Since 2013, the concentration of PM2.5 in Beijing has been has dropped significantly, there is still there is still far from the World Health Organization’s revised 2021 target of 5µg/m 3 of the average annual concentration. We analyzed the monthly, quarterly and annual trends of PM2.5 concentrations in the Beijing, China, from 2014 to 2021 using independent T-test. In addition, we proposed a new combined PM2.5 concentration prediction model, which combines mode decomposition and reconstruction algorithm, convolutional neural network and long short-term memory neural networks. The results show that: (1) From 2016 to 2018, PM2.5 concentration in Beijing decreased at a fast pace, and then the decline entered a slow period. (2) The combination of the modal decomposition algorithm and neural network model can effectively process the PM2.5 concentration data with the characteristics of nonlinearity, instability and high complexity, and is more stable and generalizable. (3) The combination of convolutional neu-ral network and long short-term memory neural network model has a higher prediction accuracy than traditional neural network models. The research results are helpful to know the possible air pollution process in advance and take corresponding measures to reduce the harm caused by air pollution in the new stage of haze pollution.