Air Quality Prediction with 1-Dimensional Convolution and Attention on Multi-modal Features

Abstract

Air pollution, especially from particulate matters, has become a serious problem in many countries. To cope with these abrupt pollutions, there have been several studies to predict the temporal concentration of air pollution using deep neural networks. However, these studies have difficulties in predicting accurately since the air quality is complexly correlated with various types of multi-modal features over a long time. In this paper, we propose a new architecture to predict air qualities of particulate matters incorporating deeply stacked 1 dimensional CNN with residual connection [I] and attention mechanism [2]. Specifically, 1-dimensional CNN extracts highlevel features with large receptive fields and attention mechanism captures complex correlation among various features. Through extensive experiments with Seoul air pollution data and public benchmarks, we verify our architecture achieves state-of-the-art result in nib.; and P.Mio prediction.