Abstract
Due to the limited understanding of the physical/chemical processes and
large uncertainties in emissions, ozone prediction task becomes more
difficult with numerical models. Deep learning provides an alternative
way. However, most of the deep learning ozone prediction models only
consider temporality and have limited capacity. Exist spatiotemporal
deep learning models generally suffer from model complexity and
inadequate spatiality learning. Thus, we propose a novel spatiotemporal
model, namely the Spatiotemporal Attentive Gated Recurrent Unit
(STAGRU), which employs double attention mechanism and Gated Recurrent
Unit (GRU) to capture spatiotemporal information. We compare STAGRU with
Seq2Seq and their single attention version on nine monitoring stations
in Nanjing. The results show that STAGRU outperforms other competitors
in terms of RMSE, R2, and SMAPE. In addition, we make
an interpretability discussion for STAGRU. The discussion reveals that
wind direction plays an important role in ozone transmission and
temporality mainly involves short-term and periodical dependency.