PM 2.5 Concentration Prediction in the Cities of China Using Multi-Scale Feature Learning Networks and Transformer Framework

Particulate matter (PM) concentration, especially PM 2.5 , is a major culprit of environmental pollution from unreasonable energy system emissions that significantly affects visibility, climate, and public health. The prediction of PM 2.5 concentration holds significant importance in the early warning and management of severe air pollution, since it enables the provision of guidance for scientific decision-making through the estimation of impending PM 2.5 concentration. However, due to diversified human activities, seasonal factors and industrial emissions, the air quality data not only show local anomalous mutability, but also global dynamic change characteristics. This hinders existing PM 2.5 prediction models from fully capturing the aforementioned characteristics, thereby deteriorating the model performance. To address these issues, this study proposes a framework integrating multi-scale temporal convolutional networks (TCNs) and a transformer network (called MSTTNet) for PM 2.5 concentration prediction. Specifically, MSTTNet uses multi-scale TCNs to capture the local correlations of meteorological and pollutant data in a fine-grained manner, while using transformers to capture the global temporal relationships. The proposed MSTTNet’s performance has been validated on various air quality benchmark datasets in the cities of China, including Beijing, Shanghai, Chengdu, and Guangzhou, by comparing to its eight compared models. Comprehensive experiments confirm that the MSTTNet model can improve the prediction performance of 2.42%, 2.17%, 2.87%, and 0.34%, respectively, with respect to four evaluation indicators (i.e., Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, and R-square), relative to the optimal baseline model. These results confirm MSTTNet’s effectiveness in improving the accuracy of PM 2.5 concentration prediction.