Multi-Pollutant Air Quality Forecasting Using Bidirectional Attention and Multi-Scale Temporal Networks

Chee-Onn Chow^1*Zi-Ang Xie¹Joon Huang Chuah^1,2Wong Jee Keen Raymond¹

• ¹University of Malaya, Kuala Lumpur, Malaysia
• ²Southern University College, Skudai, Malaysia

Accurate air quality prediction is essential for urban environmental governance and public health protection. However, many existing deep learning models are limited in their ability to capture multi-scale temporal dependencies and synergistic interactions between pollutants. This paper proposes a novel deep learning framework named Enhanced Bidirectional Attention Multi-scale Temporal Network (EBAMTN), which integrates a multi-scale Temporal Convolutional Network (TCN) with linear attention, an enhanced bidirectional Long Short-Term Memory (Bi-LSTM), and multi-head attention mechanisms. Under a multi-task learning paradigm, the model jointly forecasts P M 2.5 and P M 10 concentrations by leveraging shared temporal representations. The proposed model incorporates a parallel multi-scale TCN module equipped with linear attention to dynamically fuse features across different receptive fields. The BiLSTM component, integrated with multi-head attention, captures bidirectional temporal dependencies and highlights informative time steps with enhanced granularity. A gated fusion layer further enhances the integration of sequential and contextual features while improving interpretability. Extensive experiments on realworld datasets from Guangzhou, Beijing, and Chengdu demonstrate that EBAMTN consistently outperforms existing models, achieving R² values above 0.94 for both pollutants. These results underscore robustness, generalization ability, and deployment potential of the framework for real-time and scalable air quality forecasting in complex urban scenarios.