Development of Advanced Factor Analysis Methods for … · 2015-09-18 · Development of Advanced...

Development of Advanced Factor Analysis Methods for Carbonaceous

PM Source Identification and Apportionment

Philip K. Hopke

Center for Air Resources Engineering and ScienceClarkson University

Investigators

• This project represents a collaboration among• Pentti Paatero, University of Helsinki• Ron Henry, University of Southern California• Cliff Spiegelman, Texas A&M University

• Work was conducted by• Eugene Kim, Weixiang Zhao, Jong Hoon Lee,

David Ogulei, and Ramya Sunder Raman

Objectives

• The objective of this project was to combine the best features of the two advanced factor analysis models, UNMIX and Positive Matrix Factorization (PMF), and to test the effectiveness of this improved factor analysis methodology by analysis of the data developed in the various supersites with an emphasis on data from the New York City Supersite and other data from New York State.

Methodological Research

• Part of the effort in the project was methodological studies. • Duality of Solutions

• Singular value decomposition of the data leads to two sets of eigenvectors. One set of eigenvectors spans a space in which source compositions are points and source contributions are hyperplanes. This space is shown to be dual to the space spanned by the second set of eigenvectors of the data in which source compositions are hyperplanes and source contributions are points. The duality principle has been applied to greatly increase the computational speed of the Unmix multivariate receptor model.

Methodological Research

• G-Space Edges• Scatter plots are created of pairs of source contribution

factors. When factors are plotted in this way, unrealistic rotations appear as oblique edges that define the distribution of points away from one (or both) of the coordinate axes. With a correct rotation, the limiting edges usually coincide with the axes or lay parallel with them. Inspection of the plots helps one in choosing a realistic rotation.

G-Space Edges

• If the two factors are independent of one another, then the resulting contribution values should completely fill the scatter plot and there should be no correlation between them.

G-Space Edges

0 1000 2000 3000 4000 50000

0.5

1

1.5

2

2.5

3

3.5

x 104

Biomass burning

Mot

or v

ehic

les

G-Space Edges

0 500 1000 1500 2000 25000

1000

2000

3000

4000

5000

6000

7000

8000

Road dust

2−st

roke

eng

ines

G-Space Edges

0 1000 2000 3000 4000 5000

0

5000

10000

15000

20000

Biomass burning

Met

al s

mel

ter

G-Space Plots

• Obviously there is an edge in this plot.

• Does it make sense that these two factors are correlated?

• If not, it suggests the need for a rotation.

G-Space Edges

0 1 2 3

x 104

0

5000

10000

15000

20000

Motor vehicles

Met

al s

mel

ter

G-Space Edges

• Note that there can be points outside the apparent edge.

• These points should be checked to be sure they belong. However, it may be necessary to ignore them.

G-Space Edges

0 2000 4000 60000

0.5

1

1.5

2

2.5

3

3.5

4

x 104

Biomass burning

Met

al s

mel

ter

G-Space Edges

0 1 2 3

x 104

0

0.5

1

1.5

2

2.5

3

3.5

4

x 104

Motor vehicles

Met

al s

mel

ter

G-Space Edges

• Even after rotation, edges can persist.

G-Space Edges

0 1000 2000 3000 4000 50000

0.5

1

1.5

2

2.5

3

3.5

4

4.5

x 104

Biomass burning

Mot

or v

ehic

les

Applications

• Particle Composition Data• Use of IMPROVE carbon fractions • Use of STN data

Watson, J.G.; Chow, J.C.; Lowenthal, D.H.; Pritchett, L.C.; Frazier, C.A.; Neuroth, G.R.; and Robbins, R. (1994). Differences in the carbon composition of source profiles for diesel- and gasoline-powered vehicles. Atmos. Environ., 28(15):2493-2505.

Gasoline-fueled vehiclesIMPROVE

• IMPROVE provides carbon fractionsDiesel-fueled vehicles

IMPROVE

0.00.10.20.3

OC1 OC2 OC3 OC4 OP EC1-OP EC2 EC3

0.0

0.4

0.8Atlanta

Con

cent

ratio

n (μ

g/μg

)

Washington, DC

OC1 OC2 OC3 OC4 OP EC1-OP EC2 EC30.00.10.20.3

Brigantine

Gasoline Vehicles

IMPROVE

Diesel Vehicles

0.0

0.4

0.8

OC1 OC2 OC3 OC4 OP EC1-OP EC2 EC3

0.0

0.3

0.6Atlanta

Con

cent

ratio

n (μ

g/μ g

)

Washington, DC

OC1 OC2 OC3 OC4 OP EC1-OP EC2 EC30.0

0.3

0.6Brigantine

STN-NYCGasoline Vehicles

STN-NYCDiesel Vehicles

Gasoline – Diesel Split

• Can gasoline vehicular emissions be separated from diesel emissions?• Shah et al. (Environ. Sci. Technol. 38 (9), 2544-2550,

2004) show that stop and go and creeping diesel vehicles emit roughly 50:50 OC/EC as measured with the NIOSH protocol.

