Machine learning analysis of PM1 impact on visibility with comprehensive sensitivity evaluation of concentration, composition, and meteorological factors

Abstract

Abstract This study introduces a novel approach to visibility modelling, focusing on PM1 concentration, its chemical composition, and meteorological conditions in two distinct Polish cities, Zabrze and Warsaw. The analysis incorporates PM1 concentration measurements as well as its chemical composition and meteorological parameters, including visibility data collected during summer and winter measurement campaigns (120 samples in each city). The developed calculation procedure encompasses several key steps: formulating a visibility prediction model through machine learning, identifying data in clusters using unsupervised learning methods, and conducting global sensitivity analysis for each cluster. The multi-layer perceptron methods developed demonstrate high accuracy in predicting visibility, with R values of 0.90 for Warsaw and an RMSE of 1.52 km for Zabrze. Key findings reveal that air temperature and relative humidity significantly impact visibility, alongside PM1 concentration and specific heavy metals such as Rb, Vi, and Cd in Warsaw and Cr, Vi, and Mo in Zabrze. Cluster analysis underscores the localized and complex nature of visibility determinants, highlighting the substantial but previously underappreciated role of heavy metals. Integrating the k-means clustering and GSA methods emerges as a powerful tool for unravelling complex mechanisms of chemical compound changes in particulate matter and air, significantly influencing visibility development.