A deep learning-based hybrid method for PM_2.5 prediction in central and western China

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Accurate prediction of PM2.5 is crucial for mitigating air pollution's adverse effects.
• Existing single models face inherent limitations in prediction accuracy.
• Hybrid models offer a promising approach to overcome individual model weaknesses.

Purpose of the Study:

• To develop a hybrid deep learning model for enhanced PM2.5 prediction.
• To fuse Transformer and LSTM architectures for synergistic performance.
• To optimize the hybrid model using Particle Swarm Optimization (PSO).

Main Methods:

• Integration of Transformer and LSTM deep learning architectures.
• Application of Particle Swarm Optimization (PSO) for parameter tuning.
• Utilizing LSTM's gating mechanism and Transformer's self-attention for improved feature extraction.

Main Results:

• The proposed hybrid model significantly outperforms traditional LSTM and PSO-LSTM models.
• Key evaluation metrics (R², MAE, MBE, RMSE, MAPE) show substantial improvements.
• The model demonstrates consistent and stable performance across diverse urban environments and timeframes.

Conclusions:

• The hybrid Transformer-LSTM model with PSO optimization provides a robust approach for PM2.5 forecasting.
• This fusion strategy enhances prediction accuracy and reliability.
• The study offers a valuable tool for air quality management and public health initiatives.