A Preliminary Assessment of the Impacts of 1 Multiple Temporal-scale Variations in Particulate 2 Matter on its Source Apportionment

8 9 State Environmental Protection Key Laboratory of Urban Ambient Air Particulate 10 Matter Pollution Prevention and Control & Center for Urban Transport Emission 11 Research, College of Environmental Science and Engineering, Nankai University, 12 Tianjin 300350, China, 13 2 Chinese Research Academy of Environmental Sciences, Beijing 100012, China, 14 3 College of Software, Nankai University, Tianjin 300350, China, 15 Department of Physics, University of Nevada Reno, Reno, Nevada, USA 89557, 16 Earth and Atmospheric Sciences, Georgia Institute of Technology, Atlanta, GA 30332 17


Introduction
Aerosol pollutants have become a major problem in recent years (Huang et al., 2014;van Donkelaar et al., 2015), due to its negative influences on visibility, climate change and human health (Langridge et al., 2012;Cheng et al., 2015;Butt et al., 2016;Ding et al., 2017).Variations in aerosol concentrations and chemical species reflect influences from multiple factors, such as local emissions sources and weather conditions, etc. (Keim et al., 2005).Observed concentrations of pollutants, in general, have characteristic variations, which are influenced by data noise, source intensities, shortterm fluctuations, source seasonal variation, meteorological condition, climate, policy, and economic conditions (Milanchus et al., 1998;Wise and Comrie, 2005).
Online instruments can provide high time resolution data of particulate matter (PM) and chemical species, and these instruments have been widely applied in the detection of pollutants (Tchepel et al., 2010;Du et al., 2011;Zheng et al., 2015;Gao et al., 2016).
More and more studies on aerosol pollution have become dependent on high temporal resolution observations due to their capabilities in revealing multiple temporal-scale fluctuations of the aerosol concentrations that tend to arise from different physical, chemical and dynamical processes.For example, during a high pollution period, pollutant concentrations increase rapidly by several times over a short time period, and such an increase tends to result from changing meteorological conditions.Hogrefe et al. (2000) suggested that the time series of pollutant concentrations can be decomposed into four components.The first component is the intra-day component with periods less than 12 h and is typically linked to fast-acting, local emission sources and local-level processes (Tchepel et al., 2010).The second is the diurnal component dominated by 12-48 h periodicity.The third is the synoptic component mainly driven by 2-21 day fluctuations in weather patterns and short-term fluctuations in emissions.The last baseline component is related to the low-frequency fluctuations with periods greater than 21 days, which might including seasonal or long-term scale variation in emissions, climate, policy, etc. (Rao et al., 1997;Wise and Comrie, 2005).
Pollution sources are the key drivers of aerosol pollution.Understanding source impacts on aerosols is important for the control of air pollution (Zhao et al., 2017).
Factor analysis models are widely used for estimating source impacts.These models include principal component analysis/multiple linear regression (PCA/MLR), Unmix, positive matrix factorization (PMF), and Multilinear Engine (ME-2) (Thurston and Spengler, 1985;Paatero and Tapper, 1994;Henry and Christensen, 2010;Yin et al., 2015;Zong et al., 2016).Among these, ME-2 is a particularly useful tool and has been widely used in source apportionment studies (Paatero, 1999;Amato et al., 2009;Peng et al., 2016).Factor analysis models depend on the variation of chemical species in aerosols (which reflects the temporal variation of sources) to extract source categories and calculate their contributions.Therefore, multiple temporal-scale variations in the raw online datasets associated with various factors (e.g., data noise and weather fluctuations) can have significant impacts on the source apportionment results using factor analysis models.This is the main motivation behind our analysis, i.e., to decompose the raw online datasets into multiple temporal-scale components and then The Kolmogorov-Zurbenko (KZ) filter used in our study for extraction of a specific temporal-scale component (TS component hereafter) is a low-pass filter that has been widely used for decomposing temporal variations in O3, PM, and chemical species (Rao et al., 1997;Hogrefe et al., 2000Hogrefe et al., , 2006;;Wise and Comrie, 2005;Tchepel et al., 2010).Hogrefe et al. (2006) reported that the synoptic component associated with synoptic scale weather fluctuations has the largest relative contribution to the total variance of hourly PM2.5 concentrations, and the relative contributions of other components to total PM2.5 mass concentrations varies by chemical species in PM2.5.In addition, the noise of data might impact the analysis, and efforts have been made to remove the noise (Kuebler et al., 2001;Tchepel et al., 2010;Henneman et al., 2015).
