A Conv-Attention Seasonal Aggregation Network (CASA-Net) for PM2.5 Concentration Forecasting Using US EPA Monitoring Data

Abstract

The prediction of fine particulate matter, or fine particulate (PM₂.₅), is very important for air quality management and protecting human health. Current deep-learning approaches, however, cannot capture the local time-specific interactions and spatiotemporal correlations across long time periods, or are computationally too expensive to effectively model the seasonal periodic variations. This study builds on these insights to suggest the Convolutional Attention Seasonal Aggregation Network (CASA-Net) combining Conv1D feature extraction, Bidirectional Long Short-Term Memory (BiLSTM), Multi-Head Self-Attention (MHSA), residual learning, and sinusoidal Day-of-Year (DOY) seasonal embeddings. The proposed model was tested against the day-by-day PM₂.₅ data for the Yorkshires provided by the United States Agency for Protection of the Environment (EPA) in the Air Quality System (AQS). The proposed model was tested using the Daily PM₂.₅ data for Yorkshires from the United States Agency for Protection of the Environment (EPA) Air Quality System (AQS) with 133,625 sliding window observations from 543 quality-controlled stations. To prevent temporal leakage and provide reliable model assessment, the block cross-validation approach with a train--validation--test split of 70:15:15, which is geographically diverse, was used. The experimental results show that CASA-Net can significantly reduce RMSE, MAE, and MAPE values from machine learning-based models and the latest deep learning baseline by getting RMSE of 3.1847 µg/m³, MAE of 1.7214 µg/m³, R² of 0.8134, MAPE of 10.84%, and a Pearson correlation coefficient (r) of 0.9041. CASA-Net outperforms Informer in terms of the RMSE reduction (18.22%) and the increase in R² value (0.1503), as well as fewer required parameters (4.9× fewer) and faster training and inference time, and estimated fewer required GFLOPs per forward pass (8.8× fewer). Moreover, the robustness, computational efficiency, and application for next-day PM (2.5) forecasting and air quality decision-making are also demonstrated by applying statistical significance testing, uncertainty quantification, ablation studies, feature attribution, and attention analysis to CASA-Net.