Prediction of urban ultrafine particle emission fluxes using generalized additive models

Abstract

Ultrafine particles (UFP) are abundant in urban atmospheres. To assess the strength and temporal variation of urban UFP emission sources, information on the surface-atmosphere exchange, i.e. the turbulent vertical flux of particles, is vital. A three-year time series of UFP emission fluxes (FUFP) observed at an urban site in Berlin, Germany, using the eddy covariance technique was utilized to develop and evaluate generalized additive models (GAM) for FUFP. GAM allow to account for non-linear relationships between response and predictor variables. Two separate models for summer and winter were developed. The predictors that most strongly influenced modelled FUFP in the summer model were traffic activity, friction velocity, land use, air temperature and PM10 concentration, whereas the winter model additionally incorporated relative humidity. The GAM were evaluated by ten-fold cross-validation for the first two study years, and by predicting the third year based on the model trained with observational data of the first two years. The coefficients of determination of the two validation methods were R2 = 0.52 (uncertainty of −47 to 88% for FUFP) and R2 = 0.48 (−45 to 82% for FUFP) for the winter model, whereas the summer model yielded R2 = 0.48 and 0.44 (uncertainty of −51 to 102%). GAM were shown to successfully capture the non-linear relationships between predictor variables and FUFP for the three-year data set at this urban site.