STF-Net: An improved depth network based on spatio-temporal data fusion for PM2.5 concentration prediction

Abstract

Accurate PM2.5 concentration prediction is essential and important for air pollution assessment. The air pollutant concentration (e.g. PM2.5 concentration) monitoring network consists of numerous stations deployed in different geographic locations, which generate multiple time series with dynamic spatio-temporal correlations. However, much existing work mainly focus on the observations themselves and ignore the dynamic spatio-temporal correlations among them, which is necessary for the accurate PM2.5 concentration prediction. On the one hand, geospatially distributed stations can provide data and additional hints for spatio-temporal dependency modeling of regional PM2.5 concentrations. On the other hand, the stations with high impact can be screened according to the correlation among values, which can reduce the computational complexity and improve the performance of the model. Inspired by the two insights mentioned above, in this paper, we design an improved depth network based on spatio-temporal data fusion model (short for, STF-Net) to predict future PM2.5 concentrations. Specifically, STF-Net is proposed based on a multilevel recurrent neural network (RNN) and temporal convolutional network that considers multiple stations’ values and other critical external data. And it is made up of two key components: (1) a data fusion component is mainly modeling the dynamic spatial correlation between the station-level data; (2) a prediction component aims to learn the temporal pattern of time series. The experiments with two real-world datasets demonstrate our method outperforms six baseline methods.