Apportionment of Pm2.5 and Air Toxics in Detroit Michigan Final Report


Sonoma Technology, Inc.

1360 Redwood Way, Suite C
Petaluma, CA 94954-1169
707/665-9900
FAX 707/665-9800
www.sonomatech.com

APPORTIONMENT OF PM2 5 AND
AIR TOXICS IN DETROIT, MICHIGAN

FINAL REPORT
STI-906201.06-3103-FR

By:

Juli I. Rubin
Steven G. Brown
Katherine S. Wade
Hilary R. Hafner
Sonoma Technology, Inc.
1360 Redwood Way, Suite C
Petaluma, CA 94954-1169

Prepared for:

Dennis Doll
Ellen Bald ridge
U.S. Environmental Protection Agency
USEPA Mailroom, Mail Stop C304-04
Research Triangle Park, NC 27711

December 29, 2006

-------

-------
TABLE OF CONTENTS

Section	Page

LIST OF FIGURES	v

LIST OF TABLES	ix

EXECUTIVE SUMMARY	ES-1

1.	INTRODUCTION	1-1

2.	DATA AND METHODS	2-1

2.1	STN PM2.5 Data	2-1

2.1.1	Sampling and Analysis Details	2-1

2.1.2	Measurement Uncertainties	2-3

2.1.3	Data Used for Analysis and Source Apportionment	2-6

2.2	SANDWICH Data	2-8

2.3	PMF	2-10

2.3.1	Workings of PMF	2-10

2.3.2	Final Data Set Development	2-11

2.3.3	Using PMF Output	2-11

2.4	Wind and Potential Source Contribution Function Analysis	2-12

3.	RESULTS	3-1

3.1	Data Analysis	3-1

3.2	Source Apportionment of STN PM2.5 Data	3-8

3.3	Source Apportionment of SANDWICH Data	3-25

3.4	Source Apportionment of STN and Air Toxics Data	3-28

4.	SUMMARY AND CONCLUSIONS	4-1

4.1	New STN Uncertainties	4-1

4.2	PMF on STN and SANDWICH Data Sets	4-2

4.3	Linking PMF Factors to Sources	4-2

4.4	PMF Runs with PM2.5 Data and Air Toxics—Exploratory	4-3

5.	REFERENCES	5-1

iii

-------

-------
LIST OF FIGURES

Figure	Page

2-1. Detroit, Michigan, STN monitoring sites (2000 through 2005)	2-2

2-2. Time series of OC filter blank values at the three Detroit-area STN sites

(Allen Park, Luna Pier, and Dearborn) and the average values used for blank
correction	2-3

2-3. Potassium concentration versus uncertainty at all three Detroit STN sites

(July 2003 to October 2005)	2-4

2-4. Tin concentration versus uncertainty at all three Detroit STN sites

(July 2003 to October 2005)	2-5

2-5. Sodium ion concentration versus uncertainty at all three Detroit STN sites

(July 2003 to October 2005)	2-5

2-6.	Meteorological monitoring sites located near the three STN sites	2-13

3-1.	Luna Pier average ambient composition for STN and SANDWICH data

(May 2002 through 2005)	3-2

3-2. Allen Park average ambient composition for STN and SANDWICH data

(2000 through 2005)	3-3

3-3. Dearborn average ambient composition for STN and SANDWICH data

(May 2002 through December 2005)	3-3

3-4. Yearly ambient PM25 composition for STN data at Allen Park, Dearborn,

and Luna Pier	3-4

3-5. Seasonal ambient PM25 composition for STN data at Allen Park, Dearborn,

and Luna Pier	3-5

3-6. Yearly winter ambient PM25 composition for STN data at Allen Park,

Dearborn and Luna Pier	3-6

3-7. Yearly summer ambient PM25 composition for STN data at Allen Park,

Dearborn, and Luna Pier	3-7

3-8. Scatter plot by season of manganese, zinc, nickel, and chromium at Allen Park

(2000 through 2005)	3-7

3-9. Scatter plot by season of silicon and calcium at Luna Pier and Allen Park	3-8

3-10. Luna Pier PMF factor profiles and time series for STN data (May 2002

through December 2005)	3-10

v

-------
3-11. Allen Park PMF factor profiles and time series for STN data

(2000 through 2005)	3-11

3-12. Dearborn PMF factor profiles and time series for STN data

(May 2002 through December 2005)	3-12

3-13. Average composition of 8-factor PMF results at Luna Pier for STN data

(May 2002 through December 2005)	3-13

3-14. Average composition of 9-factor PMF results at Allen Park for STN data

(2000 through 2005)	3-13

3-15. Average composition of 10-factor PMF results at Dearborn for STN data

(May 2002 through December 2005)	3-14

3-16. Yearly and seasonal trends in PMF mass composition at Luna Pier for

STN data (May 2002 through December 2005)	3-15

3-17. Yearly and seasonal trends in PMF mass composition at Allen Park for

STN data (2000 through 2005)	3-16

3-18. Yearly and seasonal trends in PMF mass composition at Dearborn for

STN data (May 2002 through December 2005)	3-17

3-19. Proportional PM2.5 point source emissions map	3-19

3-20. Proportional zinc and manganese point source emissions map	3-20

3-21. Proportional chromium and nickel point source emissions map	3-21

3-22. Proportional copper point source emissions map	3-22

3-23. Wind roses for three STN sites: Allen Park, Dearborn, and Luna Pier	3-23

3-24. Wind roses of zinc/manganese, chromium/nickel, and copper on high days

with point source emissions	3-23

3-25. Wind roses of calcium on high days with point source emissions at

Luna Pier	3-24

3-26. Wind roses on high mixed industrial factor days with PM2.5 point source

emissions at Luna Pier	3-25

3-27. PMF results and ambient mass composition for both STN and

SANDWICH data sets at Luna Pier (May 2002 through December 2005)	3-26

3-28. PMF results and ambient mass composition for both STN and

SANDWICH data sets at Allen Park (2000 through 2005)	3-27

vi

-------
3-29. PMF results and ambient mass composition for both STN and

SANDWICH data sets at Dearborn (May 2002 through December 2005)	3-28

3-30. Comparison of average PMF mass contribution at Allen Park for a 9-factor

565 sample run and a 9-factor 158 sample run	3-29

vii

-------

-------
LIST OF TABLES

Table	Page

2-1. Summary of uncertainty methods by species and values used to estimate

uncertainties	2-7

2-2.	SNRs and percent below detection for species measured at three STN sites

(Allen Park, Dearborn, and Luna Pier)	2-9

3-1.	Number of factors used for final PMF runs at the three STN sites for both

STN data and SANDWICH data	3-9

4-1.	Comparison of PMF results using recently updated STN uncertainties and

PMF results from previous studies at Allen Park	4-2

IX

-------

-------
EXECUTIVE SUMMARY

Source apportionment of speciated PM2.5 and air toxics data using positive matrix
factorization (PMF) methods was applied to three Speciation Trends Network (STN) sites in
Detroit, Michigan. The goals of the analysis were threefold: (1) to determine the effects of using
recently updated STN uncertainties in PMF, (2) to determine the effects of using SANDWICH
data on PMF, and (3) to explore PMF applications to combined PM2.5 and air toxics data sets.
The three selected STN sites—Luna Pier, Allen Park, and Dearborn—had 220, 565, and 190
samples dating from May 2002 to December 2005, December 2000 to December 2005, and May
2002 to December 2005, respectively. PMF was applied at each site with (1) STN data and
recently updated uncertainties, (2) SANDWICH data, and (3) a combination of STN data and air
toxics data. Additional data analysis techniques were applied to both STN and SANDWICH
data sets including analysis of ambient data composition, yearly and seasonal trends, and species
correlations to help evaluate PMF results.

The majority of the PM2.5 mass was apportioned to ammonium sulfate, ammonium
nitrate, and mobile sources at all three sites. The PMF results using recently updated
uncertainties were compared with previous PMF efforts at Allen Park; results were similar across
studies with the exception that previous PMF studies at Allen Park were able to separate out a
diesel component from the mobile source factor. Differences are likely due to the larger relative
uncertainties applied to metal species (i.e., the new uncertainties) without any changes to the
uncertainties applied to the carbon species in the current study.

PMF runs conducted using STN and SANDWICH data sets produced similar factors with
similar mass apportionment, on average. Major differences included larger mass apportioned to
ammonium sulfate at the expense of ammonium nitrate when using the SANDWICH data.

These results are as expected because the SANDWICH data set has higher sulfate concentrations
and lower nitrate concentrations than the STN data set.

PMF runs conducted with combined PM2.5 and air toxics data did not provide additional
insight into mobile or other source contributions. This is likely due to the limited amount of
collocated PM2.5 and air toxics data, making it difficult to produce meaningful results; with a
larger data set results may be improved.

ES-1

-------

-------
1. INTRODUCTION

The U.S. Environmental Protection Agency (EPA) is interested in analyzing the
relationships among pollutants and their precursor species. Increased awareness of interactions
among secondary pollutants and the need to identify common culpable sources to assist with
developing cost effective and efficient control strategies make it increasingly important that
viable tools be available for attributing pollution sources to ambient measurement data.

