Forecasting Air Poluttion using Transformer Models

Abstract

This thesis explores the application of transformer-based models for forecasting air pollution levels, specifically PM2.5 and PM10, in Oslo, Norway. The primary objectives were to evaluate the effectiveness of transformers in integrating spatial and temporal dimensions, compare their performance against traditional and modern forecasting methods, and explore the role of synthetic data in enhancing model accuracy. Through experiments, PatchTST emerged as the most effective transformer model, outperforming other transformer based models and the MLP based TSMixer. The study revealed that granular, hourly data significantly improves forecasting accuracy. Recommendations include optimizing model parameters, incorporating additional meteorological variables, and extending evaluations to enhance generalizability. Future research directions suggest exploring advanced transformer variants and integrating multi-modal data for long-term and more accurate air pollution forecasting.