A balanced social LSTM for PM_2.5 concentration prediction based on local spatiotemporal correlation

Reliable prediction for the concentration of PM_2.5 has become a hot topic in pollution prevention. However, the prediction for PM_2.5 concentration remains a challenge, one of the reasons is that current prediction methods do not consider the relevance of PM_2.5 concentration among surrounding areas. In this paper, we propose the assumption that the PM_2.5 concentration has spatial interaction, which includes two parts: 1) The PM_2.5 concentrations observed by adjacent stations usually present relevant trends; 2) Stations with higher PM_2.5 concentration tend to show higher influences on neighboring areas. Based on the spatial interaction assumption, we propose a balanced social long short-term memory (BS-LSTM) neural network for the prediction of PM_2.5 concentration. BS-LSTM is composed of two kernel components: a social-LSTM based prediction model and a new balanced mean squared error (B-MSE) based loss function. On the one hand, to capture the spatiotemporal correlation of the PM_2.5 concentration among adjacent stations, we develop a social-LSTM based model which has advantages in describing the trend information of neighboring locations. On the other hand, considering the unbalanced influence caused by various local pollution levels, we design a new B-MSE loss function to assign different attention to the observation stations. In the experiments, we evaluate the proposed method on two real-world PM_2.5 datasets. The results indicate that BS-LSTM is promising, especially in the case of heavy pollution.