Machine Learning Driven Prediction and Analysis of NO2 and its Catalyst Based Reduction in Urban Environments

Abstract This study employed machine learning (ML) to predict nitrogen dioxide (NO₂) pollution in Marylebone Road, London a high-traffic urban corridor using historical data from 2015 to 2022 to forecast concentrations for the period January 2023 to January 2025. Four ML models were developed and evaluated: Linear Regression, Random Forest, LightGBM, and an Ensemble Stacking model. These models incorporated meteorological and pollutant data and were assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²). The Ensemble Stacking model outperformed the others, achieving an R² of 0.9723, MAE of 3.91 µg/m³, and RMSE of 6.25 µg/m³. In comparison, the Linear Regression model showed the lowest performance (R² = 0.8307, MAE = 11.55, RMSE = 15.45), while Random Forest (R² = 0.9232) and LightGBM (R² = 0.9719) demonstrated intermediate accuracy. The best-performing ensemble model was further used to simulate NO₂ trends with and without titanium dioxide (TiO₂) catalyst intervention, assuming a 28% NO₂ reduction. Temporal analysis revealed that NO, NO₂, and NOₓ concentrations peaked during colder months (November–January) and weekdays. Correlation analysis showed a weak negative relationship between NO₂ and ozone (O₃) (R² = 0.26), moderate positive correlations with black carbon (BC) (R² = 0.597) and sulfur dioxide (SO₂) (R² = 0.654), and a very weak positive correlation with particulate matter (PM2.5) (R² = 0.143). The study concludes that ensemble stacked ML models are effective for predicting NO₂ concentrations and that TiO₂ nanocatalyst interventions hold promise for reducing NO₂, BC, and SO₂ levels in urban environments.