A novel hybrid ensemble model for hourly PM2.5 concentration forecasting

PM2.5 is the main constituent of haze whose equivalent diameter is less than or equal to 2.5 mu m. Highly concentrated PM2.5 suspended in the air for a long time may cause serious air pollution. To capture the complicated multi-scale factors related to PM2.5 concentration, a novel hybrid ensemble model is proposed for PM2.5 concentration forecasting. Firstly, the complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is utilized to decompose the nonlinear and non-stationary features in the PM2.5 data. Secondly, the fuzzy C-means (FCM) clustering method divides the decomposition components with similar characteristics into groups. Thirdly, the long short-term memory (LSTM) network is explored for the sub-models forecasting. Finally, the empirical study focuses on the PM2.5 concentration forecasting of Xinyang, which is one of the top 10 livable cities in China. The forecast results of the three cases in this paper show that the CEEMDAN-FCM-LSTM model at least reduces the root-mean-square error (RMSE) by 46.35%, mean absolute percentage error by 58.73%, and mean absolute error by 61.78%, and improves the coefficient of determination R-2 by 7.28%, compared to the single LSTM model.