Air pollution is composed of chemical species originating from natural and manmade
emissions that can be transported from their original source areas. Sources typically emit a
variety of pollutants, so efficient emission control strategies are needed to address multiple
pollutants to bring air pollution concentrations below mandated health standards (e.g., 15 |ag/m3
annual average for PM2.5 and 0.080 ppm for 8-hr ozone). These multiple pollutant control
strategies depend on the ability to determine the relationships between emissions sources and
elevated levels of air pollution observed at ambient monitoring sites. Source attribution needs to
be performed across the range of pollutants to identify common sources among pollutants.
Findings will support control strategy development.

The purpose of this work assignment (WA) was to perform source apportionment of
multiple pollutants in selected cities using positive matrix factorization (PMF). This WA builds
on past source apportionment efforts applied to speciated PM2.5 data alone. Results are intended
to help the EPA better understand the common sources of PM2.5, ozone precursors, and air
toxics. Results will also be useful to other ongoing investigations aimed at understanding the
links among pollutants.

1-1

-------

-------
2. DATA AND METHODS

In this WA, PM2.5 data from the Speciation Trends Network (STN) for sites in Detroit,
Michigan, were combined with collocated air toxics data. This section summarizes the available
data, uncertainty values and their quantification, the species selected for use in PMF, an
overview of PMF, and discussion of a post-analysis technique called Potential Source
Contribution Function (PSCF).

2.1 STN PM2.5 DATA

2.1.1 Sampling and Analysis Details

Integrated 24-hr PM2.5 samples were collected as part of the STN at three sites in the
Detroit, Michigan, area: Allen Park (December 2000-December 2005, 565 samples, l-in-3 day);
Dearborn (May 2002-December 2005, 190, l-in-6 day); and Luna Pier (May 2002-December
2005, 220 samples, l-in-6 day). A map of the area is shown in Figure 2-1. Samples were
collected by the STN using the MetOne Spiral Aerosol Speciation Samplers (SASS) and were
taken on a l-in-3 or l-in-6 day schedule. Field blanks were collected for l-in-10 routine
samples, and trip blanks were collected for l-in-30 routine samples (Research Triangle Institute,
2004). Blank correction for organic carbon (OC) is important for STN data (Kim et al., 2005c;
Subramanian et al., 2004; Rice, 2004) and the average blank value over all blanks at each site
was used to blank correct the OC concentrations. Only small seasonal variations in the blank
values were observed, and with a small set of blanks, applying a seasonal blank correction would
introduce additional bias and artificial trends in the ambient data; thus, only the average over the
entire period was used. Figure 2-2 shows the trends in OC blanks by site; the values used were
0.99, 1.32, and 1.16 (J,g/m3 for Allen Park, Dearborn, and Luna Pier, respectively.

STN PM2.5 samples were collected on Teflon, nylon, and quartz filters. The Teflon filter
was used for mass concentrations and analyzed via x-ray fluorescence (XRF) for the elements.
The nylon filter was analyzed for the ions sulfate, nitrate, ammonium, sodium, and potassium via
ion chromatography (IC). The quartz filter was analyzed by the Research Triangle Institute
(RTI) via the National Institute for Occupational Safety and Health/Thermal Optical
Transmittance (NIOSH/TOT) protocol (NIOSH, 1999; Birch and Carey, 1996) for OC and
elemental carbon (EC).

2-1

-------
BlSktOBAff'TERN h TIONA L

.Detroit-Metro JA re.

\imM\

Arm y&borJ

^OEtmi 7 fVlETRQE©&$\ NWMyf\'E^^4

Legend

S3 Monitoring Sites

Luna Pier?

Jo I Rio.

Figure 2-1. Detroit, Michigan, STN monitoring sites (2000 through 2005).

2-2

-------
Allen Park Average = 0.99 +/- 0.44
Luna Pier Average = 1.16 +/- 0.47
Dearborn Average = 1.32 +/- 0.45

A * ¦ •

A. . >

: . *
^ .

. a

A .

~ Allen Park

Allen Park Average

A Luna Pier

Luna Pier Average

¦ Dearborn
Dearborn Average

Figure 2-2. Time series of OC filter blank values at the three Detroit-area STN
sites (Allen Park, Luna Pier, and Dearborn) and the average values used for blank
correction.

2.1.2 Measurement Uncertainties

In addition to concentrations, analysts need to understand the associated uncertainties.
Uncertainties reported by RTI for speciated PM2.5 data are currently being updated to ensure
consistency among the estimation methods used by the laboratories in STN. At present, the
EPA's Air Quality System (AQS) only reports updated uncertainties for data measured between
July 2003 and October 2005. It was, therefore, necessary to use the recently updated
uncertainties to extrapolate uncertainties for the remaining samples without updated values.

For species with updated uncertainties, relationships were examined between
concentration and uncertainty. Measurements from the three selected Detroit sites (Allen Park,
Dearborn, and Luna Pier) were combined because relationships between concentration and
uncertainty were found to overlap. The expected results of such plots are that a constant
uncertainty for concentrations at or below the detection limit will be observed because the
uncertainty is absolute. Above the detection limit, it is expected that an increase in uncertainty
with concentration will be found because a relative component of the uncertainty is introduced.
These patterns were observed for species such as potassium (Figure 2-3); however, some of the
speciated metals did not follow the expected trend. No consistent relationship was found for
selected speciated metals including titanium, vanadium, tin, and selenium (e.g., tin is shown in

2-3

-------
Figure 2-4). Additionally, some metal species such as manganese, arsenic, lead, and bromine
appeared to have two separate trend lines (e.g., sodium ion is shown in Figure 2-5). For species
that showed two trends, the data were examined to determine if the trends were caused by
different measurements, time periods, or other patterns. No identifiable patterns were found. In
discussions with RTI, it was noted that samples were analyzed by different XRF instruments
with different uncertainty relationships. However, the data are reported to AQS with one method
code; AQS does not have identifiers at the instrument level.

0.012

0.01

1 0.008
"Si

£ 0.006

S

"3

u

a 0.004
s

0.002

0

Figure 2-3. Potassium concentration versus uncertainty at all three Detroit STN
sites (July 2003 to October 2005).

0	0.02 0.04 0.06 0.08 0.1 0.12 0.14

mass (ug/m3)

2-4

-------
0.014
0.012
0.01
0.008

~ ~ ~

/~

~ «
~

~ ~~
~

1 0.006

» ~ ~
fN* *

~

~~

0.004

0.002

0.01	0.02	0.03	0.04

mass (ug/m3)

0.05

0.06

Figure 2-4. Tin concentration versus uncertainty at all three Detroit STN sites
(July 2003 to October 2005).



WW*



4** ~ *

0.05 0.1 0.15 0.2
mass (ug/m3)

0.25

0.3

0.35

Figure 2-5. Sodium ion concentration versus uncertainty at all three Detroit STN
sites (July 2003 to October 2005).

2-5

-------
Different methods for estimating uncertainties were applied to address the different
relationships found for the speciated data. For species that followed the expected curve, as
shown for potassium (Figure 2-3), uncertainties were assumed to be constant below a determined
cutoff point. Above the cutoff point, a linear regression was used to relate uncertainty to
concentration. For the remaining species, the uncertainty was set to the maximum reported
uncertainty (excluding outliers) because it is always best to assume a larger uncertainty for
source apportionment purposes (Hafner, 2005). A species and uncertainty method summary is
provided in Table 2-1.

The uncertainty development methods described could not be used for the carbon and ion
species because the uncertainties for these species were not updated in AQS; only species
measured by XRF were updated. As a result, the root median squared percent error (RMSPE)
(Equation 2-1) was used to estimate the uncertainties using collocated measurements. This
method is similar to the more common root mean squared error (RMSE) with two exceptions:
(1) the error is a percent of the original value instead of an absolute error, and (2) the median of
the squared errors is used instead of the mean. Using the percent enables the output to be applied
directly to the concentrations to determine sample-specific uncertainties. Using the median
excludes outlying values that heavily influence an RMSE but are not considered useful for PMF.

_ _

RMSPE =

median

2(xi -x2)

_ Ol +*2) _

where x and >' are collocated measurements.

2.1.3 Data Used for Analysis and Source Apportionment

Data from the STN program are routinely screened and validated before being made
publicly available. Additional quality control (QC) checks were performed on the data prior to
source apportionment, including comparison of reconstructed fine mass to measured mass and
comparison of XRF sulfur to IC sulfate. Approximately 10% of samples from each site did not
pass these checks and were excluded from the PMF analysis. Those samples for which all
species were missing were also excluded from the analysis.

2-6

-------
Table 2-1. Summary of uncertainty methods by species and values used to estimate
uncertainties. Values include maximum value reported (excluding outliers), the
regression cutoff point, pre-cutoff uncertainty, slope and intercept for calculating post-
cutoff uncertainties, and RMSPE values.