Our earlier studies demonstrated that the time resolution of the data could influence the source apportionment results (Peng et al., 2016).Tchepel et al, (2010) used the KZ filter to remove noise in the PM data, fed filtered PM data into air quality models and showed that the model performance improved.In this study, we make a preliminary assessment of the impacts of multiple temporal-scale variations in PM data on source apportionment with PM2.5 (PM with an aerodynamic diameter less than 2.5 μm) and its chemical species observed in Beijing, China.Wavelet analysis was first used to evaluate the periodicities of the PM and chemical species concentrations Twenty-three chemical species were selected for analysis, including NH4 + , Na + , Mg 2+ , Cl -, NO3 -, SO4 2-, K, Ca, Cr, Mn, Fe, Ni, Cu, Zn, As, Se, Ag, Cd, Ba, Hg, Pb, OC and EC.The principles of these instruments and QA/QC are described in detail by Gao et al. (2016).

Source Impact Model
ME-2, a general factor analysis model developed by Paatero (1999), was applied to estimate the impacts of source categories at a location of interest.It is a general solver of widely different multilinear and quasi-multilinear problems (Ramadan et al., 2003) with the ability to deal with models consisting of a sum of products of unknowns.
Instead of being restricted to a specific structure, ME-2 is defined in a ''script file'' that is written in a special-purpose programming language.It has efficient performance as it runs in DOS, which made it faster than those models with graphical interfaces (Ramadan et al., 2003).ME-2 decomposes original matrix  (×) into source impact matrix  (×) and source chemical species (source profile) matrix  (×) , as follow: Variable  (×) is the chemical species concentrations (unit: μg m -3 ) in PM2.5 that are observed at the receptor site;  (×) is the residual matrix;  and  are the sample size and chemical species number, respectively; and  is the number of sources.
The basic principle of ME-2 also can be expressed as follow: where   is the element in matrix  (×) , which is the measured concentration of the  ℎ specie in the  ℎ sample (μg m -3 );   is the element in matrix  (×) and is the impact of the  ℎ source on the  ℎ sample;   is the element of matrix of  (×) and is the concentration of the  ℎ specie in the  ℎ source (source profile); and   is the element in the residual matrix (Hopke, 2003).
In ME-2, a priori information (e.g. chemical profiles and ratios) can be incorporated as a target to be approximately accomplished.be handled in form of auxiliary equations (Paatero, 1999).Auxiliary equations are included as additional terms ux Q in an enhanced object function enh Q (Amato et al., 2009;Amato and Hopke, 2012), the equation can be written as follows: The term   is described as follows: where   is the uncertainty in the  ℎ species for the  ℎ sample;   has the same meaning as is described in Eq.( 2).
One of the simplest forms of the auxiliary equation is the ''pulling equation'' (Paatero and Hopke, 2009), consisting of pulling   (for instance) toward the specific target value   : where    is the uncertainty connected to the pulling equation or softness of the pull; and   is the element of factor loading.The task of ME-2 is to calculate a minimum  ℎ value or balance the minimization of the values   and   in the iterative process (Paatero and Hopke, 2009).
For ME-2, it requires that every element in the input dataset (matrix X) be a nonnegative value.original data.

Temporal Scale Analysis
The KZ filter is a widely applied filtering technique due to its powerful separation characteristics, simplicity, and ability to handle missing data (Rao et al., 1997;Hogrefe et al., 2006).The principle of KZ filter is described as follow: is the length of the moving average window, which is an odd number;  (+) is the ( + ) ℎ original value,   is the average value.Then the   as the input data and calculate according to Eq. ( 6).After k times (number of iterations) calculation,   () is expressed as: () is removed the variations that frequency lower than w (cutoff frequency).
is the number of iterations.Before conducting the KZ filter, the data is log transformed for variance stabilization (Hogrefe et al., 2000).The separation point w , between the high-frequency and low-frequency component, is a function of the filter parameters m and k (Rao et al., 1997).The equation can be written as follows: Selecting proper filter parameters m and k could remove the temporal component at a specific frequency from the original dataset.

