Application of an Ensemble-Trained Source Apportionment Approach at a Site Impacted by Multiple Point Sources

Abstract

Four receptor models and a chemical transport model were used to quantify PM2.5 source impacts at the St. Louis Supersite (STL-SS) between June 2001 and May 2003. The receptor models used two semi-independent data sets, with the first including ions and trace elements and the second including 1-in-6 day particle-bound organics. Since each source apportionment (SA) technique has limitations, this work compares results from the five different SA approaches to better understand the biases and limitations of each. The source impacts calculated by these models were then integrated into a constrained, ensemble-trained SA approach. The ensemble method offers several improvements over the five individual SA techniques at the STL-SS. Primarily, the ensemble method calculates source impacts on days when individual models either do not converge to a solution or do not have adequate input data to develop source impact estimates. When compared with a chemical mass balance approach using measurement-based source profiles, the ensemble method improves fit statistics, reducing chi-squared values and improving PM2.5 mass reconstruction. Compared to other receptor models, the ensemble method also calculates zero or negative impacts from major emissions sources (e.g., secondary organic carbon (SOC) and diesel vehicles) for fewer days. One limitation of this analysis was that a composite metals profile was used in the ensemble analysis. Although STL-SS is impacted by multiple metals processing point sources, several of the initial SA methods could not resolve individual metals processing impacts. The results of this analysis also reveal some of the subjectivities associated with applying specific SA models at the STL-SS. For instance, Positive Matrix Factorization results are very sensitive to both the fitting species and number of factors selected by the user. Conversely, Chemical Mass Balance results are sensitive to the source profiles used to represent local metals processing emissions. Additionally, the different SA approaches predict different impacts for the same source on a given day, with correlation coefficients ranging from 0.034 to 0.65 for gasoline vehicles, -0.54-0.48 for diesel vehicles, -0.29-0.81 for dust, -0.34-0.89 for biomass burning, 0.38-0.49 for metals processing, and -0.25-0.51 for SOC. These issues emphasize the value of using several different SA techniques at a given receptor site, either by comparing source impacts predicted by different models or by using an ensemble-based technique.