Abstract:  Haze is a hot issue closely related to social development. In order to predict PM2.5 concentration and provide basis for its effective prevention and control,  the PM2.5 prediction model VMDLSTM is proposed based on the combination of the improved VMD (variational modal decomposition) and LSTM (long and shortterm memory) neural network. Firstly, the threshold method is used to determine the decomposition number of VMD method, then the historical data is decomposed into different sequences, further each sequence is predicted, and the final prediction result is obtained by summing the prediction results of each sequence. The VMDLSTM model is applied to the shortterm prediction of PM2.5 series in Beijing, and its result is compared with nine models such as ARIMA (autoregressive integrated moving average), RFR (random forest regression)，LSSVR（least squares support vector regression）, LSTM and so on,  by using seven evaluation indexes. The comparison results show that among the five error evaluation indexes, the VMDLSTM model performs best,  with only one error index ranking second. In the protocol index evaluation, the VMDLSTM model is closest to 1 and has the highest accuracy; The mean square error of VMDLSTM model is 41.10, the root mean square error is 6.42, the mean absolute error is 5.79, and the protocol index is 0.97. The mean square error range of RFR，VMDLSSVR，ARIMA，SVR，and LSTM models is from 60.72 to 1 058.07, the root mean square error range is from 7.79 to 32.53, the mean absolute error range is from 7.45 to 26.14, and the protocol index range is from 0.39 to 0.95. The VMDLSTM model proposed in this paper has the highest accuracy.

Key words:  , VMD, LSTM neural network, threshold value method, PM2.5, shortterm prediction