Longer Time Span Air Pollution Prediction: The Attention and Autoencoder Hybrid Learning Model

Air pollution has become a critical issue in human’s life. Predicting the changing trends of air pollutants would be of great help for public health and natural environments. Current methods focus on the prediction accuracy and retain the forecasting time span within 12 hours. Shorter time span decreases the practicability of these perditions, even with higher accuracy. This study proposes an attention and autoencoder (A&A) hybrid learning approach to obtain a longer period of air pollution changing trends while holding the same high accuracy. Since pollutant concentration forecast highly relates to time changing, quite different from normal prediction problems like autotranslation, we integrate “time decay factor” into the traditional attention mechanism. The time decay factor can alleviate the impact of the value observed from a longer time before while increasing the impact of the value from a closer time point. We also utilize the hidden states in the decoder to build connection between history values and current ones. Thus, the proposed model can extract the changing trend of a longer history time span while coping with abrupt changes within a shorter time span. A set of experiments demonstrate that the A&A learning approach can obtain the changing trend of air pollutants, like PM2.5, during a longer time span of 12, 24, or even 48 hours. The approach is also tested under different pollutant concentrations and different periods and the results validate its robustness and generality.