Abstract Urban air pollution is a health hazard linked to anthropogenic emissions. Reliable evaluation of changes in pollutants due to altered emissions requires considering meteorological and other variability influencing concentrations. Here, a combination of ensemble learning, competitive learning and unsupervised clustering is proposed and applied to leverage the change analysis of particulate matter (PM2.5) and other pollutants. Machine Learning (ML) algorithms Random Forest (RF) and Self-Organizing Map (SOM) were trained with historical meteorological data, pollutant concentrations and traffic indicators. The importance of different variables for local PM2.5 was determined with RF. SOM was configured for multivariable cluster analysis. The trained SOM enabled predicting a cluster for new data representing conditions with shifted anthropogenic activity. The prediction forms a benchmark for the analysed period with maximized meteorological similarity, which facilitates identifying changes in ambient pollutants due to changed emissions. The method was applied to data from the start of COVID-19 pandemic, 3/2020, when emissions suddenly decreased. For measurements from Helsinki, Finland, the SOM yielded a statistically significant change in PM2.5 (−0.7%), NO2 (−33%) and O3 (+17%). Comparing data from 3/2020 to data from 3/2017–2019 produced different results (PM2.5 −1.7%, NO2 −37%, O3 −4.0%). Statistical indicators confirmed better compatibility between the analysed period and its benchmark when using the SOM prediction instead of calendar-based selection: Average RMSRE was 19%-points lower and Willmott’s dr 41% higher with SOM than with 3/2017–2019. Based on the case study and method evaluation, using ML for multivariate analysis of changed air pollution is feasible and yields meaningful results.