Forecasting air pollution concentration using long Short-Term Memory Networks

A deep learning prediction model integrating multi-source spatiotemporal features and seasonal trend decomposition was proposed． The model utilizes the graph convolutional network (GCN) and the time-domain convolutional network with attention mechanism to extract the spatial-dependent and temporal features of particulate matter 2.5 (PM2.5) concentration; after the Hadamard product fusion of the two, they are concatenated with the key meteorological and pollutant factors selected by random forest to form multi-source spatiotemporal features． Then, the seasonal trend decomposition algorithm is used to decompose the PM2.5 concentration sequence of the target site into trend, season, and residual sub-sequences; the multi-source spatiotemporal features are concatenated with each sub-sequence respectively, and input into the long short-term memory (LSTM) network - sequence-to-sequence - attention mechanism model for prediction, obtaining the final prediction result of PM2.5 concentration． The multi-source spatiotemporal feature seasonal trend decomposition model is named M5． The variant models obtained after removing the time, space, spatiotemporal feature fusion and seasonal trend decomposition feature extraction modules are named M1, M2, M3, and M4． Compared with M1 to M4, the root mean square error (RMSE) values of M5 in the 1-hour prediction are reduced by 13.1%, 16.6%, 14.8%, and 9.7% respectively; in the 3-hour prediction, they are reduced by 18.7%, 12.0%, 9.6%, and 13.1% respectively; in the 6-hour prediction, they are reduced by 15.8%, 8.7%, 10.7%, and 20.5% respectively; compared with M1 to M4, in each sliding window end observation time, the overall RMSE of the error square summary calculation of M5’s prediction from the 1st to the 6th hour is reduced by 17.2%, 10.8%, 10.3%, and 16.5% respectively．