TS Component Removed Datasets
To exam the impact of the four TS components on the source impacts, datasets without (− ) (RI dataset) is the concentration dataset with the intra-day TS component removed from the original dataset (μg m -3 ) and it contains the diurnal, synoptic, and baseline TS components. ( ) (RD dataset),  ( ) (RS dataset), and  ( ) (RBL dataset) are the datasets with the diurnal, synoptic, and baseline TS components singly removed from the original dataset, respectively.As there were many negative values in the RBL dataset, these data were analyzed by PCA to qualitatively identify the sources of PM2.5.The original, RI, RD, and RS datasets were run by ME-2 for the source apportionment, and their results were compared.Also, few negative values (very low count) were replaced with a value equal to half of the detection limits.The average absolute error (AAE, see the SI) and correlation analysis were employed to compare the differences in the source impacts between original dataset and datasets with removed TS components.AAE was employed and is calculated as follows (Javitz et al., 1988): where,   is the  value for the  ℎ species and  is the number of samples.
is the concentration (μg m -3 ) of the  ℎ species for the  ℎ sample from the RI, RD, or RS datasets.  is the concentration (μg m -3 ) of the  ℎ species for the  ℎ sample from the original dataset.The larger AAE value and the lower correlation coefficients (r) indicate a larger difference in source impacts between the original dataset and modified datasets, suggesting that the corresponding TS component has a larger influence on the observed concentrations.

TS Component Influence on Concentrations
The influence of each TS component on the pollutant concentration variation and the concentration levels were investigated.The original dataset was decomposed into intraday, diurnal, synoptic, and baseline TS components by using the KZ filter (Figure 1).
The variation analysis was then employed to study each TS component contribution to the total variance of PM2.5 and the chemical species concentrations (Table 1).We placed emphasis on investigating PM2.5 and the source markers (e.g.SO4 2-, Ca, OC, etc.), because the variation of those markers can reflect the source emission pattern to some extent.
The sample size of the four TS components was less than original dataset, because the KZ filter was iterated with a moving average with a specified length and resulted in missing head and tail data of the original data.The same period (from 24 July 2014 to 10 August 2014) of the original datasets with the same size were selected for comparison and analysis.Among all the species, PM2.5 and NO3 -had similar trends: the diurnal and synoptic TS components had larger amplitudes and higher relative contributions to the total variance of PM2.5 (diurnal: 36%, synoptic: 32%) and NO3 - (diurnal: 36%, synoptic: 32%) than the intra-day and baseline TS components.SO4 2- and NH4 + showed similar variability: synoptic TS component had the largest amplitude and had the largest relative contributions to the total variance of SO4 2-(48%) and NH4 + (54%) concentrations, followed by the baseline, diurnal and intra-day TS components.
OC was relatively different from the species mentioned above.For OC, the relative contribution of the synoptic TS component was the largest (56%), followed by diurnal (23%), baseline (12%) and intra-day TS components (9%).Species from primary emission sources (such as EC, Ca, Fe, etc.) showed different patterns, compared with the secondary species (NO3 -, SO4 2-, NH4 + ) discussed above.For EC and Ca, the diurnal TS component had the largest relative contribution to the total variance of concentrations, accounting for 47% and 45%, respectively.The synoptic (28%) and intra-day (40%) TS component was the second largest contributor to the total variance of EC and Ca concentrations, respectively.For Fe, diurnal and synoptic TS components had larger amplitudes and higher relative contributions to the total variance than intraday and baseline TS components.For other elements, diurnal or intra-day TS components had the largest amplitudes and were the larger contributors to the total variance of the concentrations.Secondary organic carbon (SOC) also has been estimated using the OC/EC ratio (see Supporting Information), and the influence of TS component on the SOC was investigated (Table 1).The average concentration of SOC were mainly influenced by chemical reaction (photochemical, liquid phase or heterogeneous reaction) and meteorological conditions (Buzcu et al., 2006;Jung et al., 2010, Martin et al., 2014).Therefore, species with similar TS component contributions trends may have similar sources or influencing factors.
