Regime-aware hybrid ensemble learning for adaptive [Formula: see text] forecasting in urban environments

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Accurate forecasting of fine particulate matter (PM2.5) is crucial for public health and environmental protection.
• Existing models struggle with dynamic meteorological conditions, limiting their effectiveness.
• Urban air quality management requires advanced predictive tools.

Purpose of the Study:

• To develop a novel hybrid ensemble framework for adaptive PM2.5 forecasting.
• To enable context-specific model dominance using meteorological dispersion parameters.
• To improve the accuracy and adaptability of air quality prediction models.

Main Methods:

• Developed a hybrid ensemble framework adaptively switching between CatBoost and TabNet.
• Utilized meteorological dispersion parameters (ventilation coefficient, relative humidity) as switching triggers.
• Evaluated the model using the OpenAQ dataset for hourly PM2.5 concentrations in Delhi, India.

Main Results:

• The ensemble model achieved superior performance, with RMSE of 15.54, MAE of 11.02, MAPE of 12.2%, and R² of 0.86.
• Outperformed baseline models (LSTM, XGBoost, LightGBM) with statistically significant improvements.
• Demonstrated efficient inference with only a 14% increase over CatBoost alone.

Conclusions:

• The adaptive ensemble framework offers a scalable and accurate solution for urban air quality management.
• The methodology is extendable to other pollutants, regions, and ensemble combinations.
• Provides a practical tool for data-driven environmental policy and public health protection.