Comparison of PM2.5 prediction performance of the three deep learning models: A case study of Seoul, Daejeon, and Busan

Abstract

In this study, the PM2.5, prediction performances of long short-term memory (LSTM), gated recurrent unit (GRU), and bidirectional LSTM (Bi-LSTM) were compared using data from Seoul, Daejeon, and Busan, which are representative cities in Korea. The data analysis period was from 9:00 on May 16, 2014, to 23:00 on December 31, 2021, based on data at 1 h intervals. The causal factors affecting the change in PM2.5 of three cities in Korea, and five major cities in China were determined. The analysis revealed that the three models showed similarly high performances in short-term prediction within 24 h (R2 ≥ 0.9). The Bi-LSTM model using both past and future time information showed high prediction accuracy for long-term prediction (R2 ≥ 0.6). Using the PM2.5 data of the five major Chinese cities, it was confirmed that the accuracy of the PM2.5 prediction model for Seoul, Daejeon, and Busan improved. The deep learning model showed a high accuracy even when the Fine Dust Act measures were implemented. This study can facilitate governments to prepare measures against air pollution with a high regional prediction performance by identifying the causal factors affecting PM2.5, specific to the city, and designing different models for each forecasting period.