A CEEMDAN-GNN-transformer hybrid model for air quality index forecasting: a case study of Chang’an town, Dongguan, China

Air pollution has emerged as a pressing global environmental issue, and accurate forecasting plays a critical role in environmental governance and public health protection. This study proposes an enhanced air quality forecasting model based on a hybrid CEEMDAN-GNN-Transformer architecture, and conducts an empirical analysis using data from Chang’an Town, Dongguan, China. The proposed model first employs Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to extract multi-scale temporal features and mitigate non-stationary noise in the data. Then, a Graph Neural Network (GNN) is applied to capture the spatial dependencies among various air pollutants. Finally, a Transformer model is utilized to model complex temporal dependencies and improve the capture of long-term trends. The research uses historical air quality monitoring data from 2015 to 2024, including concentrations of PM 2.5 , PM 10 , SO 2 , CO, NO 2 , and O 3 as input features, with the Air Quality Index (AQI) as the prediction target. Model performance is enhanced through ablation studies and hyperparameter tuning, and is compared against several mainstream baseline models. Experimental results demonstrate that the proposed CEEMDAN-GNN-Transformer model outperforms traditional approaches in terms of MAE, MSE, and R 2 metrics, achieving superior prediction accuracy and robustness. This study not only contributes to the theoretical advancement of air quality forecasting methodologies but also provides a more precise predictive tool for environmental management and public health risk prevention, offering significant practical value.