Development and assessment of a receptor source apportionment model based on four nonnegative matrix factorization algorithms

Abstract

This study developed a receptor model, comprising four non-negative matrix factorization algorithms: the multiplicative update method; the optimal gradient method; the highly efficient, monotonic, fixed-point method; and the conjugate gradient method. The feasibility and performance of the developed model for emission source apportionment were assessed, using both a synthetic dataset, and an ambient PM2.5 dataset. The results from the US EPA's positive matrix factorization (PMF) 5.0 model were used for the assessment. Modeled results for the synthetic data showed that the range of factor contributions to most matrix elements solved by the four algorithms covered actual values. Modeled results, using the ambient dataset as the input, showed that the four algorithms in the developed model, and the PMF model, identified the same eight emission sources, and apportioned similar source contributions to PM2.5. Comparisons between the modeled organic carbon, and the elemental carbon source apportionments and radiocarbon measurements, suggested that combined application of multiple algorithms could satisfactorily apportion emission source contributions for one, or a few, specified samples among a receptor dataset, thus confirming the excellent source apportionment ability of the proposed model.