alawi, S.M., 2002. Assessment and prediction oftropospheric ozone concentration levels using artificial neural networks.Environ. Model. & Soft. 17, 219–228.

Abdul-Wahab, S.A., Bakheitb, C.S., Al-Alawi, S.M., 2005. Principal componentand multiple regression analysis in modelling of ground-level ozone andfactors affecting its concentrations. Environ. Model. & Soft. 20, 263–1271.

Argiriou, A.A., 2007. Use of neural networks for tropospheric ozone timeseries approximation and forecasting – a review. Atmos. Chem. & Phys.Diss. 7, 5739–5767.

Blankinship, D.J., 1996. A discussion of the spatial and temporal variability ofozone concentrations along the front range of Colorado, University ofColorado at Boulder. Program in Atmospheric and Oceanic Sciences.

Brauer, M., Brook, J.R., 1997. Ozone personal exposures and health effects forselected groups residing in the FraserValley. Atmos. Environ. 31, 2113–2121.

Brulfert, G., Galvez, O., Yang, F., Sloan, J.J., 2007. A regional modelling study ofthe high ozone episode of June 2001 in southern Ontari. Atmos. Environ.41, 3777–3788.

Buhr, M., Parrish, D.D., Elliot, J., Holloway, J., Carpenter, J., Goldan, P., Kuster,W., Trainer, M., Montzka, S., McKeen, S., Fehsenfeld, F., 1995. Evaluationof ozone precursor source types using principal component analysis ofambient air measurements in rural Alabama. J. Geophys. Res. 100,22853–22860.

Chaloulakou, A., Assimacopoulos, D., Lekkas, T., 1999. Forecasting dailymaximum ozone concentrations in the Athens basin. Environ. Monit.Assess. 56, 97–112.

Chatterton, T., Dorling, S., Lovett, A., Stephenson, M., 2000. Air quality inNorwich, UK:multi-scalemodeling to assess the significance of city, countyand regional pollution sources. Environ. Monit. Assess. 65, 425–433.

Chelani, A.B., 2009. Statistical persistence analysis of hourly ground levelozone concentrations in Delhi. Atmos. Res. 92, 244–250.

Clark, T.L., Karl, T.R., 1982. Application of prognostic meteorological variablesto forecasts of daily maximum one-hour ozone concentrations in thenortheastern United States. J. Appl. Meteor. 21, 1662–1671.

Cox, W.M., Chu, S.H., 1991. Using meteorological/ozone relationships toestablish an index of potential ozone severity. 84th AWMA AnnualMeeting and Exhibition, AWMA, Vancouver, British Columbia, Canada.

Debaje, S.B., Kakadeb, A.D., 2009. Surface ozone variability over westernMaharashtra. India. J. Hazard. Mater. 161, 686–700.

Ding, S., Zhang, P., Ding, E., Yin, S., Naik, A., Deng, P., Gui, W., 2010. On theapplication of PCA technique to fault diagnosis. Tsınghua Scı. & Technol.15, 138–144.

Please cite this article as: Özbay, B., et al., Multivariate methodoi:10.1016/j.atmosres.2011.06.005

Duan, J., Tan, J., Yang, L., Wu, S., Hao, J., 2008. Concentration, sources andozone formation potential of volatile organic compounds (VOCs) duringozone episode in Beijing. Atmos. Res. 88, 25–35.

Hsieha, K.-L., Yang, I.-C., 2008. Incorporating PCA and fuzzy-ART techniquesinto achieve organism classification based on codon usage consideration.Comput. Biol. Med. 38, 886–893.

Jun, T., Xia, Z.G., Wang, H., Li, W., 2007. Temporal variations in surface ozoneand its precursors and meteorological effects at an urban site in China.Atmos. Res. 85, 310–337.

Keskin, G.A., Ilhan, S., Özkan, Ç., 2010. The fuzzy ART algorithm: acategorization method for supplier evaluation and selection. Expet SystAppl. 37, 1235–1240.

Kovač-Andrić, E., Brana, J., Gvozdić, V., 2009. Impact of meteorological factorson ozone concentrations modelled by time series analysis andmultivariate statistical methods. Ecol. Inf. 4, 117–122.

Lam, K.C., Tao, R., Lam, M.C.K., 2010. A material supplier selection model forproperty developers using Fuzzy Principal Component Analysis. Auto. inConst. 19, 608–618.

Lavecchia, C., Angelino, E., Bedogni, M., Bravetti, E., Gualdi, R., Lanzani, G.,Musitelli, A., Valentini, M., 1996. The ozone patterns in the aerologicalbasin of Milan (Italy). Envıron. Softw. 11, 73–80.