To investigate the influence of the TS components on concentration levels, partial statistical analysis and AAE analysis were performed on the PM2.5 and source markers (NO3 -, SO4 2-, NH4 + , Ca, Fe, OC, and EC) from five ambient datasets (including the original, RI, RD, RS, and RBL datasets).The results are shown in Figure 2 and Table S1.

Source Impacts on PM2.5 Concentrations
Four datasets, including the original, RI, RD and RS datasets, were respectively introduced into ME-2 to identify the sources of PM2.5.The RBL was analyzed by PCA, as several negative values were in this dataset (the ME-2 model only allows nonnegative input values).The source apportionment results were explored, including source profiles and source impacts, to investigate the source impacts on the PM2.5 concentrations under the influence of different TS components.
For all four datasets, 3 to 7 factors were tested to determine the optimal number of factors (source categories).There are some criteria for choosing the appropriate number of factors, including the Q values, physical meaningfulness of the factor profiles, the reasonableness of source impacts, and goodness of fit for PM2.5 and chemical species concentrations.After testing, four source categories were identified using ME-2 from the original, RD, and RS datasets; five sources were obtained from the RI dataset.When the calculated Q value close to the theoretical Q (Qthe), the corresponding results might be acceptable (Hopke, 2003).The Q values of each solution calculated by ME-2 are displayed in Table S3 and were close to the theoretical Q, further suggesting that the results are acceptable.
Performance of ME-2 was evaluated by analyzing the goodness of fit for the modeled and measured PM2.5 and chemical species mass concentrations (slope, r).
Figure S3 illustrates the slope and r results, respectively.For PM2.5, the slopes (ranging from 1.0 to 1.1) and r (ranging from 0.8 to 1) were close to 1, suggesting the good performance of ME-2 obtained for the five runs.For original dataset run, 13 out of 23 chemical species (e.g.SO4 For the original dataset, four factor profiles were obtained (Figure 3).characterized by Ca and Fe, which is linked to crustal dust (Pant and Harrison, 2012).
Factor 2 had OC and EC, which are markers of vehicle emissions (Ramadan et al., 2003).Factor 3 was identified as coal combustion due to high loadings of OC, EC and Ca (Ramadan et al., 2003).Factor 4 was secondary formation due to elevated SO4 2-, NO3 -and NH4 + (Pant and Harrison, 2012).According to the previous studies (Yu et al., 2013), coal combustion, secondary formation, vehicle emissions and crustal dust were the dominating sources of PM2.5 in Beijing.Five sources were obtained from the RI dataset, including coal combustion, crustal dust, secondary formation, secondary nitrate and vehicle emissions (Figure 3).For the RD and RS datasets, coal combustion, secondary formation, secondary nitrate and vehicle emissions were identified (Figure 3), and crustal dust was not identified and was mixed with vehicle emissions from the two datasets.It was an expected result that ME-2 failed to identify the crustal dust source for the RD dataset.Because ME-2 extracts factors based on the chemical species variation pattern (the marker species variation can reflect the source emission pattern over the time), crustal dust markers (Ca and Fe) lost much variance and could not reflect the expected pattern after removing the diurnal TS component (the largest contributor to the total variance of Ca and Fe) (Table 1).As for the solution for RS dataset, the sulfate source (not the secondary formation) and the nitrate source were distinguished as different factors (Figure 3).For the solutions of the original, RI, and RD datasets, secondary formation of sulfate and nitrate source were mixed together and extracted as one factor.We found that after removing the similar part of the variance (the information filtered by KZ filter) of sulfate and nitrate, the difference between sulfate The correlation was lowest (0.49) for the RS dataset compared with other datasets (0.68, 0.72, and 0.85 in the original, RI, and RD datasets, respectively).
