Advanced Air Quality Forecasting Using an Enhanced Temporal Attention-Driven Graph Convolutional Long Short-Term Memory Model With Seasonal-Trend Decomposition

Abstract

This study presents the Improved Residual Spatial Attention-Temporal Convolutional Network (IRSA-TCN), an advanced framework for enhancing air quality forecasting across multiple pollutants. By integrating Residual Spatial Attention (RSA) with Graph Convolutional Long Short-Term Memory (GCLSTM) and Seasonal-Trend decomposition using Loess (STL), the IRSA-TCN model effectively captures intricate spatial and temporal patterns inherent in environmental data. The model demonstrates significant improvements in predictive accuracy, achieving reductions of 10-15% in error metrics such as Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) compared to traditional forecasting methods. Specifically, the IRSA-TCN model excels in predicting key pollutants, including nitrogen dioxide (NO2), carbon monoxide (CO), benzene, and nitrogen oxides (NOx), showcasing its capability to account for seasonal variations and complex interdependencies among pollutants. The findings underscore the model’s potential as a robust tool for environmental monitoring and management, providing actionable insights for policymakers and stakeholders. Future research will focus on enhancing the model’s adaptability through multi-scale spatial attention mechanisms and exploring hybrid architectures with Graph Neural Networks (GNNs) to further refine its applicability in real-time air quality forecasting scenarios. This work significantly contributes to the academic discourse on advanced analytical techniques in environmental science.

Keywords

• Air quality
• improved residual spatial attention
• temporal convolutional network
• seasonal-trend decomposition with loess
• reptile search algorithm