High-resolution real-time air quality forecasting can alert decision-makers and residents about forthcoming air pollution events and refine air quality management. The Environmental Protection Administration in Taiwan has deployed numerous low-cost air quality microsensors near industrial zones lately to facilitate local air quality monitoring. Nevertheless, the frequent occurrence of missing sensor data due to problems of mobile transmission, frontend/backend device malfunction, or other unforeseen issues would raise difficulty in making quick responses to air pollution incidents. This study proposed a hybrid deep learning model (AE-CNN-BP) collaborating an Autoencoder (AE), a Convolutional Neural Network (CNN), and a Back Propagation Neural Network (BPNN) to effectively extract crucial features from big data for making successive high-spatiotemporal-resolution forecasts of PM2.5 concentrations 4 h ahead. The proposed model was trained and tested in three industrial zones densely installed with microsensors in Kaohsiung City of Taiwan. A high pollution incident was selected to evaluate model performance. The results show that the proposed model could reliably produce nice high-spatiotemporal-resolution forecasts for 12 air quality monitoring stations and 485 microsensors, with Coefficient of Determination (R2) values and Root Mean Squared Error (RMSE) of 0.82 (0.76) and 11.05 (12.75) μg/m3 in the training (testing) stage, respectively. For the selected incident, the Mean Absolute Percentage Error (MAPE) values of the proposed model were 22.3% and 27.1% at T+1 and T+4, respectively. This study demonstrates that the proposed deep learning model based on ensemble datasets of sparsely distributed monitoring stations and densely deployed microsensors can offer reliable high-spatiotemporal-resolution air quality forecasts, benefiting environmental studies and informed policymaking by accounting for local-scale variations in PM2.5 concentrations.
