Forecasting PM10 hourly concentrations in northern Italy: Insights on models performance and PM10 drivers through self-organizing maps

Abstract

A linear and an artificial neural network (ANN) statistical model have been developed and validated for short-term forecasting of PM 10 hourly concentrations in the city of Brescia (Italy). PM 10 observed concentrations were biased by less than 1% by each model, though the ANN outperformed the linear model, as exhibiting NRMSE of 0.48 vs. 0.53, and r 2 of 0.57 vs. 0.48. The self-organizing maps (SOMs) showed that both models predictions exhibit the same clustering as the observations, with the ANN at worst capable of under-estimating clustered PM 10 peak concentrations by 5.8 μg/m 3 . In Brescia, PM 10 most critical conditions were detected in wintertime in the early morning or late afternoon under unfavourable meteorological conditions, i.e. reduced advection enhancing PM 10 stagnation, and lack of precipitations capable of reducing PM 10 resuspension. Under these conditions, PM 10 accumulation is driven by local anthropogenic emissions ascribing to two main sources: heating plants, responsible of emissions of primary PM 10 (mostly PM 2.5 , likely resulting from wood and biomass burning); and road traffic (basically diesel vehicles), mainly responsible of emissions of secondary PM 10 precursors (mostly NO x ), and secondly of primary PM 10 emissions. The SOM analysis clearly indicated that PM 10 most critical conditions are driven by the secondary rather primary PM 10 component. • A linear and an ANN model developed and validated to forecast PM 10 1-h concentrations in Brescia (Italy). • The ANN outperforms the linear model, with NRMSE of 0.48 vs. 0.53, and r 2 of 0.57 vs. 0.48 • SOMs proved to recognise the cluster structure of all drivers controlling PM 10 concentrations. • Heating plants and road traffic were confirmed as main anthropogenic emission sources. • The SOM analysis indicated that PM 10 most critical conditions are driven by the secondary rather primary PM 10 component.