Short-term air quality prediction using a multi-scale attention fusion model with 3DIGAT-CBAM-BiLSTM based on spatio-temporal correlation

Air Quality Index (AQI) prediction is crucial for environmental management and public health. However, most existing studies focus on single site modeling, neglecting the complex spatial correlations of meteorological factors and air pollutants. Therefore, a multi-scale spatio-temporal prediction model, 3DIGAT-CBAM-BiLSTM, is proposed to fully capture the spatio-temporal evolution characteristics of AQI. To reduce the interference of redundant information, the Maximum Information Coefficient and Dynamic Time Series Trend Correlation Method are employed to select the neighboring sites and influencing factors that are highly correlated with the AQI of the target site. The original air quality data is decomposed and reconstructed into high-frequency, low-frequency, and trend-term subsequences using Multivariate Variational Mode Decomposition and Sample Entropy to enhance prediction accuracy. To forecast the three-dimensional spatial tensors of these reconstructed subsequences based on time steps, monitoring sites, and influencing factors, we propose the 3DIGAT-CBAM-BiLSTM model. The spatial dependencies between sites are effectively captured by the Improved Graph Attention Network, which constructs a graph adjacency matrix based on MIC and geographic distance. Meanwhile, the Convolutional Block Attention Mechanism enhances the focus on important sites and features by combining channel and spatial attention. Furthermore, the Bidirectional Long Short-Term Memory network extracts global temporal patterns. The experimental results on the Beijing dataset show that the proposed model achieves a relative reduction of 8.53 % in RMSE and 5.83 % in MAE compared with the optimal baseline model, demonstrating clear performance improvements and offering a novel approach for modeling complex spatio-temporal data.