Species

Uncertainty
Method

Maximum
Value
(Ug/m3)

Regression
Cutoff
(Ug/m3)

Precutoff
Uncertainty
(Ug/m3)

Slope

Intercept
(Ug/m3)

RMSPE
(%)

Aluminum PM2 5

max value

0.0160

--

--

--

--

--

Antimony PM2 5

max value

0.0325

--

--

--

--

--

Arsenic PM2 5

max value

0.0028

--

--

--

--

--

Barium PM2 5

max value

0.0733

--

--

--

--

--

Bromine PM2 5

max value

0.0017

--

--

--

--

--

Cadmium PM2 5

max value

0.0142

--

--

--

--

--

Cerium PM2 5

max value

0.1083

--

--

--

--

--

Lead PM2 5

max value

0.0044

--

--

--

--

--

Magnesium PM2 5

max value

0.0375

--

--

--

--

--

Manganese PM2 5

max value

0.0017

--

--

--

--

--

Samarium PM2 5

max value

0.0077

--

--

--

--

--

Selenium PM2 5

max value

0.0022

--

--

--

--

--

Strontium PM2 5

max value

0.0032

--

--

--

--

--

Terbium PM2 5

max value

0.0092

--

--

--

--

--

Tin PM2 5

max value

0.0120

--

--

--

--

--

Titanium PM2 5

max value

0.0026

--

--

--

--

--

Tungsten PM2 5

max value

0.0183

--

--

--

--

--

Vanadium PM2 5

max value

0.0018

--

--

--

--

--

Calcium PM2 5

trend lines

--

0.0300

0.0026

0.0684

0.0009

--

Chlorine PM2 5

trend lines

--

0.0060

0.0036

0.0555

0.0043

--

Chromium PM2 5

trend lines

--

0.0040

0.0008

0.0629

0.0006

--

Copper PM2 5

trend lines

--

0.0080

0.0008

0.0652

0.0004

--

Iron PM2 5

trend lines

--

0.0066

0.0009

0.0690

0.0005

--

Nickel PM2 5

trend lines

--

0.0045

0.0006

0.1699

-0.0003

--

Potassium PM2 5

trend lines

--

0.0400

0.0036

0.0695

0.0010

--

Silicon PM2 5

trend lines

--

0.0600

0.0030

0.0791

0.0021

--

Sulfur PM2 5

trend lines

--

0

0

0.0711

0.0009

--

Zinc PM2 5

trend lines

--

0.0095

0.0009

0.0691

0.0003

--

Ammonium Ion PM2 5

RMSPE

--

—

--

--

--

7

EC PM2 5

RMSPE

--

—

--

--

--

10

OC PM2 5

RMSPE

--

—

--

--

--

33

Potassium Ion PM2 5

RMSPE

--

--

--

--

--

0.7

Sulfate PM2 5

RMSPE

--

—

--

--

--

10

Total Nitrate PM2 5

RMSPE

"

--

--

--

--

8

RMSPE = root median squared percent error

2-7

-------
To determine which species would be included as variables in the PMF analysis, two
criteria were used: signal-to-noise ratio (SNR) and percent below detection limit (BDL). SNR
can be used as a criterion to determine the "strong", "weak", and "bad" variables (Paatero et al.,
2003). SNR is automatically calculated by EPA PMF 1.1 as

SNR, =1/2

(2-2)

z=l

where xtj is the concentration for species j on day i and .v,7 is the uncertainty. In general, the
species j is defined to be "strong" if SNR > 2, "weak" if 0.2 < SNR < 2, and "bad" if SNR < 0.2
(Paatero, Hopke 2003).

Table 2-2 summarizes the PM25 measurements with the SNR and percent BDL values
for the species with at least 10% of the data above detection for each site. In general, species
having more than 70% BDL were discarded. Weak variables that have less than 70% BDL were
included in the analysis, but those variables were down-weighted by a factor of three in PMF
calculations. No sodium or chlorine data were used in this analysis because the impact of sea
and road salt on PM25 and air toxics was expected to be small and confidence (or SNR) in the
sodium and chlorine data was often low. Because potassium ion (K+) was mostly BDL (over
70%) of the time), elemental potassium, which was mostly above detection, was used.

2.2 SANDWICH DATA

While STN measures PM25 mass and the species that comprise the mass, the
measurements are often slightly different than the Federal Reference Method (FRM) PM25 mass
measurements, which are the metric for regulations. To translate the STN measurements into
"FRM equivalent" measurements, the Sulfate, Adjusted Nitrate, Derived Water, Inferred
Carbonaceous mass and estimated aerosol acidity (H+) material balance approach
(SANDWICH) was developed (Frank, 2006). This method assumes PM2 5 on the filter is broken
down as follows:

PM2.5 = N03(FRM) + S04 + NH4 + H20 + Crustal + TCM + Blank + Other (2-3)

The measurement of total carbonaceous material (EC and OC) has a higher analytical
uncertainty than the other components of PM2 5; therefore, the SANDWICH method uses the
other components to calculate EC and OC. This method should eliminate blank corrections and
artifacts on the filters. In addition, N03(frm) (nitrate measured by the FRM) is the retained NO3
predicted using N03(stn) (nitrate measured by the STN) and temperature and relative humidity.
The result of these adjustments is usually a higher sulfate mass and a lower nitrate mass overall.

To compare analyses based on the STN measurements and ensure that results are
applicable in a regulatory sense, SANDWICH data were also examined as part of this project.
This is the first known application of PMF to SANDWICH data.

2-8

-------
Table 2-2. SNRs and percent below detection for species measured at three STN
sites (Allen Park, Dearborn, and Luna Pier). For SNR, values in bold are
considered "strong"; the remaining values are considered "weak". Percent below
detection values in bold have greater than 70% BDL.

Species

Signal-to-Noise Ratio (SNR)

% Below Detection Limit (BDL)

Allen Park

Dearborn

Luna Pier

Allen Park

Dearborn

Luna Pier

Arsenic PM2.5

__

0.67

0.49

__

33.58

67.73

Aluminum PM25

0.83

2.32

0.94

71.63

21.02

65.91

Antimony PM2 5

—

0.28

—

—

90.19

—

Bromine PM25

0.97

1.14

0.90

44.17

21.25

39.09

Calcium PM25

3.47

2.71

5.72

1.44

7.72

3.18

Cadmium PM25

—

0.29

—

—

55.23

—

Cerium PM2 5

—

0.27

—

—

22.55

—

Chromium PM2 5

1.57

1.84

1.44

68.94

14.53

67.27

Copper PM2 5

3.33

1.58

1.59

28.01

11.12

54.09

Chlorine PM2 5

3.81

3.83

1.50

64.99

9.06

67.27

Iron PM2 5

3.65

2.58

6.75

0.72

4.50

0.91

Lead PM25

0.59

1.42

0.48

68.94

18.09

61.82

Magnesium PM2 5

—

1.23

—

—

22.42

—

Manganese PM25

1.16

3.93

0.63

45.60

6.72

56.82

Nickel PM25

0.98

1.01

0.86

71.81

19.03

70.91

Samarium PM2 5

—

0.70

—

—

18.59

—

Selenium PM2 5

0.50

0.71

0.86

83.30

30.34

67.73

Strontium PM2 5

—

0.51

—

—

30.15

—

Terbium PM2 5

—

0.77

—

—

50.99

—

Tin PM2 5

0.36

0.32

—

93.90

64.16

—

Titanium PM2 5

0.88

1.59

0.64

52.42

19.26

67.27

Tungsten PM2 5

—

0.25

—

—

77.89

—

Vanadium PM2 5

0.51

0.78

—

84.56

26.24

—

Silicon PM2 5

2.85

2.32

3.93

10.23

9.40

15.45

Zinc PM2.5

4.07

3.19

5.55

2.51

7.34

4.09

Sulfur PM2.5

4.36

2.52

6.98

1.08

7.38

0.45

Potassium PM2 5

1.46

2.41

1.13

1.62

7.59

2.27

Sodium PM2 5

__

—

—

__

—

75.00

Ammonium Ion PM2 5

2.45

16.65

2.98

1.62

3.10

0.45

Sodium Ion PM2 5

2.43

1.18

1.15

16.34

41.24

18.18

Potassium Ion PM2 5

1.14

3.01

1.59

57.81

10.54

63.64

oc pm2.5

1.66

5.16

1.60

1.26

10.08

0.00

Total Nitrate PM2 5

3.00

12.50

3.64

0.90

4.08

0.00

EC PM2.5

1.54

4.56

1.71

3.95

11.05

13.64

Sulfate PM25

2.86

16.66

3.62

0.18

3.10

0.00

2-9

-------
2.3 PMF

2.3.1 Workings of PMF

PMF, described in detail elsewhere (Paatero, 1997; Paatero and Tapper, 1994) and briefly
covered here, was used in this WA. PMF is an advanced multivariate receptor modeling
technique that calculates site-specific source profiles with time variations of these sources based
on correlations imbedded in ambient data. PMF has been successfully applied to PM data, air
toxics data, and volatile organic compound (VOC) data in several studies (Anttila et al., 1995;
Begum et al., 2005; Buzcu et al., 2003; Juntto and Paatero, 1994; Kim et al., 2003; Kim et al.,
2004a; Kim et al., 2004b; Kim and Hopke, 2004a, b; Kim et al., 2005a; Kim et al., 2005b; Larsen
and Baker, 2003; Lee et al., 1999; Poirot et al., 2001; Polissar et al., 2001; Ramadan et al., 2000;
Zhou et al., 2004).

