Constructing and Verifying a Generalized Model for PM10 Prediction Based on Atmospheric Pollutants and Meteorological Factors

Accurate prediction of PM10 concentrations is essential for safeguarding public health and forms the foundation for effective collaborative control of regional air pollution. The generalizability of PM10 prediction models is often limited by their localized training and variable selection. Consequently, their predictive performance deteriorates in new geographical contexts and fails to achieve the generalizability required for nationwide deployment. A novel framework—sifting influencing factors, comparing basic models, optimizing parameters intelligently, and verifying cross-regional universality—is proposed because of the limitations. Based on correlation analysis and literature evidence, five atmospheric pollutant variables and four meteorological factors were identified as key predictors. Seven machine learning algorithms were compared using multiple metrics, and Support Vector Regression (SVR) performed best across 82 cities (R2 = 0.801, RMSE = 21.396 μg/m3, MAE = 12.365 μg/m3). The Sparrow Search Algorithm (SSA) then optimized SVR’s penalty and kernel parameters, raising R2 to 0.836 (RMSE = 19.104 μg/m3, MAE = 11.544 μg/m3). Independent validation across seven heterogeneous regions yielded R2 values from 0.778 to 0.962 (mean 0.854), with RMSE of 16.712 μg/m3 and MAE of 8.910 μg/m3. These results demonstrate strong generalization of the SSA-SVR model across multiple heterogeneous regions, supporting its potential nationwide applicability to PM10 prediction.