A secondary-decomposition-ensemble learning paradigm for forecasting PM2.5 concentration

To design high-accuracy tools for hourly PM2.5 concentration forecasting, we propose a new method based on the secondary-decomposition-ensemble learning paradigm. Prior to forecasting, the original PM2.5 concentration series are processed using secondary-decomposition (SD): (1) wavelet packet decomposition (WPD) is used to decompose the time series into low-frequency components and high-frequency components; (2) the high-frequency components are further decomposed by the complementary ensemble empirical mode decomposition (CEEMD) algorithm. Then Phase space reconstruction (PSR) is utilized to determine the optimal input form of each intrinsic mode function (IMF). The least square support vector regression (LSSVR) model, optimized by the chaotic particle swarm optimization method combined with the gravitation search algorithm (CPSOGSA), is employed to model all reconstructed components independently. Finally, the predict results of these components are integrated into an aggregated output as the final prediction, utilizing another LSSVR optimized by CPSOGSA as an ensemble forecasting tool. Our empirical results show that this method outperforms the benchmark methods in both level and directional forecasting accuracy.