Given a data matrix X consisting of the concentration measurements of n chemical
species in m samples and their corresponding uncertainties, the objective of PMF is to determine
the number of factors p, the chemical composition profile of each factor, and the contribution of
each factor to each sample. PMF factorizes the data matrix X into two matrices according to

~X(m by n) G(m by p) F(p by n) E(m by n)	(^ "4 )

where G represents the contribution of each factor to each ambient sample, and describes the
time variations of the factors because each ambient sample is an observation at different times.
F is a matrix of chemical composition profiles of each factor. F and G are both forced to be non-
negative in order to make physical sense (i.e., factors cannot have negative species
concentrations and ambient samples cannot have a negative factor contribution). E is an m by n
residual matrix of random errors. The elements of the residual matrix, e^, can be defined as

p

ev=XV~YuSrkfk1	(2"5)

k=1

where i = 1,... ,m j = 1,...,«). In PMF, the sum of the squares of residuals, elh weighted
inversely by the variation of the data points, Sjf, is minimized according to the following
constrained weighted least-squares model:

9

n m

minimize Q = ^ J] ¦(2-6)

2=1 j=1 Sij

The objective is to determine the matrices G and F by minimizing Q. Equation 2-6 is
solved using a unique iterative algorithm in which matrices G and F vary simultaneously at each
iteration step (Paatero, 1997). Theoretically, if the uncertainties correctly characterize the data
and every point is perfectly modeled, Q should be approximately the number of species
multiplied by the number of observations, minus the number of factors multiplied by the number
of species (i.e., the number of data points). In these analyses, Q from the modeling was required
to be within 50% of the theoretical Q to ensure a reasonable fit for all observations.

2-10

-------
The EPA's Office of Research and Development has recently developed a standalone
version of PMF (EPA PMF) that has been freely distributed to the air quality management
community (Eberly, 2005). EPA PMF version 1.1 is a graphical user interface that has been
developed based on the PMF model and solved using the multi-linear engine as implemented in
the program ME-2 (Paatero, 1999). EPA PMF operates in a robust mode, meaning that
"outliers" are not allowed to overly influence the fitting of the contributions and profiles.

2.3.2	Final Data Set Development

In PMF, each data point is weighed individually, allowing the user to adjust the influence
of each point depending on the confidence in the data. This feature is an advantage of PMF
because samples with some species missing or below the minimum detection limit (MDL) can be
used in the analysis, with associated uncertainty adjusted so that these data points are given less
weight in the model solution. Data below MDL were substituted with the maximum MDL
reported for the given species divided by two and missing data were substituted with median
concentrations (Poirot et al., 2001; Polissar et al., 2001; Song et al., 2001). The maximum MDL
was used for substitution because use of the sample-specific MDL could introduce a false source
of variability in the data. Uncertainties for values below MDL were calculated as 5/6*max MDL
and for missing values as 4*median concentration. For samples above detection, updated
uncertainties were used, if reported, and the remaining uncertainties were estimated based on
methods discussed in Section 2.1.2. Some species were given less or more weight by increasing
or decreasing their uncertainty, which resulted in better modeling of individual species and the
total mass.

2.3.3	Using PMF Output

Source contributions of each factor can be determined using PMF output and matrices G
and F. Each element of the matrix G (,) is a normalized source contribution of a factor k for a
given sample i. Each element of F (f:i) is a mass contribution of a chemical species j to a
factor k. Using matrix F, the reconstructed mass (mk) for an individual factor i can be calculated
as

«•„ = L/„	<2-7)

;=i

The normalized source contributions of matrix G can then be converted to meaningful mass units
by multiplying the individual elements of G and the corresponding factor mass

Crk=8rk*mk	(2"8)

Uncertainties in the EPA PMF solution are estimated using a bootstrapping technique,
which is a re-sampling method in which "new" data sets are generated that are consistent with
original data, each data set is decomposed into F and G matrices, and the resulting F and G
matrices are compared with the base run (Eberly, 2005). Instead of inspecting point estimates,
this method allows the analyst to review the confidence intervals for each species to obtain more

2-11

-------
robust profiles. Output of the bootstrapping analysis consists of box whisker plots of species for
each profile by both percent of species and concentration. The box shows where the middle
50% of the bootstrap values exist—the tighter the box, the more certainty in the profile, and the
more consistent the results are across the bootstraps. In this study, 200 bootstrap runs were
performed for the final analysis results.

2.4 WIND AND POTENTIAL SOURCE CONTRIBUTION FUNCTION ANALYSIS

A Potential Source Contribution Function (PSCF) (Draxler and Hess, 1997) was applied
to help interpret the PMF results. The transport patterns on days with the highest 10%
concentration of a given factor were compared with the climatological transport patterns. This
comparison highlights the differences in transport and areas of influence between the general
transport pattern (i.e., the climatology) and high concentration days of a given factor. Using the
National Oceanic and Atmospheric Administration (NOAA) Hybrid Single-Particle Lagrangian
Integrated Trajectory (HYSPLIT) model, (Draxler and Hess, 1997) 72-hr backward trajectories
were run for all sample dates at three ending heights (100, 300, and 1,000 meters). Ensemble
backward trajectories were run every six hours to account for variability over a 24-hr sampling
period. If a trajectory endpoint of the air parcel lies in the if1 cell, the trajectory is assumed to
collect PM2.5 emitted in the cell. Once the PM2.5 is incorporated into the air parcel, it is assumed
to be transported along the trajectory to the monitoring site. PSCF,( is the conditional probability
that an air parcel that passed through the if1 cell had a high concentration upon arrival at the
monitoring site defined as

PSCF. = (2-9)

where n,, is the total number of endpoints that fall in the if1 cell and m,, is the number of
endpoints in the same cell that are associated with samples that exceeded the threshold criterion.
In this study, the average contribution of each source was used as the threshold criterion. The
sources are likely to be located in the areas that have high PSCF values (Draxler and Hess,
1997). Emissions data, including point sources and fire locations, were overlaid on the PSCF
maps to identify specific emissions sources in likely source areas.

While trajectories provide useful information on regional transport, wind roses can
provide insight into local transport. For the 10% of sample days with the highest concentrations,
wind roses were examined for selected species. Hourly wind data were not available at the STN
monitoring sites; as a result, wind data were used from nearby locations. A map of the study
sites and the locations from which wind data were obtained is found in Figure 2-6. Site KDTW
was used for Allen Park and Dearborn wind roses and site KDUH was used for Luna Pier. The
wind roses were combined with point source emissions data to understand the link between PMF
factors and specific sources.

2-12

-------
fester,



Detr.o it. Met i.o. A re a

TDETR'OlT MEJJttzmiiTti N WAYNE^Ul

KONZ

flvecK -,y-f
J KjlKal

•w^djgsjd '¦

Legend

O Wind Sites
S3 Monitoring Sites

KTTF

[Toledo;

Figure 2-6. Meteorological monitoring sites located near the three STN sites.

2-13

-------

-------
3. RESULTS

This section includes summaries of the results of the ambient data analysis and the PMF
runs for both STN and SANDWICH PM2.5 data sets, and analysis of trajectories and wind as
discussed in Section 2. Additional exploratory results are presented for PMF runs conducted
with both STN PM2.5 data and air toxics data.

3.1 DATA ANALYSIS

PMF should be considered one of a set of complementary analyses that help analysts
understand and quantify source contributions to ambient concentrations. Prior to running PMF,
it is essential to analyze ambient data to better understand the results of the source
apportionment. These analyses can provide information on which factors are expected and can
provide checks for determining whether results are sensible. Consequently, the composition of
the ambient data and the correlations between various species were examined.

The average ambient PM2.5 composition of data at Luna Pier, Allen Park, and Dearborn
are displayed in pie charts in Figures 3-1 through 3-3. Each figure includes two pie charts, one
with STN data and one with SANDWICH data. Allen Park and Dearborn have very similar
ambient PM2.5 compositions, with Dearborn having a slightly higher sulfate concentration and
Allen Park having more organic mass (OM= blank corrected OC* 1.8). At Dearborn, a larger
fraction of the mass is unaccounted for (grey portion of the pie) and there is more soil mass.
When comparing the STN and SANDWICH compositions, a larger fraction of the mass is
associated with ammonium sulfate with SANDWICH data and there is a corresponding decrease
in ammonium nitrate. There is also a decrease in the organic mass portion with SANDWICH
data relative to STN.

Yearly and seasonal (quarterly) trends in the ambient data were examined for significant
differences year to year or season to season. Yearly trends by season were also reviewed.

Yearly trends in ambient composition for all three sites are shown in Figure 3-4 and at all sites
there is a clear decrease in PM2.5 mass in 2004. Seasonal comparisons (Figure 3-5) showed a
shift in composition from higher ammonium sulfate in the summer to higher ammonium nitrate
in the winter, as expected. An increase in OM was also seen in the winter months, but overall,
there is a decrease in total PM2.5 mass. Yearly comparisons of winter season ambient data
(Figure 3-6) again show decreased PM2.5 mass in 2004, however, at Dearborn low mass is also
observed in 2005. The higher 2003 mass has a larger organic mass compared with the two lower
years. The winter data from 2002 were excluded for both Dearborn and Luna Pier because of
insufficient data. Yearly comparisons of summer season ambient data (Figure 3-7) show
decreased PM2.5 mass in 2004 caused by a decrease in ammonium sulfate relative to other years.

Select scatter plots by season are shown in Figures 3-8 and 3-9. Zinc and manganese
were found to be correlated at both Allen Park and Dearborn and nickel/chromium correlations
were found at all three sites. As a result, we expect to find these species grouped in the same
PMF factors. Another correlation found was between silicon and calcium, which is often
associated with a soil source. However, as shown in Figure 3-9, the plot between silicon and

3-1

-------
calcium is more scattered at Luna Pier than at the other two sites; the scatter indicates a second
source of calcium at Luna Pier.

