PM2.5 Forecasting with Machine Learning: An Analysis of Meteorological and Atmospheric Factors

Abstract

This thesis examines the efficacy of 13 machine learning and deep learning models for hourly PM2.5 forecasting in Debrecen, Hungary, utilizing meteorological and atmospheric data from January 2024. The research utilizes a four-stage workflow — baseline modeling, feature engineering, hyperparameter tuning, and ensemble optimization — aimed at isolating and quantifying the impact of each element on prediction accuracy. The findings indicate that feature engineering, especially lag features, rolling statistics, and thermodynamic inversion indicators, yielded the most substantial enhancement, decreasing RMSE by 65% and increasing R² from 0.59 to 0.94. Regularized linear models like Lasso attained competitive accuracy with meticulously crafted features; however, deep learning models were constrained by the limited dataset size. SHAP-based interpretability analysis validated that the models identified physically significant atmospheric correlations. The thesis presents a replicable, modular pipeline suitable for air quality forecasting in data-constrained environments.