Implementing constrained multi-time approach with bootstrap analysis in ME-2: An application to PM2.5 data from...

Advanced receptor models have been recently developed and tested in order to improve the resolution of apportionment problems reducing rotational ambiguity of results and aiming at identifying a larger number of sources. In particular, multi-time model is a factor analysis method able to compute source profiles and contributions using aerosol compositional data with different time resolutions. Unlike traditional factor analysis, each measured value can be inserted into multi-time model with its original time schedule, thus all temporal information can be effectively used in the modelling process. In this work, multi-time model was expanded in order to impose constraints on modelled factors aiming at improving the source identification. Moreover, as far as we know for the first time, a suitable bootstrap technique was implemented in the multi-time scheme to estimate the uncertainty of the final constrained solutions. These implemented approaches were tested on a PM2.5 (particulate matter with aerodynamic diameter <2.5 μm) dataset composed of 24-h samples collected during one year and hourly data sampled in parallel for two shorter periods in Florence (Italy). The daily samples were chemically characterised for elements, ions and carbonaceous components while elemental concentrations only were available for high-time resolved samples. The application of the advanced model revealed the major contribution from traffic (accounting for 37% of PM2.5 as annual average) and allowed an accurate characterisation of involved emission processes. In particular, exhaust and non-exhaust emissions were identified. The constraints imposed in the continuation run led to a better description of the factor associated to nitrates and also of biomass burning profile and the bootstrap results gave useful information to assess the reliability of source apportionment solutions. Finally, the comparison with the results computed by ME-2 base model applied to daily and hourly compositional data separately demonstrated the advantages provided by the multi-time approach.