Citywide PM2.5 Concentration Prediction Using Deep Learning Model

With the accelerated pace of urbanization and modernization, the problem of air pollution is becoming increasingly serious worldwide. While monitoring and forecasting air pollutant concentration are important for both the government and the public, this study focuses on the problem of citywide short-term and fine-grained PM2.5 (the particulate matter with an aerodynamic equivalent diameter of less than or equal to 2.5 microns in ambient air) concentration forecasting. This work designs a model based on a deep neural network that can extract the effects of temporal influences, spatial transport, and meteorological conditions on the variation of air pollutant concentrations. We construct a dataset based on real historical PM2.5 concentrations in Beijing. For the 1-hour prediction, the RMAE value and the MAE value of the results on the test set are 3.79 and 2.68. In addition, we evaluate the model from the perspective of level predicting accuracy in response to measuring its guiding value in practical forecasting operations. Experiments show that our model performs better than baseline models in both 3-hour prediction and heavy pollution event forecast.