Impacts of Temporal Resolution and Spatial Information on Neural-Network-Based PM2.5 Prediction Model

Taking Beijing as an example and using the data of air quality monitoring stations from 2015 to 2018, the impacts of temporal resolution and spatial information on the PM2.5 concentration prediction were analyzed by a BP neural network, an LSTM network, and a CNN-LSTM hybrid model. The results show that neural network models are generally better than the multi-linear regression model. Increasing the temporal resolution of the input data can significantly improve the accuracy of the predicted daily average PM2.5 concentration. When the temporal resolution of the input data increases from one day to 6 hours, the mean absolute error of the LSTM model reduces from 27.39 μg/m3 to 20.59 μg/m3. This improvement is more obvious when the weather is significantly getting better or getting worse. The distribution of PM2.5 concentration in North China has distinct spatial and temporal characteristics. The first spatial mode is a uniformly increasing or decreasing mode, and the second one is a north/south dipole mode. The analysis shows that the concentration of PM2.5 in Beijing is related to the PM2.5 in Inner Mongolia, Hebei, and Tianjin of the previous day. The CNN-LSTM hybrid model, trained with the spatial-temporal information of PM2.5 in North China, can further improve the predictability of PM2.5 in Beijing. It further reduces the mean absolute error to 17.36 μg/m3. © 2020 Peking University.