The accurate understanding of the real-time dynamic evolution patterns of the air quality index (AQI) is crucial for air pollution control and urban public health governance. Therefore, we extend the AQI feature information from the temporal dimension to the spatio-temporal dimension by constructing a three-dimensional space tensor. Furthermore, a spatio-temporal forecasting network model based on the multi-scale three-dimensional-convolutional neural networks-convolutional block attention mechanism (3D-CNN-CBAM) is proposed to improve the AQI prediction accuracy. First, the effective screening is performed on the influencing factors data. Second, the AQI and its influencing factors are decomposed into subsequences with different modes. Then, based on the three aspects of time, spatial geographic location, and influencing factors, a three-dimensional spatial tensor is constructed to reflect the spatio-temporal characteristic evolution of AQI data. Third, a 3D-CNN-CBAM network model is designed to predict the subsequences, which can effectively extract the crucial spatio-temporal correlational features between the AQI and its influencing factors. By learning the importance of local AQI sequence features, the model can assign different weights to spatio-temporal information, thereby enhancing the influence of critical information and reducing the interference of redundant information. The proposed method is applied to the daily AQI prediction in the three major urban agglomerations from 2019 to 2024. The results show that the proposed method is suitable for predicting the AQI with spatio-temporal attributes and exhibits higher forecasting accuracy and applicability compared to existing methods.