Air Quality Index (AQI) variations possess a dual temporal and spatial nature, while exhibiting nonlinear, non-stationary, high-noise and high-volatility characteristics. Existing AQI prediction methods struggle to fully extract its spatiotemporal features and achieve stable and effective forecasting. Therefore, a spatiotemporal data-driven AQI forecasting method based on multi-scale 3D-CNN-CBAM is proposed in this paper. The method can effectively identify and extract spatiotemporal information embedded in AQI data by constructing a 3D spatial tensor. First, considering the spatial domain weight between the local and neighboring regions, mutual information (MI) is applied to select influencing factors of AQI. Second, the multivariate empirical mode decomposition (MEMD) and sample entropy (SE) methods are used to decompose the historical data and the influencing factor series into more regular high-frequency series, low-frequency series, and trend terms, respectively. Third, based on the decomposed regional AQI data, atmospheric pollutant concentration values, and meteorological factor variables, subsequence three-dimensional spatial tensors are constructed to reflect their spatiotemporal evolution. Furthermore, a three-dimensional convolutional attention mechanism (3D-CNN-CBAM) network model is developed to predict the subsequence. This model effectively extracts key spatiotemporal correlation features between AQI and its influencing factors while reducing the interference of noise information on the fitting performance. Finally, the destination AQI prediction results are obtained by integrating the subsequence predictions. The proposed method is applied to the daily AQI prediction in the Yangtze River Delta urban agglomeration from 2019 to 2023. The results show that this method is suitable for predicting air quality index with spatiotemporal attributes and exhibits higher predictive accuracy and applicability compared to existing methods.