Predicting urban air pollution time series with machine learning models

Urban air pollution time series exhibit strong temporal variability, including long-term trends, daily and seasonal cycles, and infrequent but systematic events such as holidays. These characteristics make Machine Learning based modelling sensitive to preprocessing choices and the length of the training period. In this work, we analyze how detrending and multi-year training influence the ability of Machine Learning models to reproduce recurrent temporal patterns and rare events. Using several years of nitrogen dioxide and ozone measurements from monitoring stations in Berlin, combined with meteorological and traffic data, we compare model behaviour under different training setups. The trained models reach high prediction accuracy for NO2 and O3 with normalized mean biases of 4% and 3% respectively. The results show that removing long-term trends improves the stability of learned temporal patterns, while short training periods limit the representation of rare but systematic events. Additionally, a traffic scenario analysis for NO₂ is presented, including simulations with 50% and 0% traffic levels, to investigate how reductions in traffic influence the predicted NO₂ concentrations.