Spatio-attention embedded recurrent neural network for air quality prediction

Predicting the air quality index (AQI) has been regarded as a critical problem for environmental control management. Many factors over time and space may relate to the diffusion of pollutants. In other words, there exist very intricate spatio-temporal interactions among the characteristic for revealing diffusion of pollutants. Recently, some relevant works studied the topic of AQI prediction considering spatial and temporal correlations simultaneously, but most of them either ignore geospatially topological structures to learn spatio-temporal dependency or utilize sub-modules separately to encode the spatial and temporal information. Unfortunately, ignoring geospatially topological structures or correlations among spatial properties and temporal dependencies leads that the AQI prediction model cannot deal with the prediction task well. Although few existing works utilize an integrated model to encode the spatial and temporal information with geospatially topological structures, these works focus only on catching inter-station relationships among stations, which only can reflect the spatio-temporal dependency. In this paper, we propose a spatio-attention embedded recurrent neural network, called SpAttRNN, for air quality prediction by comprehensively leveraging dynamic spatio-temporal correlations among monitoring stations to tackle the intricate spatio-temporal interactions. To capture spatial properties, a self-loop-normalized adjacency matrix is utilized to enhance graph-based attention cells to learn the relationships between monitoring stations. Moreover, the proposed graph-based attention cell is embedded into a recurrent neural network to learn dynamic spatio-temporal correlations from multiple monitoring stations simultaneously. We conducted comprehensive experimental evaluations on two real-world air quality datasets collected from Beijing's real-time air quality monitoring system. The results demonstrate that our method outperforms other state-of-the-art methods significantly. In particular, SpAttRNN delivers up to 15% improvement in terms of the mean square error (MAE) on PM_2.5 prediction compared with other existing methods.