Using multiple linear regression method (MLR) and BP neural network method (BPNN), PM2.5 concentration prediction models were built by taking 1 h, 3 h, 6 h, 12 h, 24 h, 48 h as prediction time from 2015 to 2018 in Guigang. The accuracy of the models were tested and compared. The results showed that both MLR and BPNN could be used in PM2.5 concentration prediction. The prediction accuracy declined as the prediction time increased. The mean absolute error(MAE) of the prediction models by MLR and BPNN were 4.01 μg/m3~15.48 μg/m3 and 3.89 μg/m3~15.63 μg/m3, respectively. Using wavelet data analysis method for optimizing the contaminant data and making the model again, both W MLR and W BPNN were optimized. The optimization was significant when the prediction time was from 3 h to 24 h. By W MLR and W BPNN models, the MAE reduced 1.6%~13.5% and 0.8%~9.8%, respectively, and W BPNN model was superior to W MLR model in prediction. |