Sensitivity Analysis for a Statistical Prediction Model Utilized for Particle Number Concentrations with a Focus on Submicron Size Fractions

Abstract. Long-term particle number size distribution (PNSD) observations are sparse and costly to maintain, limiting their use for prediction and process attribution. We develop a Bayesian regression with autoregressive (AR) errors and sequential prediction, estimating posteriors via Markov chain Monte Carlo (MCMC). The study uses multi-year data set from 1996 till 2024, hourly observations from three Finnish SMEAR stations—SMEAR I (Värriö), SMEAR II (Hyytiälä), and SMEAR III (Helsinki)—with predictors drawn from routinely measured meteorology and trace gases. The developed model was applied on 5 different scenarios with different aims.Scenario I showed that expanding the historical training period improve or stabilize prediction performance. Scenario II demonstrated that predictive performance depends strongly on the training year. Scenario III showed that adding increasingly old observations provide no additional benefit once recent observations were already included. Scenario IV demonstrated that progressively updating the model with more recent data generally improved predictive performance. Scenario V showed that increasing the length of consecutive training windows generally reduced prediction errors. Addtionally, the model captures both diurnal and seasonal variability and achieves high coefficients of determination (R2) at 1 h resolution indicating that most observed variability is explained. These results suggest a successful prediction for particle number concentrations. The potential of this approach is to make future forecasting for particle number concentrations for the up coming years.