An adaptive spatio-temporal neural network for PM2.5 concentration forecasting

The prediction of PM2.5 concentration is essential and important for improving people's physical and mental health, as one of the main sources of air pollution, it limits the sustainable development of urban areas. In this paper, we propose a hybrid prediction model to predict PM$_{2.5}$ concentration for the next 72 hours by employing the spatio-temporal information of monitoring stations, called adaptive spatio-temporal prediction network (ASTP-NET), which consists of four components: feature extraction module, spatial attention module, Multi-Layer Bi-LSTM and future time features. The feature extraction module conducts adaptive feature extraction while preserving the inherent spatio-temporal features of different stations. The spatial attention network aims to obtain the spatial attention weights of the auxiliary and target stations based on the target station features. Multi-Layer Bi-LSTM with the transformation mechanism is executed to capture the entangled temporal features of all stations from their weighting fusion features. The purpose of adding future time feature module is to enhance the prediction performance through incorporating prior knowledge and expanding the time perception horizon. The evaluation of ASTP-NET is implemented based on the air quality dataset of Xi'an and the results show our model outperforms other latest state-of-the-art methods. The methodology can be applied to predict multivariate time series data in different domains.