Long time series ozone prediction in China: A novel dynamic spatiotemporal deep learning approach

Abstract

Ozone pollution is a global environmental problem becoming increasingly prominent in China. It is of great significance to achieve long-term and high-precision ground-level ozone prediction on large scales to improve the efficiency of environmental governance. In this paper, we developed a dynamic graph convolutional and sequence to sequence embedded with the attention mechanism model (DG-ASeqseq) for predicting daily maximum 8-h average ozone (MDA8 O3) concentrations over China the next seven days. In the proposed approach, changeable spatial correlations are modelled by graph convolutional operations on dynamic graphs constructed based on multiple information of historical change, and temporal correlations in long time series are modelled through the sequence to sequence networks embedded with the attention mechanism. Results show the reliability and effectiveness of the proposed model, and it is superior to other benchmark models in simulating long-term spatiotemporal variations of O3 concentrations in large scale areas. Moreover, the proposed model has good prediction capability in severe O3 pollution events. Advancement in this methodology could provide guidance for the government's coordinated control of regional pollution to help improve air quality and jointly safeguard global climate security.