Short-term source apportionment of fine particulate matter with time-dependent profiles using SoFi Pro: exploring the reliability of rolling positive matrix factorization (PMF) applied to bihourly molecular and elemental tracer data

Abstract

Positive matrix factorization (PMF) has been widely used to apportion the sources of fine particulate matter (PM2.5) by utilizing PM chemical speciation data measured at the receptor site(s). Traditional PMF, which typically relies on long-term observational datasets of daily or lower time resolution to meet the required sample size, has its reliability undermined by changes in source profiles; thus, it is inherently ill-suited for apportioning sporadic sources or ephemeral pollution events. In this study, we explored short-term source apportionment of PM2.5 using a set of bihourly chemical speciation data over a period of 37 d in the winter of 2019–2020. PMF run with campaign-wide data as input (PMFref) was initially conducted to obtain reference profiles for the primary source factors. Subsequently, short-term PMF analysis was performed using the Source Finder Professional (SoFi Pro). The analysis sets a window length of 18 d and constrained the primary source profiles using the a-value approach embedded in SoFi Pro software. Rolling PMF was then conducted with a fixed window length of 18 d and a step of 1 d using the remaining dataset. By applying the a-value constraints to the primary sources, the rolling PMF effectively reproduced the individual primary sources, as evidenced by the slope values close to unity (i.e., 0.9–1.0). However, the estimation for the firework emission factor in the rolling PMF was lower compared with PMFref (slope: 0.8). These results suggest the unique advantage of short-term PMF analysis in accurately apportioning sporadic sources. Although the total secondary sources were well modeled (slope: 1.0), larger biases were observed for individual secondary sources. The variation in source profiles indicated higher variabilities for the secondary sources, with average relative differences ranging from 42 % to 173 %, while the primary source profiles exhibited much smaller variabilities (relative differences of 8 %–26 %). This study suggests that short-term PMF analysis with the a-value constraints in SoFi Pro can be utilized to apportion primary sources accurately, while future efforts are needed to improve the prediction of individual secondary sources. Additionally, future rapid source apportionment analysis can benefit from utilizing a library of source profiles derived from existing measurement data, thereby significantly reducing the time lag associated with receptor modeling source apportionment techniques.