Extraction of multi-scale features enhances the deep learning-based daily PM_2.5 forecasting in cities

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Accurate daily PM2.5 concentration forecasting is vital for effective air quality management.
• Existing models may struggle with the complex, non-linear dynamics of atmospheric pollutants.

Purpose of the Study:

• To propose and evaluate a novel hybrid model for enhanced PM2.5 forecasting.
• To assess the model's performance across diverse geographical and economic conditions.

Main Methods:

• Utilized a two-stage decomposition technique: Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Variational Mode Decomposition (VMD).
• Employed a Long Short-Term Memory (LSTM) deep learning network for time-series prediction.
• Validated the hybrid CEEMDAN-VMD-LSTM model in five cities with varying characteristics.

Main Results:

• PM2.5 levels were significantly higher in inland cities (66.98 ± 0.76 μg m⁻³) compared to coastal cities (40.46 ± 0.40 μg m⁻³).
• The secondary decomposition approach demonstrably improved prediction accuracy and stability.
• The CEEMDAN-VMD-LSTM model achieved superior prediction performance (R² = 0.9803 ± 0.01) versus benchmark models (R² = 0.7537 ± 0.03).

Conclusions:

• The hybrid CEEMDAN-VMD-LSTM model offers a robust solution for PM2.5 forecasting.
• This approach effectively captures complex correlations within time-series data for improved air quality assessment.
• The findings support the model's utility in identifying patterns for better atmospheric environment governance.