Location Embedding and Relation Modeling Based Time Series Prediction

In recent years, air pollution has became one of the serious pollution problems with the rapid development of urbanization and industrialization. The analysis and assessment of air quality has drawn much attention. Nowadays, monitoring stations in a city can provide real-time air quality which is source data of prediction. Compared to the normal multivariate time series, there are relations between air monitoring stations. Therefore, it's necessary to take spatial influence into consideration when making forecast. Most commonly used methods have focused on the context of the temporal dimension. The mixed contexts have not been sufficiently explored and the relation between stations is not clear and simple enough. To solve these problems, we introduce Temporal-spatial Prediction Network (TSPN). It uses a graph structure to mine relations between stations, thus forming an implicit spatial context in addition to the original temporal context. TSPN combines GNN and LSTM to deal with mixed contexts. The experimental results on Harbin city show that TSPN achieves better performance, with minimum 2.1% improvement on R2 compared to other baselines.