Evaluation of a statistical forecast model for size-fractionated urban particle number concentrations using data from five European cities

In this study we evaluated a forecast model for urban aerosol number concentrations against measurements made at five European cities: Helsinki, Stockholm, Copenhagen, Leipzig, and Athens. This model requires a learning data set with particle number concentrations, traffic densities and local meteorology. Additionally, in the forecasting process it requires the same parameters from the past week and forecasted values of weather and traffic. The model performance was tested based on the R2, index of agreement (IA), mean square error (MSE), and bias. We tested three modelling approaches: one with fixed parameterisation and two with optimisations based on either the Deviance or the Akaike Information Criterion. Based on the hourly one-day forecasts at the urban background sites the IA ranged from 0.65 to 0.80 for accumulation mode particles and from 0.68 to 0.87 for ultrafine particles. The model performance was the best for Helsinki and Stockholm and the worst for Leipzig and Copenhagen. The main reason is the more pronounced diurnal variation in Helsinki and Stockholm. Another reason is that the traffic data for Leipzig and Copenhagen were not as complete as for the other cities. The three approaches yielded similar model performances, hence we conclude that the simplest one based on a fixed parametrisation is to be preferred.