A novel model for hourly PM_2.5 concentration prediction based on CART and EELM

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Hourly PM2.5 concentrations exhibit complex, multi-pattern changes.
• Accurate prediction requires localized models but faces challenges with global-local duality.

Purpose of the Study:

• To develop a novel hierarchical prediction model for hourly PM2.5 concentrations.
• To address the global-local duality inherent in localized prediction methods.

Main Methods:

• A hybrid approach combining Classification and Regression Trees (CART) for hierarchical data splitting.
• Ensemble Extreme Learning Machines (EELM) are trained at each node and leaf of the CART-generated tree.
• Model selection at each leaf considers both global and path-specific local EELM models.

Main Results:

• The developed CART-EELM model effectively handles multiple change patterns in PM2.5 data.
• It demonstrates superior performance compared to global models (Random Forest, v-SVR, EELM) and traditional local models (season, k-means).
• The method successfully resolves the global-local duality issue.

Conclusions:

• The novel hierarchical CART-EELM model offers improved accuracy for hourly PM2.5 prediction.
• This approach enhances the capability to model complex air pollutant concentration dynamics.
• The study provides a robust framework for environmental data prediction challenges.