0.21,1%-.

3.27,22%	1.82,12%

STNIIata	SANDWICH Data

~ Ammonium Sulfate ¦ Ammonium Nitrate ¦ OM (blank coll ected OC x 1.8) ¦ Soil ¦ EC
¦ Other Species ~ Difference in sum of species and filter mass

Figure 3-1. Luna Pier average ambient composition for STN and SANDWICH
data (May 2002 through December 2005).

3-2

-------
SANDWICH Data

~ Ammonium Sulfate ¦ Ammonium Nitrate ¦ OM (blank collected OC xl.8) ¦ Soil ¦ EC
¦ Other Species ~ Difference in sum of species and filter mass

Figure 3-2. Allen Park average ambient composition for STN and SANDWICH
data (December 2000 through December 2005).

~ Ammonium Sulfate ~ Ammonium Nitrate ~ OM (blank coll ected OC xl.8) ¦ Soil ¦ EC
¦ Other Species ~ Difference in sum of species and filter mass

Figure 3-3. Dearborn average ambient composition for STN and SANDWICH
data (May 2002 through December 2005).

7.91,36%

2.47,11%

4.58,21%

0.30,1 %
1.04,5%

0.97,4%

4.97,22%

3-3

-------


25

20

15



S 10

Allen Park

Dearborn

0

o

Luna Pier

¦	Other Species

¦	Soil

¦	EC

¦	Organic Mass (OM)

¦	Ammonium Nitrate
~ Ammonium Sulfate

2002 2003 2004 2005 2002* 2003 2004 2005 2002* 2003 2004 2005

Figure 3-4. Yearly ambient PM2.5 composition for STN data at Allen Park,
Dearborn, and Luna Pier. An asterisk (*) indicates incomplete data collected in
winter months.

3-4

-------
Allen Park





Dearborn

0

Q

0

0

Luna Pier





¦	Other Species

¦	Soil

¦	EC

¦	Organic Mass (OM)

¦	Ammonium Nitrate
~ Ammonium Sulfate

^ ^ ^ ^
S>	jv .ov ^	,os xy

Figure 3-5. Seasonal ambient PM2.5 composition for STN data at Allen Park,
Dearborn, and Luna Pier (Allen Park: December 2000 through December 2005,
Dearborn and Luna Pier: May 2002 through December 2005).

3-5

-------
Dearborn*

0

2 B

Luna Pier*

2002 2003 2004 2005 2003 2004 2005 2003 2004 2005

¦	Other Species

¦	Soil

¦	EC

¦	Organic Mass (OM)

¦	Ammonium Nitrate
~ Ammonium Sulfate

Figure 3-6. Yearly winter ambient PM2.5 composition for STN data at Allen Park,
Dearborn, and Luna Pier. An asterisk (*) indicates incomplete data were
collected in winter months.

3-6

-------
30

25

20



OX

0 15

10

Allen Park

Dearborn

o

I

0

g

Luna Pier

¦	Other Species

¦	Soil

¦	EC

¦	Organic Mass (OM)

¦	Ammonium Nitrate
~ Ammonium Sulfate

2002 2003 2004 2005 2002 2003 2004 2005 2002 2003 2004 2005

Figure 3-7. Yearly summer ambient PM2.5 composition for STN data at Allen
Park, Dearborn, and Luna Pier.

D.03

D.00

E

0.00 0 01 0.02 Hi.03 0.04 0.05 0.06 0.07 0.C8
Chromium (ug/m3)

0.03

0.02

u
c
a
ro
c

0.01

0.00

O.OO 0.C5 0.10 0.15
Znc (ig/m3)

D.20

SEASON

o fall
x spring
+ summer
a winter

Figure 3-8. Scatter plot by season of (a) nickel and chromium and (b) manganese
and zinc at Allen Park (December 2000 through December 2005).

3-7

-------
1.00

0.1 0.2 0.3
Silicon (ug/m3)

0.75

ra
3

ro
O

0.50

0.25

0.00

—i 1 1 r

Allen Park

"i r

i i i i i i i i

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Silicon (ug/m3)

SEASON

o fall
x spring

summer
a winter

Figure 3-9. Scatter plot by season of silicon and calcium at Luna Pier and Allen
Park (Luna Pier: May 2002 through December 2005, Allen Park: December
2000 through December 2005). Lines indicate edges or general trends.

3.2 SOURCE APPORTIONMENT OF STN PM2 5 DATA

For all three sites, PMF was run using 7 to 11 factors and additional runs were conducted
as sensitivity tests for inclusion of selected species. The sensitivity runs included species such as
chlorine and various metals that had a large fraction of samples below detection. Species such as
chlorine had very low R2 values between measured and modeled results (0.11 at Allen Park) and
had little effect on the overall results; they were therefore, excluded. Additional sensitivity runs
were conducted to test the use of elemental potassium instead of potassium ion. Potassium ion,
which is a typical tracer for wood burning, was mostly below detection whereas elemental
potassium, not unique to wood burning, was mostly above detection. The results of the PMF run
showed a strong correlation between potassium and potassium ion with both species coming out
in the same factors, indicating that elemental potassium was useful as a wood smoke tracer for
these sites.

Additional runs were conducted using data collected prior to July 2003 and using data
collected from July 2003 to October 2005. Samples taken between July 2003 and October 2005
had recently updated uncertainties reported in AQS. For samples taken prior to July 2003,
uncertainties were estimated based on methods previously discussed. The same factors and
similar mass attributed to each factor were found using either set of data, showing that the
estimated uncertainties do not change the solution.

The final number of runs and the species used at each site were based on model
performance criteria including the Q-value, convergence, species correlations, and mass
recovery. The number of factors for both STN and SANDWICH data is summarized in
Table 3-1. For each site and data set, 10 random runs were conducted for the final number of
factors to ensure robust results. Over the 10 runs, Q values were stable at all three sites and were
within 50% of the theoretical Q values. Residuals of the PMF results were also within

3-8

-------
recommended bounds (-3 < residual < 3) (Paatero et al., 2003) excluding a few outliers. At
Allen Park, Luna Pier, and Dearborn, there were 11,1, and 4 outliers, respectively.

Table 3-1. Number of factors used for final PMF runs at the three STN sites for
both STN data and SANDWICH data.

Sites

STN Data

SANDWICH Data

Luna Pier

Allen Park

Dearborn

Ideally, tracers for PMF sources would be mostly above detection limits and unique to a
particular source. Of the species within the STN data sets, iron and zinc are mostly above
detection and are unique to industrial sources in the Detroit area. Sources of these species
should, therefore, be easily quantifiable. Other metals such as nickel, chromium, and manganese
have more than 50% of the samples below detection. These species are nearly unique tracers and
can still be quantified. For sources such as wood burning, the only unique metal tracer is
potassium. Potassium has other sources including soil; consequently, it will be hard to isolate a
wood burning source.

The factor profiles from the PMF runs are presented in Figures 3-10 through 3-12. All
three sites had some similar factors including ammonium sulfate, ammonium nitrate, soil, and
mobile sources. The presence of sulfate, nitrate, and soil is consistent with the ambient data
composition previously examined. In addition, we expect to find an influence of mobile sources
as all three sites are surrounded by interstates and various highways as shown in Figure 2-1.
Correlated species were grouped together in factors as indicated in Section 3-1. Silicon/calcium
and nickel/chromium factors were identified in addition to a second calcium factor at Luna Pier.
The identification of two industrial factors at Luna Pier may be attributed to the proximity of this
site to both Toledo and Detroit (Figure 2-1). Of all three sites, Dearborn had the most factors
resolved. This is expected due to the complexity of sources around the site.

The mass composition of the PMF results is presented in pie charts in Figures 3-13
through 3-15 for STN runs. At all three sites, ammonium nitrate, ammonium sulfate, and
organic carbon are the largest portions of the mass, which is in agreement with the ambient data.
Yearly and seasonal trends were examined for the PMF results and are presented in Figures 3-16
through 3-18. Average yearly trends in the PMF results showed decreased mass in 2004,
consistent with the ambient data. At all sites, there is in increase in apportioned mass in the
summertime with increased ammonium sulfate and organic carbon, also consistent with the
ambient data.

3-9

-------
5 a

O iA

-Or- O C
« tfl Z N

Steal

sserMsemim

SSS2S " "

sls^isssSs

dfjrsrsrrdi2

rT-rWfliflSWTi"

MINI Mi

3. i¦:. ¦¦:¦: ,J in.d.' \¦...¦ .¦ £ .: tV_V JV '!i:^

AmmMii/ate

ATM

Mixed Irrdu strial

', i¦ V'. i ¦' ¦. - ¦'•¦¦ V ¦ -,V1' ; ,V- '. : V ,:¦

Gene-af Mobile

Zn Smelter

wajwi lU A\k MvmV'w/l fu.

Cement

,-^W^V-Aa-v-. ¦¦V'^'Va/ U,^T"fti'A^AA :2k£z£i

3 t # Js n • J; • *¦ -tl ¦ • p Jt

° 'TtVff tTtfiTrr- nvrfrVh YH Tf

A44«444viaa
-------
N N Q

ggSSS«55|S220

sill

II) IA IA IA IA O

S5S5S5

lfll» r r

Figure 3-11. Allen Park PMF factor profiles and time series for STN data
(December 2000 through 2005).

