Fourier-based multiharmonic temporal encoding for enhanced PM2.5 estimation in machine learning models

Accurately estimating PM2.5 concentrations has always been difficult due to its complicated periodicity. On the foundation of the two-stage model, we introduce an enhanced Fourier-based temporal encoding method that is enriched with more than one harmonic for efficient capturing of seasonal and sub-seasonal patterns. We were quite optimistic about XGBoost and 10-fold cross-validation result: our findings show reduction from RMSE = 9.545μg/m³ up to 9.3284μg/m³ , decrease in MAE scores from 4.973μg/m³ to 4.8889 μg/m³ , R² increases from 0.907 up to 0.9118. Our results show that the proposed Fourier time coding method can better capture the detailed seasonal and sub-seasonal variations of PM2.5 concentration, which significantly improve the prediction accuracy and provide valuable insights for future air pollution modeling.