The PCA results of the RBL dataset are listed in Table S4.Seven factors were extracted and accounted for 80.9% of the total variance, which had corresponding eigenvalues larger than 1 (can be considered as the potential sources).Factor 1 (19.6% of the variance) had high loadings for heavy metals, such as As, Se, Pb, etc. Factor 2 had high loadings for Ca and Ba, and relatively high loadings for Mn, Fe, and EC.This factor might be associated with crustal dust and had a 14.3% contribution to the variance (Pant and Harrison, 2012).High loadings were observed for SO4 2-, NO3 -, and NH4 + in factor 3, which is associated with secondary formation (Pant and Harrison, 2012).Factor 4 had relatively high loadings for Cu and Cl -, and factor 5 and 6 were characterized by heavy metals.After removing the baseline dataset, heavy metals and crustal dust were the dominant sources of PM2.5.According to the results of the RBL dataset, we found the evidence that intra-day and diurnal TS components had larger relative contributions to the total variance of element (heavy metals) concentrations, as these elements are mainly emitted from primary sources.
Because there were different sample sizes for the four datasets, we selected the same period of results (from 24 July 2014 to 10 August 2014) to study the influence of TS components on the source variation (Figure 4).The time series of source impacts from the RI, RD, and RS datasets were respectively used for correlation analysis with the corresponding results from the original dataset (Table S5).Vehicle emissions solutions from the RI (r = 0.45) and RD (r = 0.51) datasets had a higher correlation than the RS dataset (r = 0.25), suggesting that intra-day and diurnal TS components had stronger influences on the source pattern (variation) of vehicle emissions.For coal combustion, results from the RI, RD, and RS datasets had similar correlation coefficients (ranging from 0.74 to 0.82).The sulfate source was identified from the RS dataset (Figure 3), however, the sulfate source had the lowest correlation with secondary formation (Table S5) solutions from the original dataset.The correlation analysis of the nitrate source produced similar results, where the lowest correlation coefficient occurred between solutions from the RS dataset and original dataset, suggesting that secondary source impact variation is dominantly affected by synoptic scale influences.
To further investigate source impacts on PM2.5 from different TS components, we discussed the average impacts of individual source categories on PM2.5 from the datasets with removed TS components (Table 2).
Vehicle emissions, crustal dust, and coal combustion were combined together for the analysis (called as TPS: total primary sources), because crustal dust was mixed with vehicle emissions and coal dust for the RD and RS datasets, as mentioned above.
Secondary formation and nitrate source were also plus together for the discussion (called as TSS: total secondary sources).To better explore the influence of TS components, source impacts during the entire sampling period and pollution period were investigated separately.For the entire sampling period, the impacts of TPS obtained from the original, RI, RD, and RS datasets were similar to each other, ranging from 35.1 to 40.4 μg m -3 .This was an expected result because the intra-day, diurnal, and synoptic TS components had small influence on the concentrations levels of primary source markers (OC, EC, elements), as shown in Figure 2. The TSS solutions from the original, RI, and RD datasets exhibited similar source impacts, accounting for about 30 μg m -3 , which was higher than the solution from the RS dataset (21.2 μg m -3 ).
The synoptic TS component had impact on the SO4 2-, NO3 -and NH4 + concentrations, and removing this TS component may have resulted in lower impacts of the secondary sources.
During the pollution period (from 30 July to 4 August 2014, gray shadow shown in Figure 1), the highest concentration of PM2.5 was up to 183.7 μg m -3 at 1:00 am on 31 July 2014, with an average concentration of 85.5 μg m -3 .The TPS impacts derived from the original, RI, RD, and RS datasets were relatively stable, ranging from 30.3 μg m -3 to 37.4 μg m -3 (Table 2).The TSS impacts from the RS dataset (29.3 μg m -3 ) were lower than the solutions from the original, RI, and RD runs (about 51 μg m -3 ).The synoptic TS component increased the NO3 -, SO4 2-, and NH4 + concentrations (Figure 1), accounting for 58%, 57%, and 50% of their original average concentrations, respectively.This implies that the synoptic TS component had a larger impact on the secondary source than the primary source during the pollution period.Liu et al. (2017) reported that some haze episodes in North China Plain (including Tianjin) resulted from elevated relative humidity (RH) and stagnant weather conditions.The study proposed an inorganic aerosol formation mechanism for which the elevated RH and the inorganic fraction increased the aerosol liquid water content (LWC), then the liquid particles would uptake pollutants to form the aerosols.In this work, elevated RH and reduced wind speeds have been observed during the pollution period (Figure S4).To further confirm the assumption, the aerosol LWC was estimated using ISORROPIA II model (Guo et al., 2015).The LWC and total ions (NO3 -+SO4 2-+NH4 + ) depending on RH are shown in Figure S5.Under similar RH conditions, the total ions (ranged from 93 to 121 μg m -3 ) and LWC concentrations (ranged from 48 to 513 μg m -3 ) during the pollution period were higher than the corresponding values (total ions concentrations: 22 to 38 μg m -3 ; LWC: concentrations 12 to 108 μg m -3 ) during the non-pollution, suggesting high total ion and LWC concentrations.When the RH lower than 90% during the pollution period, the total ions concentrations remained relatively stable (about 100μg m -3 ), while the LWC concentrations increased.When the RH was higher than 90%, the total ions and LWC concentrations increased to 120 and 513 μg m -3 during the pollution period, respectively, suggesting the drastically increasing in LWC concentrations may have led to the elevated ions concentrations.Therefore, the stagnant weather conditions and the elevated RH may increase the inorganic concentrations during the pollution period in this work.
