Abstract：Predicting PM2.5 24 hours in advance will avoid the harm from serious air pollution. To enhance the efficiency and generalization ability of deep learning prediction models for PM2.5 24 hours in advance prediction， a sub-sampled layer based on support vector regression (SVR) was attached to traditional recurrent Neural Network (RNN) to perform nonlinear features extraction and dimension reduction. And then a multi-kernel convolutional Neural Network (CNN) was employed to enhance feature expression. At last， a gated recurrent unit (GRU) network was employed to provide high stability of time sequence prediction using its ability in long time information memory. The air quality data and meteorological data of Wuhan and its surrounding 13 cities from January 1， 2015 to April 10， 2020 were employed to test the SVR-CNN-GRU. The experiment results showed that SVR-CNN-GRU exceed RNN， SVR and random forest methods in higher prediction accuracy and stronger generalization ability whose R2 is 0.97. The proposed method would provide high accuracy prediction for early warning 24 hours in advance.
2022, Vol. 56 
