Air pollution remains a significant issue in China, especially in regions with large environmental disparities. The aim of this work is to select a model to predict multi step PM \(_{\varvec{2.5}}\) concentration at various stations with differing natural and industrial environments. To this end, we integrated air pollution data and meteorological data from the ministry of ecology and environment (MEE) and the national climatic data center (NCDC), and selected eight stations from the Fenwei plain. We find that convolutional neural network with long short-term memory network under variational mode decomposition (CNN-LSTM-VMD) outperforms other decomposition algorithms and models in terms of regression scores. We employed the standards of the environmental protection agency (EPA) and the ministry of ecology and environment (MEE) to evaluate the methods, and CNN-LSTM-VMD achieved an accuracy of 81.18%-97.93% across various stations. Despite identified limitations, the method is capable of accurately and efficiently predicting day-ahead air pollution values at various stations with differing situation.
