2026-06-23T22:14:29.809Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:1:11943 2026-05-27T12:47:42Z 1

Advanced AI approaches for understanding benzene pollution in urban atmosphere 2025 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/1/11943 https://www.croris.hr/crosbi/publikacija/prilog-skup/906627 G. Jovanović S. Herceg Romanić S. Davila I. Bešlić G. Pehnec T. Bezdan M. Perišić A. Stojić Benzene is a toxic, carcinogenic volatile organic compound (VOC), recognized as a Group 1 human carcinogen by the IARC. Despite often being present at low ambient levels, prolonged exposure can cause serious health effects. In urban settings, benzene primarily originates from traffic, fossil fuel combustion, and industrial processes, and also contributes to the formation of O3 and secondary organic aerosols. This study introduces an advanced artificial intelligence (AI) framework for understanding the interactions between benzene and co-occurring air pollutants in an urban environment. By integrating machine learning, (ML) metaheuristic optimization, and explainable AI techniques (XAI), the proposed model captures complex, nonlinear relationships within high-resolution environmental datasets. The analysis draws on seven years (2017–2023) of hourly data from an urban background monitoring station in northern Zagreb, Croatia, encompassing concentrations of benzene, SO₂, NO₂, CO, and O₃, alongside 29 meteorological variables obtained from the Global Data Assimilation System (GDAS). Benzene was selected as the target variable due to its toxicological and regulatory importance, while other pollutants and meteorological variables served as predictors. After testing multiple ML algorithms, the optimal model was fine-tuned using metaheuristic optimization. To interpret the model’s behavior, XAI techniques, specifically SHAP (SHapley Additive exPlanations) and SAGE (Shapley Additive Global importancE), were applied, offering detailed insight into the individual and combined effects of predictors. Findings revealed that co-pollutants, particularly CO and NO₂, are key drivers of benzene variability. SHAP values indicated synergistic effects at elevated concentrations, suggesting common combustion-related sources. These results highlight the added value of XAI tools in unveiling hidden dependencies among pollutants, offering a transparent, data-driven foundation for refining emission inventories and shaping targeted mitigation strategies.Results showed that co-pollutants, particularly CO and NO2 emerged as dominant predictors of the benzene patterns in air. SHAP analysis revealed that low concentrations of both gases reduce predicted benzene more effectively in combination than individually, while high levels of both suggest episodic coemissions, likely linked to combustion sources. These results highlight the added value of XAI tools in unveiling hidden dependencies among pollutants, offering a transparent, data-driven foundation for refining emission inventories and shaping targeted mitigation strategies. conferenceObject AIR PROTECTION 2025 - Book of Abstracts