Abstract: PM2.5 data is a kind of time series data, which has strong time-series and non-linear characteristics. Traditional time series modeling algorithms include long short-term memory artificial neural network (LSTM), recurrent neural network (RNN), encoder-decoder neural network (Seq2Seq) and other methods. In this paper, we propose a PM2.5 prediction algorithm based on Seq2Seq network fused with attention mechanism (Seq2Seq+Attention), where the cell unit of Seq2Seq is LSTM, which can fully extract the effective feature information of the input and enhance the learning ability and prediction effect of the network. Prediction tests were conducted using PM2.5 data from ten cities in East China from January 2019 to August 2021, and the tests compared the accuracy of PM2.5 numerical prediction of LSTM, Seq2Seq and Seq2Seq+Attention methods within 24 hours. The results show that the Seq2Seq+Attention method outperforms the other methods in terms of prediction effectiveness, and its 24-hour prediction score is also higher than the other methods.

Key words: PM2.5 prediction, Seq2seq, attention mechanism, deep learning, time series