DF-OSELM: a dynamic feedback feature learning model for air quality online prediction

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Accurate air quality forecasting is vital for public health and pollution risk mitigation.
• Existing models struggle with adaptability, computational speed, and interpretability.

Purpose of the Study:

• To develop an advanced online sequential extreme learning machine for real-time air quality prediction.
• To enhance model adaptability, efficiency, and interpretability.

Main Methods:

• Proposed a Dynamic Feedback Feature Learning Online Sequential Extreme Learning Machine (DF-OSELM).
• Integrated dual Extreme Learning Machine Autoencoders (ELM-AEs), a normalization layer, and a recurrent feedback mechanism.
• Trained the model online using 10,000 hourly air quality samples (PM₂, PM₁₀, SO₂, NO₂).

Main Results:

• DF-OSELM achieved superior predictive performance (NRMSE < 0.1, R² > 0.99), outperforming baseline models.
• Ablation studies validated the importance of normalization and dual autoencoder mechanisms.
• Uncertainty quantification provided reliable confidence intervals, and SHAP analysis identified key predictors.

Conclusions:

• DF-OSELM offers a balanced approach to accuracy, efficiency (update time < 3ms), and interpretability for real-time air quality monitoring.
• The model is suitable for large-scale environmental platforms and risk assessment.