Clustering of emission–dispersion dynamics in a state space defined by pollution sources and meteorological variables

The complex relationship between air pollution sources, the meteorological conditions governing their dispersion, and the resulting pollutant concentrations is an interesting scientific topic with significant implications for air resource management. A widely adopted approach to exploring these interactions is the interpretation of statistical models which simulate them. Dispersion conditions can vary markedly—from cold, calm nights with stable atmosphere to stormy periods characterised by intense turbulence and convection. Deeper insights can be obtained by interpreting statistical models trained on data subsets that correspond to specific emission–dispersion conditions. To achieve such distinct subsets, we present a robust and reproducible methodology for partitioning the multidimensional state space, defined by pollution sources and meteorological variables, into a large number of clusters at each monitoring location. Our methodology ensures a systematic and objective analysis across multiple stations. We analyse four years of hourly data, integrating high-resolution meteorological outputs from a numerical weather prediction model with traffic volume data used as proxies for emissions or precursors of NO, NO₂, NO_x, PM_2.5, and O₃ concentrations. These pollutants were observed at 85 air quality monitoring stations across Israel. The resulting clusters capture sub-daily temporal patterns that are indicative of distinct emission–dispersion scenarios in the region. We demonstrate that statistical models trained on these clustered subsets consistently outperform models trained on the full-period datasets. This highlights the value of our clustering approach in improving both predictive performance and scientific understanding of air pollution dispersion dynamics.