A Lightweight Transformer-Based Model for Multivariate Air Pollution Forecasting: A Case Study on CO and NO₂ Prediction

Taghandiki, Kazem; Bohloli, Ali

doi:10.48301/vjtve.2026.536830.1132

A Lightweight Transformer-Based Model for Multivariate Air Pollution Forecasting: A Case Study on CO and NO₂ Prediction

Articles in Press

Document Type : Scientific Review

Authors

Kazem Taghandiki ¹

Ali Bohloli ²

¹ Department of Computer Engineering, Technical and Vocational University (TVU), Tehran, Iran

² Assistant Professor, Department of Computer Engineering, Technical and Vocational University (TVU), Tehran, Iran

10.48301/vjtve.2026.536830.1132

Abstract

Air pollution is a major environmental challenge in densely populated cities, with serious impacts on human health and ecosystems. Accurate prediction of pollutant concentrations such as carbon monoxide (CO) and nitrogen dioxide (NO₂) is essential for environmental planning and air quality management. In this study, we propose an innovative and lightweight framework based on the Transformer architecture for simultaneous prediction of CO and NO₂ using time-series data. The methodology includes data cleaning, handling noisy values, normalization, and creating time windows. The model uses historical data in 24-hour sliding windows to predict pollutant levels for the next hour. Experimental results show that our model achieves high accuracy while being computationally efficient, outperforming conventional models such as LSTM and CNN-LSTM. With an R² score above 0.92 for both pollutants, the model demonstrates strong performance on multivariate environmental data. Additionally, comparative analysis highlights the potential of this lightweight Transformer model for use in real-time air quality monitoring systems and smart city applications.

Keywords

Air pollution

Transformer

Time series

CO prediction

NO₂ prediction

Subjects