Deep Aggregation seq2seq Network With Time Feature Fusion for Air Pollutant Concentration Prediction in Smart Cities

Abstract

Air pollution poses significant risks to environmental quality and public health. Precise forecasting of air pollutant concentrations is crucial for safeguarding public health. The emission and diffusion of air pollutants is a dynamic process that changes over time and has significant seasonal characteristics. By leveraging time attributes such as month, day of the month, and hour, the precision and dependability of forecasting models can be enhanced. Therefore, this paper proposes a deep aggregation seq2seq network with time feature fusion for air pollutant concentration prediction. This network first effectively integrates temporal feature encoding with historical air pollutant concentration data through a cross attention network, and then excavates hidden features through deep aggregation seq2seq network. The encoder part of the network can extract the temporal correlation of fusion features, while the decoder part can generate them through recursive aggregation. The future prediction values fully utilize the local features and overall recursion of historical information, improving the accuracy of prediction. In this study, we conduct simulations on the actual datasets of PM2.5 and SO2, two air pollutants, in Beijing's Changping and Shunyi. The findings reveal that our model reduces the Mean Absolute Error by 5% to 10% compared to existing state-of-the-art models.