3-11

-------
*00
30
1 60

t 40

20
0

too
- ao
§ 60

t 40

T a0
R fl°

LiJ__LJ__Ll.L_LJ__LiJ__LJ__i_iJ__LJ__LiJ__L_i_LiJ__L_i

StlMl

ftogipnai Carbon fMobi**}

. i 1

I-. -

Copp

l ¦

1 ¦ B

M t I i i i M ! ! M

lllilllill I iilllllllltliI

O
25

20
E 15
I 10
5

Rc^onal CirtHwi (Moblto)

lii^H A;5 mtoi Ml Ik i^Ai k - 4*

Capper

\ | ,i im • it/ lb! 0 i ''j L

11 1 rU - / ',

'f-

* Jfeym. A

Chrome ptabr^

;--i

4oe 4 :

Q/4

18
i3

Local Carton (dw-sufl

6
4

2 -H

Steel MaJXifaclurin? |Mn,'K>

. 4

Is :^ : jsll::: r i:4: : ::: ^:::::::: j/I

jli,i-'L,.!i,l..;'iMl. fc-ft k |l,!L..r^ f I¦.-n¦ ^Sf !^. if

'r' ApwSCM
«

ITil''

3 20

Anm*K»

| 2D

'jitLfcU'V ^

¦: ¦ . raioif ¦: ¦

if!ISII!2!iHIII!!lllf!I|!!

Figure 3-12. Dearborn PMF factor profiles and time series for STN data
(May 2002 through December 2005).

3-12

-------
0.22. 2%-i
0.59. I"..

~ Ammonium Sulfate

¦ Ammonium Nitrate

¦ General Mobile

¦ Soil

~ Zinc Smelter

¦ Chrome Plating (from
Toledo)

~ Mixed Industrial
(from Toledo)

¦ Cement Plant

Figure 3-13. Average composition of 8-factor PMF results at Luna Pier for STN
data (May 2002 through December 2005).

~ Ammonium Sulfate

¦ Ammonium Nitrate

¦ General Mobile

¦ Soil

~ Zn Smelter (from
Dearborn)

¦ Chrome Plating

¦ Steel

¦ Wood Burning

~ Cu Smelter (Local
Influence)

Figure 3-14. Average composition of 9-factor PMF results at Allen Park for STN
data (2000 through 2005).

3-13

-------
~ Ammonium Sulfate

¦ Ammonium Nitrate

¦ Mobile

¦ Diesel (w/Lead)

¦ Soil

~ Zinc Smelter

¦ Chrome Plating

¦ Steel

~ Steel Manufacturing (Mn)

¦ Copper Smelter

Figure 3-15. Average composition of 10-factor PMF results at Dearborn for STN
data (May 2002 through December 2005).

3-14

-------
Seisonil Comparison of PMF Contribution itLuniPier _ir. Yeiriy Comparison of PMF Contribution it Lum Pier

2S t 25

Winter Spring Summer Fill
Yeiriy Compirison of Avenge PMF Factor Contribution

forthe Winter Seison it Lum Pier 25

2002 2003 2004 2005

Yeiriy Compirison of Avenge PMF Fietor Contribution
forthe Summer Season it Luna Pier

2003

2004

2005

2002 2003 2004 2005

~ Ammonium Sulfate ¦ Ammonium Nitrate ¦ General Mobile HSoil

~ Zn Smelter ¦ Chrome Plating ~ Mixed Industrial ¦ Cement Plant

Figure 3-16. Yearly and seasonal trends in PMF mass composition at Luna Pier
for STN data (May 2002 through December 2005).

3-15

-------
Seasonal Comparison of Average PMF Factor
Contributions it Allen Park

Yearly Comparison of Average PMF Faetor
Contributions at All en Park

Winter Spring Summer Fill
Yearly Comparison of Average PMF Faetor Contributions
for the Winter Season at Allen Park

2001 2002 2003 2004 200S

Yearly Comparison of Average PMF Faetor
Contributions for the Sum mer Season at AIIhi Park

2001 2002 2003 2004 2005

2001 2002

2003

2004

2005

~ Ammonium Sulfate

¦ Ammonium Nitrate

¦ General Mobile

¦ Soil

~ Zn Smelter

¦ Chrome Plating

¦ Steel

¦ Wood Burning

~ Cu Smelter

Figure 3-17. Yearly and seasonal trends in PMF mass composition at Allen Park
for STN data (December 2000 through December 2005).

3-16

-------
Seasonal Comparison of PMF Contributions Yearly Comparison of PMF Factor Contribution

30 t 3D

S?20

315

: 1 o

1111 i i 11

D-I 1 1 1

- „ 2002 2003 2004 2005

winter spring summer fill

Average PMF Faetor Contribution by Year for the Average PMF Faetor Contribution by Year for the

Winter Season at Dearborn

Summer Season atDeariiorn

2003

2004

2005

2002 2003 2004 2005

~ Ammonium Sulfate ¦Ammonium Nitrate ¦ Mobile ¦ Diesel (w/Lead) ¦ Soil

~ Zn Smelter ¦ Chrome Plating ¦ Steel ~ Steel Manufacturing (Mn) ¦ Cu Smelter

Figure 3-18. Yearly and seasonal trends in PMF mass composition at Dearborn
for STN data (May 2002 through December 2005).

Total PM2.5 point source emissions in the Detroit area are shown in Figure 3-19. It is
important to note that the Dearborn site is located close to many large point sources while the
Luna Pier site is located between several large point sources to the northeast and many smaller
sources from the Toledo area to the south. Point source emissions were also examined for
representative species of the PMF factors identified including chromium, nickel, zinc,
manganese, copper, and calcium. Point source emissions are shown in Figures 3-20 (zinc and
manganese), 3-21 (chromium and nickel), and 3-22 (copper). The point source emissions
combined with wind roses provide information on the sources associated with the PMF factors.
Representative wind roses for a typical day at the three STN sites are shown in Figure 3-23.
Typical wind roses provide a comparison for the high pollutant day wind roses. The typical
wind roses show winds coming from all directions at Allen Park and Dearborn and winds
coming from two directions at Luna Pier. Wind roses on high pollutant days of zinc/manganese,
chromium/nickel, and copper are shown in Figure 3-24. High zinc days at Luna Pier occur
when winds are from the Toledo area while high zinc/manganese days at Dearborn and Allen

3-17

-------
Park occur when winds are from industrial sources between the two sites. On high
nickel/chromium days at Dearborn, the wind roses point toward two small point sources directly
by the site. Allen Park appears to be most affected by the Detroit Edison power plant southeast
of the site for nickel and chromium emissions. For high copper days, both Allen Park and
Dearborn are most affected by point sources to the southeast. Dearborn is also located near two
much smaller copper sources. The source farther from Dearborn appears to have the most
influence on high copper days. Calcium point sources are shown in Figure 3-25 along with a
wind rose for high calcium days at Luna Pier. The point sources show several cement,
limestone, and gypsum facilities surrounding Luna Pier, likely explaining the non-soil calcium
source found at this site. Wind roses for high emission days associated with the two industrial
factors at Luna Pier (Figure 3-26) confirm previous suspicions that the two factors are coming
from different directions. The mixed industrial factor at Luna Pier points to the large sources
northeast of the site and the iron/chromium/nickel factor is influenced by winds from the Toledo
area.

3-18

-------
Figure 3-19. Proportional PM2.5 point source emissions map (U.S. Environmental
Protection Agency, 2002).

3-19

-------
D.e t i;o i t' M et i jcTA'i e a

Dearborn,

"Arin.Arbor

Allen Park

j Belleville iL R^ j
w®ro>^e7j$8(5$36&w* yn'e dbo.—%

^ £xl _«y ry

isfield

Luna Pier

"Toledo"

Kilometers

Zinc (Tons/yr)

100

O 1,000

Legend

£3 Monitoring Sites
Manganese (Tons/yr)

100

00,000

Figure 3-20. Proportional zinc and manganese point source emissions map (U.S.
Environmental Protection Agency, 2002).

3-20

-------
D e troi t" M e t iio-A'r e a"

Dearborn;

Belleville IL Rop-1
K.VE.

£3 Monitoring Sites
Chromium (Tons/yr)

Nickel (Tons/yi)

Luna Pier

Terrtpetaince /

iMev/ /

[Toledo*

Kilometers

Figure 3-21. Proportional chromium and nickel point source emissions map (U.S.
Environmental Protection Agency, 2002).

3-21

-------
Urfrarft

D e tro i t' M e t iio-A'r e a

Dearborn;

J Belleville IL. Rqjth
s^EiTK&lTWETpfm

kWKkK4 YNEi

Luna Pier

Terrtpetaince /

iMev/ /

"Toledo

Kilometers.

Legend

£3 Monitoring Sites
Copper (Tons yr)

• 100

Figure 3-22. Proportional copper point source emissions map (U.S.
Environmental Protection Agency, 2002).

3-22

-------
Allen Park

Dearborn

Lima Pier

25%
-20% ¦
15%
10%

25%
20%
15%
10%
%

WS (mIs)
> 30
20-30
15-20
10-15

15-10
2.5-5
0-2.5

Figure 3-23. Wind roses for three STN sites: Allen Park, Dearborn, and Luna
Pier (Allen Park: 2000 through 2005, Dearborn and Luna Pier: May 2002
through December 2005).