• Problems of “smokers” looking like “diesel” emissions


• Thus, the “diesel” profile tends to reflect the emissions from vehicles moving at highway speed (i.e., min OC/EC ratio)

• “Gasoline” reflects the maximum OC/EC ratio

• However, the oil additive trace elements tend to go into the “diesel” profile.


• Does the choice of IMPROVE or NIOSH protocols affect the apportionment and the assignment of mass to “diesel,” “gasoline,” and other major carbonaceous aerosol sources like biomass burning.

• We have an opportunity to make such a comparison using data from the St.Louis-Midwest Supersite.

Comparison of PMF using either IMPROVE or NIOSH Data

• Using daily integrated PM2.5 samples obtained at the St. Louis-Midwest Supersite, OC/EC analyses were performed by the two protocols:• OC-EC were originally analyzed at UW-Madison

with the ACE-Asia variant of the NIOSH protocol• Subsequently, the same samples were analyzed at

DRI using the IMPROVE protocol


• Analysis was undertaken for three sets of PM2.5 speciation data at the St. Louis-Midwest Supersite in which each set differs only in the carbon concentrations. • The first set (679 samples for 31 species) has eight carbon

fractions (OC1 to OC4, OP, and EC1 to EC3) from the IMPROVE protocol.

• The second set (679 samples for 25 species) included only the total IMPROVE OC (TOC = OC1 + OC2 + OC3 + OC4 + OP) and EC fractions (TEC = EC1 - OP + EC2 + EC3), respectively.

• The last set (679 samples for 25 species) contains OC and EC concentration obtained by the NIOSH analysis.


• Solutions with 11 factors, 10 factors, and 10 factors were obtained by IMPROVE carbon fractions, IMPROVE TOC-TEC values, and NIOSH OC-EC values, respectively, for the St. Louis Supersite PM2.5.

OC1

OC2

OC3

OC4

OP EC1

EC2

EC3

SO4

NO3

NH4

Al

As

Ba

Ca

Co

Cu

Fe

K Mn

Ni

Pb

Rb

Se

Si

Sr

Ti

V Zn

Zr

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

Con

cent

ratio

n (μ

g μg

-1)

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

OC1

OC2

OC3

OC4 O

PEC

1EC

2EC

3SO

4NO

3NH

4 Al

As

Ba

Ca

Co

Cu

Fe K M

n Ni

Pb

Rb

Se

Si

Sr

Ti V Zn

Zr 0.001

0.010.1

1

OC EC SO4

NO3

NH4

Al

As

Ba

Ca

Co

Cu

Fe

K Mn

Ni

Pb

Rb

Se

Si

Sr

Ti

V Zn

Zr

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

OC EC SO4

NO3

NH4 Al

As

Ba

Ca

Co

Cu

Fe K M

n Ni

Pb

Rb

Se

Si

Sr

Ti

V

Zn

Zr 0.001

0.010.1

1

OC EC SO4

NO3

NH4

Al

As

Ba

Ca

Co

Cu

Fe

K Mn

Ni

Pb

Rb

Se

Si

Sr

Ti

V Zn

Zr

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

0.0010.01

0.11

OC EC SO4

NO3

NH4 Al

As

Ba

Ca

Co

Cu

Fe K M

n Ni

Pb

Rb

Se

Si

Sr

Ti

V

Zn

Zr 0.001

0.010.1

1

Airborne soil

Copper production

Diesel emission/Road dust

Gasoline exhaust

Lead smelting

Secondary nitrate

Secondary sulfate

Steel processing

Wood smoke

Zinc smelting

Carbon-rich sulfate

(a) (b) (c)