Overall, removing the different TS components had little influence on primary source impact levels, suggesting that primary source impact levels were mainly influenced by the source emissions.The secondary source impact levels were mainly influenced by synoptic influences and source emissions.
The BL dataset may be linked with source emissions, source seasonal variance, and long-term meteorological fluctuations.To study the source impacts from the baseline TS component, ME-2 was applied to the baseline dataset.Four sources were identified, including the nitrate source, secondary formation, coal combustion, and vehicle emissions (Figure S6).During the entire sampling period (Table 3), the average TPS and TSS impacts on PM2.5 mass concentrations were 29.9 μg m -3 (57%) and 22.8 μg m -3 (43%) respectively.The average impacts of TPS and TSS during the pollution period were higher than the corresponding average impacts during the entire sampling period, which were 35.6 μg m -3 and 26.0 μg m -3 , respectively (Table 3).TPS and TSS obtained same impact percentages from the entire period and pollution period, accounting for 58% and 42% of PM2.5 mass concentrations, respectively.The time series of TPS, TSS, and PM2.5 are shown in Figure S7 and suggest that the periodicities of TPS and TSS were not synchronized.In this work, the peak of the BL TS component of PM2.5 was obtained when the both the TPS and the TSS impact levels were high.

Conclusions
In this work, KZ filter was applied to decompose the time series of PM2. were compared.We found that removing some TS components affected the source identification.Four sources were obtained from the original and RI analyses, including crustal dust, vehicle emissions, coal combustion, and secondary formation.Crustal dust was not identified by ME-2 from the RD, RS, and BL datasets, possibly due to the fact that much of the information regarding the markers of crustal dust (e.g.Ca) was lost after removing the corresponding TS components.This suggests that the diurnal and synoptic TS components of chemical species were important for identifying the crustal dust source.
For the solutions from the original, RI, RD, and RS datasets, TPS (including crustal dust, vehicle emissions, and coal combustion) were similar to each other, implying that intra-day, diurnal or synoptic TS components had little influence on the TPS impact levels.The TSS (secondary formation and nitrate source) from the original, RI, and RD datasets obtained similar source impact levels; while TSS impact from the Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License.
Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License. to estimate the influences of inclusion/exclusion of a specific temporal component on the final apportionment results.
. The time series data of the PM and chemical species were then decomposed into multiple TS components using the KZ filter.Several new datasets were created by removing individual TS Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License.components from the original receptor datasets.ME-2 or PCA analysis was conducted on the original and new datasets to assess the impacts of excluding a specific TS component on the final source apportionment results.We aim to determine what processes/sources are responsible for the main variation characteristics and overall pollution levels in this specific dataset.We also aim to determine what the implications of our results are for source apportionment analyses conducted with data from different geographical locations and under various weather/climate conditions.Ambient particles were collected in Beijing from 22 July 2014 to 12 August 2014 at CRAES (Chinese Research Academy of Environmental Sciences) in this research.And concentrations of PM2.5, inorganic ions, OC/EC and heavy metals were measured by βray monitor, model ADI 2080 online analyzer (MARGA, Applikon Analytical B.V., The Netherlands), OC/EC analyzer (Sunset Laboratory Inc, USA) and the Xact 625 automated multi-metals monitor (Copper USA), respectively, at 1 h time resolution.
Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License.was5.7 ± 3.1 μg m -3 in this work.For SOC, the diurnal and synoptic TS components had larger amplitudes (FigureS2) and higher relative contributions to the total variance of PM2.5 (diurnal: 20%, synoptic: 62%).For species showing different TS component contributions, the cause was external influencing factors.For example, variability in primary species (such as Ca, EC) concentrations was mainly caused by local emission patterns and meteorological diffusion(van Pinxteren et al., 2009); secondary species For ions, elements, OC/EC and PM2.5, the larger gap in AAE value between the concentrations of RBL and original dataset means a larger difference between them, suggesting that the baseline TS component was the largest contributor to the average concentrations of PM2.5 and chemical species.The synoptic TS component also had a relatively high contribution to the average concentrations of NO3 -, SO4 2-, and NH4 + .The average concentrations of the three ions of the RS dataset were obviously lower and had large AAE values, compared with the results of the original dataset.For PM2.5 and seven species, the correlation coefficients for the original dataset and RI, RD, RS, Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License.and RBL datasets are displayed in Table S2.The lowest correlation coefficients were obtained for the RBL TS components and the original data.Overall, baseline TS components dominating the average concentrations of PM2.5 and chemical species might imply that pollutant emissions and other long-term fluctuation factors mainly determined the pollutants level in Beijing, from 22 July 2014 to 12 August 2014.When synoptic, diurnal, and intra-day TS components mainly influenced the variation of PM2.5 and chemical species, this suggests that the short-term fluctuation (e.g.noise, weather etc.) dominantly determined the variation of pollutants.
Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License.andnitrate variation trend was more obvious, so these two sources can be distinguished by ME-2.It can be confirmed that the correlation coefficient between NO3 -and SO4 2- was highest (0.86) for the synoptic TS component compared with other TS components (0.41, 0.28, and 0.82 for intra-day, diurnal, and baseline TS components, respectively).
5 and chemical species concentrations collected in Beijing into intra-day, diurnal, synoptic, and baseline temporal-scale (TS) components.This work investigated the factors driving these variations, and influencing factors were found to vary with species.The intra-day and diurnal TS components mainly influence the fluctuation of elements concentrations Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License.(e.g.Ca, Cr, Mn, etc); diurnal and synoptic TS components mainly impacted the fluctuation of PM2.5, NO3 -, EC, and OC concentrations; baseline and synoptic TS components were the main factors contributing to SO4 2-and NH4 + variance.For the PM2.5 and all chemical species concentration levels, the baseline TS component was the dominant factor.To study the influence of different TS components on the source impacts on PM2.5, four datasets (RI, RD, RS, and RB) were created by removing one individual TS component from the original dataset each time.The original and the four modified datasets were analyzed by ME-2 and/or PCA, and the source apportionment results Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-997Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 7 March 2018 c Author(s) 2018.CC BY 4.0 License.RS dataset was lower than other three results, suggesting that TSS was mainly influenced by the synoptic TS component and source emissions.Performance of four ME-2 runs was evaluated by analyzing the goodness of fit for the modeled and measured PM2.5 and chemical species mass concentrations (slope, r).Receptor data filtering intra-day TS components by KZ filter approach can improve the performance of the model and produce reasonable source impact results, suggesting that filtering noise from the instrument is useful to data analysis.The major findings of this work are that during the whole sampling period and pollution period, TPS impact levels were mainly influenced by source emissions, and TSS impact levels were mainly influenced by synoptic scale weather fluctuations and source emissions.The future work will focus on the mechanism through which synoptic scale weather disturbances modulate the secondary species and sources.

Figure 2 .
Figure 2. The influence of different TS components on the average concentrations of PM2.5 and

Table 2 .
Average source contributions to PM2.5 (μg m -3 ) estimated by ME-2 from Beijing for the 686 original, RI, RD, and RS datasets during the entire sampling period.
a TPS is the total contributions of crustal dust, vehicle emissions, and coal combustion.b TSS is the 688 total contributions of secondary formation and nitrate source.