Multi-Site Air Quality Index Forecasting Based on Spatiotemporal Distribution and PatchTST-Enhanced: Evidence From Hebei Province in China

Abstract

Efficient and accurate air quality forecasting contributes significantly to environmental governance, health promotion, and the development of smart cities. However, few existing models can achieve multi-scale feature fusion and long sequence time series modeling in the prediction process. This study proposes a PatchTST-Enhanced model for multi-site air quality forecasting based on spatiotemporal distribution. It uses daily air quality data from 11 prefecture-level cities in Hebei Province from December 2, 2013, to October 12, 2023, to train the model. The new model demonstrates robust performance in Hebei’s air quality forecasting (PreLen =96: MSE =0.5408, MAE =0.5408, RSE =0.5408; PreLen =192: MSE =0.2795, MAE =0.2795, RSE =0.2795; PreLen =336: MSE =0.6779, MAE =0.6779, RSE =0.6779; PreLen =720: MSE =0.6779, MAE =0.6779, RSE =0.6779), and the prediction accuracy has been significantly improved compared to both the pre-optimization model and other existing models. The PatchTST-Enhanced outperforms the PatchTST and improves it through four optimization modules: CGAttention, SiLU activation, AdamW optimizer, and SmoothL1 Loss function. By incorporating spatiotemporal features, the PatchTST-Enhanced can address the challenge of combining spatial and large temporal scales in air quality forecasting. The results provide critical information to protect health and improve the environment for the public.

Keywords

• Air quality forecast
• atmospheric pollutants
• spatiotemporal distribution
• PatchTST-enhanced
• deep learning