Abstract: Accurate and reliable prediction of PM_2.5 concentrations is important for the public to effectively avoid air pollution and for governmental strategy development. However, due to the dynamic nature of atmospheric flows, the prediction of PM_2.5 concentration is characterized by great uncertainty and instability, making it difficult for a single model to efficiently extract spatio-temporal correlations. In this paper, a robust prediction system is proposed to realize accurate single-step, multi-step and trend prediction of PM_2.5 concentration. First, the article adopts a correlation analysis method to screen the meteorological and pollutant spatial information that can help predict the pollutant concentration in the target city. Then, the feature extraction capability of DenseNet is utilized to extract spatially relevant features from pollution and meteorological datasets from multiple cities; the ConvLSTM layer combines the temporal and spatial features of the pollutant data, and extracts the spatial and temporal features in order to achieve accurate pollutant prediction. Finally, the performance of the proposed prediction system is comprehensively evaluated by four accuracy indicators and three prediction experiments. In addition, the pilot study shows that the prediction system has good application prospects in early warning, regional prevention and control of air pollution, and its accuracy and stability are better than those of various baseline models.

Key words: deep learning, air pollution, DenseNet, pollutant concentration prediction