Airborne soil

Copper production


Gasoline exhaust

Lead smelting

Secondary nitrate

Secondary sulfate

Steel processing

Wood smoke

Zinc smelting

Airborne soil

Copper production


Gasoline exhaust

Lead smelting

Secondary nitrate

Secondary sulfate

Steel processing

Wood smoke

Zinc smelting



0.0010.01

0.11

OC ECSO4

NO3

NH4 Al

As

Ba

Ca

Co

Cu

Fe K M

n Ni

Pb

Rb

Se

Si

Sr

Ti V Zn

Zr

0.0010.01

0.11

IMPROVE

NIOSH

Gasoline


Diesel

0.0010.01

0.11

OC ECSO4

NO3

NH4 Al

As

Ba

Ca

Co

Cu

Fe K M

n Ni

Pb

Rb

Se

Si

Sr

Ti V Zn

Zr

0.0010.01

0.11

IMPROVE

NIOSH

0.0010.01

0.11

OC ECSO4

NO3

NH4 Al

As

Ba

Ca

Co

Cu

Fe K M

n Ni

Pb

Rb

Se

Si

Sr

Ti V Zn

Zr

0.0010.01

0.11

IMPROVE

NIOSH


Biomass Burning

0 40 80

Pred

icte

d PM

2.5 m

ass (μg

m-3

)

0

40

80y = 0.92x + 0.95R2 = 0.93

IMPROVE OC-EC

Measured PM2.5 mass (μg m-3)

0 40 800

40

80y = 0.91x + 1.13R2 = 0.92

IMPROVE C fractions

0 40 800

40

80y = 0.95x + 0.64R2 = 0.94

NIOSH OC-EC


Mass Reconstruction

0 5 100

5

10

0 15 300

15

30

Summed FractionsC-Fraction

Gasoline Diesel

ACE-Asia/NIOSH ECOC (μg m-3)

0 8 160

8

16

Wood smoke

IMPR

OV

E ( μ

g m

-3)


Contribution Comparisons

PublicationsPublications • Comparison between Conditional Probability Function and Nonparametric Regression for Fine Particle Source Directions, E.

Kim, and P.K. Hopke, Atmospheric Environ. 38: 4667–4673 (2004).• A Graphical Diagnostic Method For Assessing the Rotation In Factor Analytical Models Of Atmospheric Pollution, P.

Paatero, P.K. Hopke, B.A. Begum, S.K. Biswas, Atmospheric Environ. 39: 193-201 (2005).• Duality in Multivariate Receptor Model, Ronald C. Henry, Chemom. Intell. Lab. Syst. 77: 59-63 (2005).• Improving Source Apportionment of Fine Particles in the Eastern United States Utilizing Temperature-Resolved Carbon

Fractions. E. Kim and P.K. Hopke, J. Air & Waste Manage. Assoc. 55:1456–1463 (2005).• Estimation of Organic Carbon Blank Values and Error Structures of the Speciation Trends Network Data for Source

Apportionment, E. Kim, P.K. Hopke, and Y. Qin, J. Air Waste Manage. Assoc. 55: 1190 - 1199 (2005).• Identification of Fine Particle Sources in Mid-Atlantic US Area, E. Kim and P.K. Hopke, Water, Soil, and Air Pollution 168:

391-421 (2005).• Analysis of Indoor Particle Size Distributions from an Occupied Townhouse using PMF, D. Ogulei, P.K. Hopke, L. Wallace,

Indoor Air 16:204-215 (2006).• Source Apportionment of Baltimore Aerosol from Combined Size Distribution and Chemical Composition Data, D. Ogulei,

P.K. Hopke, L. Zhou, J.P. Pancras, N. Nair, J.M. Ondov, Atmospheric Environ. 40: S396-S410 (2006).• Comparison between Sample-Species Specific Uncertainties and Estimated Uncertainties for the Source Apportionment of

the Speciation Trends Network Data, E. Kim and P.K. Hopke, Atmospheric Environ. 41: 567-575 (2007).• Factor Analysis of Submicron Particle Size Distributions Near a Major United States-Canada Trade Bridge, D. Ogulei, P.K.

Hopke, A.R. Ferro, P.A. Jaques, J. Air Waste Manage. Assoc. 57: 190-203(2007).• Modeling Source Contributions to Submicron Particle Number Concentrations Measured in Rochester, NY, D. Ogulei, P.K.

Hopke, D.C. Chalupa, and M.J. Utell, Aerosol Sci. Technol. 41: 179-201 (2007).• Spatial Distribution of Source Locations for Particulate Nitrate and Sulfate in the Upper-Midwestern United States, W. Zhao,

P.K. Hopke, and L. Zhou, Atmospheric Environ. 41:1831-1847 (2007)..• A Computation Saving Jackknife Approach to Receptor Model Uncertainty Statements for Serially Correlated Data, C.H.

Spiegelman and E.S. Park, Chemometrics and Intelligent Laboratory Systems (in press, 2007).

QUESTIONS?

Development of Advanced Factor Analysis Methods for … · 2015-09-18 · Development of Advanced...

Documents

Transcript of Development of Advanced Factor Analysis Methods for … · 2015-09-18 · Development of Advanced...