Figure 3-24. Wind roses of zinc/manganese, chromium/nickel, and copper on
high pollutant days (U.S. Environmental Protection Agency, 2002). Units for
emissions are the same as for Figures 3-20 through 3-22.

3-23

-------
Figure 3-25. Wind roses of calcium on high concentration days with point source
emissions at Luna Pier (U.S. Environmental Protection Agency, 2002).

3-24

-------

ty ww

= Mixed Industrial

Mixed Industrial /jT

(Al/As/Cu/Ni)

(Fe/Ni/Cr) .

1 "r fc

1 una Pi*i' ^

WS (m/s)

>30

^ J /

f 1 J. 7*^1 1

20-30

S | mir) Pier'

15-20

(1 cz£\0

10-15

ri-—S-w\X Jl

5-10 *

-jr

.
,

2.5-5

— 1 L-1"i

Figure 3-26. Wind roses on high mixed industrial factor days with PM2.5 point
source emissions at Luna Pier (U.S. Environmental Protection Agency, 2002).
Units for emissions are the same as for Figures 3-19 through 3-22.

3.3 SOURCE APPORTIONMENT OF SANDWICH DATA

The PMF results using SANDWICH data are presented in Figures 3-27 through 3-29.
PMF results using STN data and ambient data composition for both data sets are also shown. In
the SANDWICH PMF results, a larger fraction of the mass is attributed to ammonium sulfate
and less to ammonium nitrate, consistent with ambient data. Better mass recovery was achieved
using the SANDWICH data set, mostly due to the difference in sulfate mass. With respect to the
number of factors, Allen Park was the only site for which the SANDWICH and STN data sets
did not agree. Using the SANDWICH data, PMF was able to split the carbon into a mobile and a
diesel source, which was not achieved with the STN data. However, neither the wood burning
nor the steel source was identified with the SANDWICH data. Overall, there is good agreement
between SANDWICH and STN results. In general, results found using STN data are useful for
FRM and SANDWICH applications. On a daily basis, the SANDWICH PMF results can be
different than the STN PMF results, but these differences are nearly all due to the differences
between SANDWICIi and regular STN data (i.e., carbon, nitrate, and sulfate concentrations are
already different in the two data sets).

3-25

-------
Luna Pier

N = 202

N = 186

Ambient Data

N = 186

N = 202

PMF Results

STN

SANDWICH

STN

SANDWICH

Ambient Data
~ Amm Sulfate ¦ Soil

¦ Atntn Nitrate ¦ Other Species

¦ OM ~ Diff in sum of species
BEC and filter mass

PMF Results

~Amm Sulfate

¦ Cement Plant

¦ Amm Nitrate

~ Zn Smelter

fl General Mobile

~ Mixed Industrial

¦ Soil

¦ Chrome Plating

~ Diff in sum of species and filter mass

Figure 3-27. PMF results and ambient mass composition for both STN and
SANDWICH data sets at Luna Pier (May 2002 through December 2005).

3-26

-------
Allen Park

STN

SANDWICH

STN

SANDWICH

Ambient Data
~ Amm Sulfate "EC

¦ Amm Nitrate ¦Soil

¦ CM ¦ Other Species

PMF Results

~ Airim Sulfate

~ Zn Smelter

¦ Ainm Nitrate

9 Chrome Plating

® General Mobile

¦ Steel

¦ Soil

¦ W o o d Burning

~ Cu Smelter

Figure 3-28. PMF results and ambient mass composition for both STN and
SANDWICH data sets at Allen Park (December 2000 through December 2005).

3-27

-------
30

Dearborn

25 -

20 -

3. 15

5 -

0 +-

N = 201

N = 190

Ambient Data

N = 190

PMF Results

STN

SANDWICH

STN

SANDWICH

Ambient Data
~ Amm Sulfate ¦ Soil

¦ Amm Nitrate ¦ Other Species

¦ OM ~ Diff in sum of species

and filter mass

PMF Results

~ Amm Sulfate

~ Zn Smelter

¦Amm Nitrate

in Chrome Plating

¦ General Mobile

¦ Steel

¦ Diesel (w/Pb)

¦ Steel (Mn)

¦ Soil

~ Cu Smelter

Figure 3-29. PMF results and ambient mass composition for both STN and
SANDWICH data sets at Dearborn (May 2002 through December 2005).

3.4 SOURCE APPORTIONMENT OF STN AND AIR TOXICS DATA

Additional PMF runs were conducted that included air toxics data. The additional runs
were conducted for Allen Park, because this site had the largest number of STN samples. The
availability of air toxics data is much more limited than speciated PM2.5 data; therefore, the
number of samples used in PMF was drastically decreased to get overlapping sample days. The
air toxics included in the analysis were benzene, o-xylene, ethylbenzene, toluene, formaldehyde,
and acetaldehyde with a total of 158 samples from April 25, 2001, through November 6, 2005.
PMF runs were first conducted using the 158 days of STN data only to ensure results were
similar to those previously found with the 565 sample runs. The factors obtained from the 158
sample PMF runs were similar to those obtained with the full set of 565 samples with slight
differences in average mass (Figure 3-30). A steel factor was not identified in the 158 sample
run, instead a mixed industrial factor of various metals was obtained. The major components of

3-28

-------
mass (ammonium sulfate, ammonium nitrate, and mobile sources), however, were apportioned
similarly between the two data sets.

16.00

14.00
12.00

-r- 10.00

3 8.00

in
re

E 6.00

4.00
2.00
0.00

~ Mixed Industrial

~ Cu Smelter

¦ Wood Burning

¦ Steel

¦ Chrome Plating

~ Zn Smelter

¦ Soil

¦ General Mobile

¦ Ammonium Nitrate

~ Ammonium Sulfate

565 sample

158 Samples

Figure 3-30. Comparison of average PMF mass contribution at Allen Park for a
9-factor 565 sample run and a 9-factor 158 sample run (April 25, 2001, through
November 6, 2005).

PMF runs were conducted using 8 to 11 factors with air toxics species included. Over all
runs conducted, OC and EC were not split into separate factors. Benzene, o-xylene,
ethylbenzene, and toluene were grouped with the steel source (iron and chromium), while
formaldehyde and acetaldehyde were grouped with the general mobile source (OC/EC) in all
runs conducted. One of the expectations of using the air toxics data with STN data was that the
additional species would help separate the mobile sources into gasoline and diesel factors. At
Allen Park, though, no additional insight into the split of mobile sources was obtained.

3-29

-------
-------
4. SUMMARY AND CONCLUSIONS

Source apportionment using PMF was applied to both STN data sets and SANDWICH
data sets at three Michigan STN sites. In addition, exploration of adding air toxics to PM2.5 data
for use in PMF was performed.

4.1 NEW STN UNCERTAINTIES

The application of PMF to STN data sets tested the use of recently updated uncertainties
that are larger than those previously used, especially for metal species. Table 4-1 shows a
comparison of PMF results using recently updated uncertainties and previous results at Allen
Park. The comparison shows good agreement in general with some differences including an
increase in the number of industrial factors isolated with the new uncertainties. Additionally,
separate mobile and diesel factors were no longer successfully isolated. As discussed in
Section 2.1.2, uncertainties for carbon and ion species were not updated. OC and EC, therefore,
have much smaller relative uncertainties than the other species used in the model. The increased
uncertainty for metal species provides more flexibility in fitting these species, and may be the
cause of the differences seen. Percent contributions were similar across studies; except for
higher mass attributed to ammonium sulfate with the current work.

The recently updated STN uncertainties for PM2.5 data were not available for the entire
data set (only for July 2003-October 2005) and as a result, extrapolation methods were used for
estimated the remaining uncertainties. PMF runs were conducted using samples collected prior
to July 2003 and using samples collected from July 2003 to October 2005. The factors identified
in the PMF runs, as well as the mass apportioned to each factor, were the same using both data
sets. The uncertainty estimations used did not affect the solution. The methods used provide
good estimates of uncertainties for PMF purposes.

4-1

-------
Table 4-1. Comparison of PMF results using recently updated STN uncertainties
and PMF results from previous studies at Allen Park. Values are the percentage
of mass attributed to each source.

Allen Park

Previous Work

Current Work

(2002-2005)

(2000-2005)

Ammonium Sulfate

Ammonium Nitrate

Soil

Mobile (OC)

Diesel (EC)

Biomass Burning

Zinc Smelter

Chrome Plating

Copper

Steel

4.2 PMF ON STN AND SANDWICH DATA SETS

SANDWICH data were developed to adjust STN data to better match FRM data and
SANDWICH has never been used for PMF purposes. PMF was run using STN data sets and
SANDWICH data sets and results were compared. The major difference between the two data
set results was that for the SANDWICH data, more mass was attributed to ammonium sulfate
and less mass was attributed to ammonium nitrate, which is consistent with ambient data. Of the
three sites considered, the same number of sources was identified using the two data sets at Luna
Pier and Dearborn, while differences were found at Allen Park, due to larger changes in the
carbon data between STN and SANDWICH data sets. SANDWICH data showed better mass
closure overall, but there was more variability in the results on a day-to-day basis, mostly due to
the differences in sulfate, nitrate, and carbon in the data sets. Overall, results found using STN
data were similar to those found using SANDWICH data, and differences between the results
were generally within the uncertainty of the application. Thus, at Detroit, the application of STN
data is likely sufficient for most applications, though this may not hold true for all areas. A
thorough comparison of STN versus SANDWICH data prior to source apportionment should be
conducted in the future to understand if differences between the data sets are important enough
to yield different source apportionment results.

