To analyze the multi-scale characteristics of PM2.5 in Changsha-Zhuzhou-Xiangtan region, China, and explain the main dynamic mechanism of PM2.5 evolution, a novel ensemble empirical mode decomposition and multifractal detrended fluctuation analysis (EEMD-MFDFA) model is proposed, and then the dynamic evolution of PM2.5 hourly average concentrations in Xiangtan, Changsha and Zhuzhou from January 1, 2015 to December 31, 2019 is studied. Through EEMD, the high-frequency modes and trend terms of PM2.5 are obtained for the three cities. The results of the trend term show a decreasing trend of PM2.5 concentrations, and the high-frequency mode of PM2.5 reflects the nonlinear characteristics of PM2.5 concentration fluctuation. Furthermore, MFDFA method is used to analyze the high-frequency cumulative mode of PM2.5. The results indicate that the high-frequency mode of PM2.5 has strong multifractal characteristics. In addition, the main sources of multifractal characteristics are studied by using shuffling procedure and phase randomization. The results indicate that the long-term persistence of PM2.5 concentration fluctuation in different time scales is the maindynamic factor for the emergence of high concentration PM2.5. Finally, the influence of meteorological conditions on the multifractal strength of high-frequency modes of PM2.5 is discussed. The results show that the multifractal strength of PM2.5 in winter was stronger than that in other seasons. The analysis shows that although the air pollution in Changsha-Zhuzhou-Xiangtan Urban Agglomeration has been effectively controlled through the air pollution action plan, in winter, the long-term persistence mechanism of air pollution plays a more dominant role in controlling the PM2.5 evolutions, and the non-linear long-term persistent mechanism of PM2.5 at different time scales can lead to the risk of emergence of high concentration PM2.5 in winter, and even more serious air pollution events. The results have great significance for studying the dynamic characteristics of regional PM2.5 multi-scale evolution and forecasting haze.