Convolution Recurrent Neural Networks Based Dynamic Transboundary Air Pollution Predictiona

Transboundary air pollution is one of the main sources of air pollution in island cities. However, the transboundary pollution confounded by local emission, meteorological conditions, and it is difficult to predict. Currently, most of urban air pollution prediction methods do not predict with transboundary air pollution. Therefore, we introduce a dynamic transboundary air pollution prediction approach based on convolutional recurrent neural networks(D-CRNN) which: (i) Divide the prediction inputs into prediction locations and transboundary air pollution sources; (ii) Use two different convolutional recurrent neural networks to solve the spatial-temporal feature of each inputs. (iii) Through a transboundary prediction network to integrate the spatial-temporal feature of prediction locations with the spatial-temporal feature of transboundary air pollution sources in a dynamic asynchronous method. Then use those mixed features to predict the air pollution. We evaluate our DCRNN model with the local atmospheric monitoring data in Kyushu, Japan and the transboundary air pollution data from 33 coastal cities in eastern Asia from January 2015 to July 2017. The results show that our D-CRNN model have achieved 86.2%, 78.6% accuracy of total prediction and transboundary air pollution in next 6 hours.