Long Short Term Memory Model for Analysis and Forecast of PM2.5

Leiming Yan, Min Zhou, Yaowen Wu, Luqi Yan

Abstract: Atmospheric PM2.5 is a pollutant that has a major impact on the atmospheric environment and human health. Based on LSTM, we construct two prediction models, Stack LSTM and Encoder-Decoder, and evaluate the prediction performance of the model through four years of meteorological data training and testing models in Nanjing, Beijing, and Sanya. In the experiment, using the meteorological factors, contaminant factors, seasonal factors, and the normalized results of PM2.5 as input, the daily average PM2.5 concentration is predicted from 1 to 3 days. Experimental results show that the LSTM model has better performance than Random Forest and Encoder-Decoder. Using Nanjing as an example, comparing the forecast results of Nanjing PM2.5 with the data released by the environmental authorities, it is found that the value of the PM2.5 concentration predicted by the LSTM model is very close to the value of PM2.5 monitored by Nanjing’s environmental authorities. In prediction of PM2.5 for three consecutive days, the Root Mean Square Error (RMSE) of the LSTM model is only 18.96. Under the LSTM model, the prediction result of the three cities are better than other prediction models, which shows that the LSTM model has a good adaptability in predicting the PM2.5 concentration.