Використання машинного навчання для прогнозування якості атмосферного повітря міста Вінниці

Abstract

This paper presents an intelligent system for forecasting PM2.5 fine particulate matter concentration in the atmospheric air of Vinnytsia city using ensemble machine learning methods. The study encompasses the development of an automated ETL pipeline for data collection and preprocessing, exploratory time series analysis, and training of a gradient boosting ensemble (LightGBM, XGBoost, CatBoost) with Bayesian hyperparameter optimization. The bestperforming LightGBM model achieved R² = 0.90 at the t+1 forecast horizon and R² = 0.45 at t+24, confirming the effectiveness of the proposed approach.