Lehman, J., Swinton, K., Bortnick, S., Hamilton, C., Baldridge, E., Eder, B., Cox,B., 2004. Spatio-temporal characterization of tropospheric ozone acrossthe eastern United States. Atmos. Environ. 38, 4357–4369.

Lippmann, M., 1989. Ozone health effects and emerging issues in relation tostandards setting. Stu. in Environ. Sci. 35, 21–33.

Lu, H.-C., Hsieh, J.-C., Chang, T.-S., 2006. Prediction of daily maximum ozoneconcentrations from meteorological conditions using a two-stage neuralnetwork. Atmos. Res. 81, 124–139.

Mayer, H., 1999. Air pollution in cities. Atmos. Environ. 33, 4029–4037.Ozden, O., Dogeroglu, T., Kara, S., 2008. Assessment of ambient air quality in

Eskişehir. Turkey. Environ. Int. 34, 678–687.Paoletti, E., 2006. Impact of ozone on Mediterranean forests: a review.

Environ. Pollut. 144, 463–474.Pearson, K., 1901. On lines and planes of closest fit to systems of points in

space. Philos. Mag. 2, 559–572.Peton, N., Dray, G., Pearson, D., Mesbah, M., Vuillot, B., 2000. Modelling and

analysis of ozone episodes. Environ. Model. Soft. 15, 647–652.Pulikesi, M., Baskaralingam, P., Rayudub, V.N., Elango, D., Ramamurthi, V.,

Sivanesan, S., 2006. Surface ozone measurements at urban coastal siteChennai. India. J. Hazard. Mater. B137, 1554–1559.

San Jose, R., Stohl, A., Karatzas, K., Bohlerd, T., James, P., Pe´ rez, J.L., 2005. Amodelling study of an extraordinary night time ozone episode overMadrid domain. Environ. Model. & Soft. 20, 587–593.

Schenone, G., Lorenzini, G., 1992. Effects of regional air pollution on crops inItaly. Agric. Ecosyst. Environ. 38, 51–59.

Seinfeld, J.H., Pandis, S.N., 2006. Atmospheric chemistry and physics: from airpollution to climate change. John Wiley & Sons, Inc., New York.

Shan, W., Yin, Y., Zhang, J., Ding, Y., 2008. Observational study of surfaceozone at an urban site in East China. Atmos. Res. 89, 252–261.

Shan, W., Yin, Y., Lu, H., Liang, S., 2009. A meteorological analysis of ozoneepisodes using HYSPLIT model and surface data. Atmos. Res. 93,767–776.

Shi, J.P., Harrison, R.M., 1997. Regression modelling of hourly NOx and NO2

concentrations in urban air in London. Atmos. Environ. 31, 4081–4094.Sousa, S.I.V., Martins, F.G., Pereira, M.C., Alvim-Ferraz, M.C.M., 2006.

Prediction of ozone concentrations in Oporto city with statisticalapproaches. Chemosphere 64, 1141–1149.

Sousa, S.I.V., Martins, F.G., Alvim-Ferraz, M.C.M., Pereira, M.C., 2007. Multiplelinear regression and artificial neural networks based on principalcomponents to predict ozone concentrations. Environ. Model. & Soft. 22,97–103.

Tarasova, O.A., Karpetchko, A.Y., 2003. Accounting for local meteorologicaleffects in the ozone time-series of Lovozero (Kola Peninsula). Atmos.Chem. Phys. Discuss. 655–676.

Tian, Y.I., Biswas, P., Pratsinis, S.E., Hsieh, W.M., 1989. Principal componentanalysis for particulate source resolution in cleanrooms. J. Environ. Sci.22–27 November/ December,.

TUIK, (Statistical Department of Turkey), [online] http://tuikapp.tuik.gov.tr/adnksdagitapp/adnks.zul. Accessed 2010.

Vaidya, O.C., Howell, G.D., Leger, D.A., 2000. Evaluation of the distribution ofmercury in lakes in Nova Scotia and Newfoundland. Water, Air, and SoilPollut. 117, 353–369.

Zhang, C., Wu, L., Luo, Y., Zhang, H., Christie, P., 2008. Identifying sources ofsoil inorganic pollutants on a regional scale using a multivariatestatistical approach: role of pollutant migration and soil physicochemicalproperties. Environ. Pollut. 151, 470–476.

ds for ground-level ozone modeling, Atmos. Res. (2011),