PM2.5 Forecasting Model Using a Combination of Deep Learning and Statistical Feature Selection

Abstract

This paper proposed a PM 2.5 forecasting model using Long Short-Term Model (LSTM) sequence to sequence combined with the statistical method. Correlation Analysis, XGBoost, and Chemical processed are used as the methods to select the essential features. The air pollution data is extracted from Taiwan Environmental Protection Agency (EPA) for the Taichung City dataset in 2014-2018. The study points out that chemical processed model of particulate matter 10 micrometers or less in diameter (PM10), Sulfur Dioxide (SO2), and Nitrogen Dioxide (NO2) have the highest accuracy or lowest Root Mean Square Error (RMSE) and more short training and testing time among the other models. The chemical processed model of PM10, SO2, and NO2 (model B) has the highest accuracy (lowest RMSE), approximately 1 point lower RMSE values, and the shortest training and testing period among the other models. Furthermore, RMSE calculations based on the stations reveal that training with the entire station dataset has a 3 point higher RMSE value than training with each station dataset.