Effective Air Quality Index (AQI) prediction supports global health, the domestic economy, and the ecosystem. Aiming at learning spatiotemporal attributes from past air quality statistics, a Robust Bootstrapped Convolutional Neural Network with a Long Short-Term Memory network (RBCNN-LSTM) was designed to predict AQI and its uncertainties. Nonetheless, the effect of various attributes on the predicted outcomes at different intervals was not analyzed, especially to predict PM2.5 concentration. Hence, this article develops an Attention-based RBCNN-LSTM (A-RBCNN-LSTM) network framework to improve the prediction of AQI and their uncertainties according to the forecasting of PM2.5 concentration over the next few days. This framework adopts the attention strategy with the RBCNN-LSTM for determining the significance of each attribute and allocating corresponding weights to all attributes and applies a multi-view fusion by sharing the weights across the views in all LSTM units to obtain the correlation between PM2.5 concentrations and other attributes. Such correlations are learned by the bootstrapped convex-CNN to predict the air quality and its uncertainties.The results show that the proposed method based on the A-RBCNN-LSTM has more competent to enhance the air quality forecasting that traditional state of art methods in terms of accuracy,precision,recall,F-Measure