Multi-Site Prediction Method for Atmospheric Pollutants Based on an Improved MLP-Mixer

The challenge in atmospheric pollutant forecasting lies in the fact that urban areas consist of multiple monitoring stations representing different regions, and the pollutant variations at these stations often differ. Employing traditional methods to design a universal predictive model results in lower accuracy of forecasted outcomes, and training specific models for different monitoring stations requires excessive time and labor costs. To address these issues, a multi-site atmospheric pollutant forecasting method based on an improved MLP-Mixer is proposed. Experiments were conducted on data sets from 12 main meteorological monitoring stations in Beijing and 145 stations in a certain industrial city. The designed model outperforms other models in predicting atmospheric pollutants for the next 24 hours, verifying the accuracy of the improved MLP-Mixer approach.