Development of an Ensemble-Trained Source Apportionment for PM2.5
SIVARAMAN BALACHANDRAN (1), Jorge Pachon (1), Roshini Shankaran (1), Armistead G. Russell (1), James A. Mulholland (1), Dongho Lee (2), Sangil Lee (3)
(1) Georgia Institute of Technology (2) Gyeongnam Province Institute of Health and Environment, Changwon, Gyeongnam, Korea (3) Center for Analytical Measurement Services, Korea Research Institute of Standards, Daejon, Korea
Abstract Number: 401
Preference: Platform Presentation
Last modified: November 9, 2009
Working Group: sq2
Most epidemiological studies associating health impacts with PM2.5 have historically focused on using total PM mass, or when available, individual PM2.5 components. Recently, there has been growing interest in associating health outcomes with sources of PM2.5, rather than individual components. However, source apportionment (SA) results can vary significantly depending on the method used. Further, there appear to be issues specific to the types of modeling used. For example, receptor-based source apportionment approaches appear to have excessive day to day variability and are limited spatially since results are representative of the location of observations. Conversely, while emissions-based chemical transport models (CTMs) provide large spatio-temporal coverage, they lack significant day-to-day variation in source impacts and are computationally intensive.
To provide better estimates of source impacts, an ensemble-trained chemical mass balance approach that utilizes the Lipschitz Generalized Optimization (CMB-LGO) has been developed. This method uses results from receptor-based models, including CMB and positive matrix factorization(PMF), and chemical transport models, in this case the community multiscale air quality (CMAQ). First, results from multiple methods are averaged to derive ensemble source impacts. Next, the ensemble source impacts along with 24- hour PM2.5 measurements at the Jefferson St. (JST) site in Atlanta, GA, were used in CMB-LGO to derive optimized, ensemble-trained source profiles (EBSPs). Here, separate profiles were developed for summer (using July 2001 data) and winter (using January 2002 data). Third, the EBSPs were then used in a CMB-LGO application for a year-long data set from JST. These results were then compared to an application of CMB-LGO using measurement-based source profiles (MBSPs). When looking at performance measures, CMB-LGO with EBSPs performed better with a lower reduced chi-square value as well as decreased zero-impact days (largely driven by a decrease in zero impact days from coal combustion). Using EBSPs also resulted in increased impacts from gasoline vehicles but decreased impacts from biomass burning and dust in the summer. In the winter, impacts from biomass burning, dust, and coal combustion increased while other organic carbon (presumably secondary organic carbon) decreased.
Subsequently, the EBSPs were applied to a 9.5 year data set from August 1999 – December 2007 at JST. Both EBSPs and MBSPs were used in CMB-LGO and compared. Similar to the year-long data set, the use of EBSPs had reduced errors and fewer zero-impact days. In addition, the 9.5 year data set allows us to assess variability and identify likely error and quantify uncertainties in both the ensemble method and the model results originally used in deriving the ensemble source impacts. The uncertainty quantification can lead to better understanding of source impacts from various methods which can inform air quality manager and health researchers in their use.