Atmospheric PM2.5 concentration is a kind of data with strong time series characteristics, so currently the prediction of PM2.5 concentration is mostly based on RNN, LSTM and other sequence models. However, RNN, LSTM and the other similar models use the same weight to calculate the input data at different times, which is not in line with the brain-like design, resulting in the low accuracy of PM2.5 concentration prediction. In view of the above problems, a PM2.5 prediction method (AT-RNN and AT-LSTM) based on Adam attention mechanism is proposed. This method firstly looks for the optimal parameters of RNN or LSTM through Adam algorithm, and introduces attention mechanism in Encoder stage to assign attention weight to input with time series characteristics, and then carries out Decoder analysis and prediction. Through the experiment, the prediction effects of BP, RNN, LSTM and AT-RNN and AT-LSTM on PM2.5 concentration in Hefei city were compared. The results show that the prediction method based on Adam attention model is more accurate than other methods, which proves the effectiveness of this method in pollutant prediction.