Multivariate Meteorological Time Series Forecasting with Transformer-based Models

Air pollution is a critical issue globally, particularly in developing countries with outdated technologies and lenient regulations. Traditional air pollutant detection methods like auto-regressive and seasonal decomposition models often fail to handle the complexity of environmental data effectively. In response, this study explores the use of advanced transformer-based models, including NLinear, DLinear, Informer and iTransformer, for predicting key air pollutants such as CO, NO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>, 0<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf>, and particulate matter. We compare these deep learning models to conventional time series methods to establish a robust multivariate forecasting framework for atmospheric conditions. Enhancements include employing a bagging strategy to improve model performance. This research contributes a transformer-based forecasting framework, demonstrating its potential to improve air pollution prediction in environmental science.