Ensemble-trained source apportionment of fine particulate matter and method uncertainty analysis

摘要

An ensemble-based approach is applied to better estimate source impacts on fine particulate matter (PM_(2.5)) and quantify uncertainties in various source apportionment (SA) methods. The approach combines source impacts from applications of four individual SA methods: three receptor-based models and one chemical transport model (CTM). Receptor models used are the chemical mass balance methods CMB-LGO (Chemical Mass Balance-Lipschitz global optimizer) and CMB-MM (molecular markers) as well as a factor analytic method, Positive Matrix Factorization (PMF). The CTM used is the Community Multiscale Air Quality (CMAQ) model. New source impact estimates and uncertainties in these estimates are calculated in a two-step process. First, an ensemble average is calculated for each source category using results from applying the four individual SA methods. The root mean square error (RMSE) between each method with respect to the average is calculated for each source category; the RMSE is then taken to be the updated uncertainty for each individual SA method. Second, these new uncertainties are used to re-estimate ensemble source impacts and uncertainties. The approach is applied to data from daily PM_(2.5) measurements at the Atlanta, GA, Jefferson Street (JST) site in July 2001 and January 2002. The procedure provides updated uncertainties for the individual SA methods that are calculated in a consistent way across methods. Overall, the ensemble has lower relative uncertainties as compared to the individual SA methods. Calculated CMB-LGO uncertainties tend to decrease from initial estimates, while PMF and CMB-MM uncertainties increase. Estimated CMAQ source impact uncertainties are comparable to other SA methods for gasoline vehicles and SOC but are larger than other methods for other sources. In addition to providing improved estimates of source impact uncertainties, the ensemble estimates do not have unrealistic extremes as compared to individual SA methods and avoids zero impact days.