4.3 LINKING PMF FACTORS TO SOURCES

Point source emissions of species that were representative of PMF factors found at the
three STN sites were examined along with meteorological data to better understand the link
between PMF factors and sources. Wind roses were developed for days on which high mass was
apportioned to various PMF factors as well as for typical days at each of the three STN sites.
Differences were noted between the wind roses for typical days at the sites and wind roses for

4-2

-------
high mass days. At both Allen Park and Dearborn, typical day wind roses show winds coming
almost equally from all directions. This is not found on high mass days. For example, on high
zinc and manganese days, winds at Allen Park are from the northeast and winds at Dearborn are
from the southwest. The winds at both sites come from the direction where industrial facilities
that perform zinc smelting are located. The differences in typical and high factor mass wind
roses indicate that high mass wind roses provide useful information for understanding which
sources are associated with which factors. Wind roses indicate that the zinc/manganese factor at
Allen Park and Dearborn is most likely associated with zinc smelting. The same analyses were
performed for the other industrial factors at Allen Park, Dearborn, and Luna Pier. Understanding
the meteorology of the sites provided useful information in understanding the appropriate
sources to associate with PMF factors.

4.4 PMF RUNS WITH PM25 DATA AND AIR TOXICS—EXPLORATORY

ANALYSES

Additional PMF runs were conducted using both STN PM2.5 data and air toxics data with
the expectation of separating the mobile sources into gasoline and diesel factors. However,
gasoline and diesel factors were not successfully isolated at Allen Park, even with the addition of
several air toxic species. The inability to separate the mobile sources was likely caused by the
limited data set. Air toxics data examined at the STN sites were mostly below detection limits
and were measured with a lower frequency than the PM2.5 data. As a result, the data set at Allen
Park was reduced from 565 samples to 158 samples. Results from PMF runs with combined
PM2.5 and air toxics data may produce some insight into mobile sources at sites where more air
toxics data are available.

4-3

-------
-------
5. REFERENCES

Anttila P., Paatero P., Tapper U., and Jarvinen O. (1995) Source identification of bulk wet
deposition in Finland by positive matrix factorization. Atmos. Environ, (published
without a date) 29 (14), 1705-1718, Pergamon Press, Ltd. (1352-2310/95).

Begum B.A., Kim E., Jeong C.H., Lee D.W., and Hopke P. (2005) Evaluation of the potential
source contribution function using the 2002 Quebec forest fire episode. Atmos. Environ.
39, 3719-3724.

Birch M.E. and Carey R.A. (1996) Elemental carbon-based method for monitoring occupational
exposures to diesel exhaust. Aerosol Sci. Technol. 25, 221-241.

Buzcu B., Fraser M.P., Kulkarni P., and Chellam S. (2003) Source identification and
apportionment of fine particulate matter in Houston, TX, using positive matrix
factorization. Environmental Engineering Science 20 (6), 533-545.

Draxler R.R. and Hess G.D. (1997) Description of the HYSPLIT 4 modeling system. Technical
memorandum by NOAA, ERL ARL-224, December 24.

Eberly S. (2005) EPAPMF 1.1 user's guide. Prepared by the U.S. Environmental Protection
Agency, National Exposure Research Laboratory, Research Triangle Park, NC, June.

Frank N.H. (2006) Retained nitrate, hydrated sulfates, and carbonaceous mass in federal

reference method PM2.5 for six eastern U.S. cities. J Air Waste Manag Assoc 56 (4), 500-
511.

Hafner H.R. (2005) Multivariate receptor modeling workbook. Work plan prepared for the U.S.
Environmental Protection Agency, Research Triangle Park, NC, by Sonoma Technology,
Inc., Petaluma, CA, STI-905103-2736-WP, April.

Juntto S. and Paatero P. (1994) Analysis of daily precipitation data by positive matrix
factorization. Environmetrics 5, 127-144.

Kim E., Larson T.V., Hopke P.K., Slaughter C., Sheppard L.E., and Claiborn C. (2003) Source
identification of PM2.5 in an arid northwest U.S. city by positive matrix factorization.
Atmos. Res. 66 (4), 291-305.

Kim E. and Hopke P.K. (2004a) Improving source identification of fine particles in a rural

northeastern U.S. area utilizing temperature-resolved carbon fractions. J. Geophys. Res.
109 (D9).

Kim E. and Hopke P.K. (2004b) Source apportionment of fine particles at Washington, DC,

utilizing temperature-resolved carbon fractions. J. Air & Waste Manage. Assoc. 54 (7),
773-785.

5-1

-------
Kim E., Hopke P.K., and Edgerton E.S. (2004a) Improving source identification of Atlanta
aerosol using temperature resolved carbon fractions in positive matrix factorization.
Atmos. Environ. 38, 3349-3362.

Kim E., Hopke P.K., Larson T. V., Maykut N.N., and Lewtas J. (2004b) Factor analysis of Seattle
fine particles. AerosolSci. Technol. 38 (7), 724-738.

Kim E., Brown S.G., Hafner H.R., and Hopke P.K. (2005a) Characterization of non-methane
volatile organic compounds sources in Houston during 2001 using positive matrix
factorization. Atmos. Environ. 39, 5934-5946.

Kim E., Hopke P.K., Kenski D.M., and Koerber M. (2005b) Sources of fine particles in a rural
midwestern U.S. area. Environ. Sci. Technol. (39), 4953-4960.

Kim E., Hopke P.K., and Qin Y. (2005c) Estimation of organic carbon blank values and error

structures of the speciation trends network data for source apportionment. J. Air & Waste
Manag. Assoc. 55, 1190-1199.

Larsen R.K. and Baker J.E. (2003) Source apportionment of polycyclic aromatic hydrocarbons in
the urban atmosphere: a comparison of three methods. Environ. Sci. Technol. 37 (9),
1873-1881.

Lee E., Chan C.K., and Paatero P. (1999) Application of positive matrix factorization in source
apportionment of particulate pollutants in Hong Kong. Atmos. Environ. 33, 3201-3212.

NIOSH (1999) Method 5040: Elemental Carbon (Diesel Exhaust). In NIOSHManual of
Analytical Methods, 4th Ed., Issue 3, National Institute of Occupational Safety and
Health, Cincinnati, OH.

Paatero P. and Tapper U. (1994) Positive matrix factorization: a non-negative factor model with
optimal utilization of error estimates of data values. Environmetries 5, 111-126.

Paatero P. (1997) Least squares formulation of robust non-negative factor analysis.

Chemometrics and Intelligent Laboratory Systems 37, 23-35.

Paatero P. (1999) The multilinear engine - A table-driven, least squares program for solving
multilinear problems, including the n-way parallel factor analysis model. Journal of
Graphical Statistics 8, 854-888.

Paatero P., Hopke P.K., and Philip K. (2003) Discarding or downweighting high-noise variables
in factor analytic models. Anal. Chim. Acta 490, 277-289.

Poirot R.L., Wishinski P.R., Hopke P.K., and Polissar A.V. (2001) Comparative application of
multiple receptor methods to identify aerosol sources in northern Vermont. Environ. Sci.
Technol. 35 (23), 4622-4636.

Polissar A.V., Hopke P.K., and Poirot R.L. (2001) Atmospheric aerosol over Vermont: chemical
composition and sources. Environ. Sci. Technol. 35 (23), 4604-4621.

5-2

-------
Ramadan Z., Song X.-H., and Hopke P.K. (2000) Identification of sources of Phoenix aerosol by
positive matrix factorization. J. Air & Waste Manage. Assoc. 50, 1308-1320.

Research Triangle Institute (2004) Extended semi-annual data summary report for the chemical
speciation of PM2 5 filter samples project. Report prepared by the Research Triangle
Institute, Research Triangle Park, NC, RTI/07565/06DS. Available on the Internet at
.

Rice J. (2004) Comparison of integrated filter and automated carbon aerosol measurements at
Research Triangle Park, North Carolina. Aerosol Science & Technology (38), 23-36.

Song X.H., Polissar A.V., and Hopke P.K. (2001) Sources of fine particle composition in the
northeastern U.S. Atmos. Environ. 35 (31), 5277-5286.

Subramanian R., Khlystov A.Y., Cabada J.C., and Robinson A.L. (2004) Positive and negative
artifacts in particulate organic carbon measurements with denuded and undenuded
sampler configurations. Aerosol Sci. Technol. 38, 27-48.

U.S. Environmental Protection Agency (2002) National Emission Inventory (NEI), air pollutant
emission trends, NEI emission trends data and estimation procedures, criteria pollutant
data, current emission trends summaries. Available on the Internet at
last accessed January 24, 2003.

Zhou L., Kim E., Hopke P.K., Stanier C.O., and Pandis S. (2004) Advanced factor analysis on
Pittsburgh particle size-distribution data. Aerosol Sci. Technol. 38 (SI), 118-132.

5-3

-------