Predicting carbonaceous aerosols and identifying their source contribution with advanced approaches

Area of Science:

• Atmospheric Chemistry
• Environmental Science
• Data Science

Background:

• Organic carbon (OC) and elemental carbon (EC) are key components of atmospheric aerosols impacting health and climate.
• Accurate prediction of carbonaceous aerosols and source apportionment are crucial for epidemiological studies and emission control.

Purpose of the Study:

• To investigate the characteristics and sources of OC and EC in Taipei's PM2.5 from 2005-2010.
• To apply machine learning for predicting OC and EC concentrations and identify their sources.
• To be the first study using machine learning and hyperparameter optimization for predicting specific aerosol contaminants.

Main Methods:

• Collected hourly average concentrations of OC and EC in Taipei's PM2.5 from 2005 to 2010.
• Employed a generalized additive model and a grey wolf optimized multilayer perceptron model for prediction.
• Utilized clustering techniques for source apportionment analysis.

Main Results:

• Average concentrations of OC and EC were 5.2 μg/m³ and 1.6 μg/m³, respectively, with significant seasonal variation in OC.
• Machine learning models explained up to 80% of the total variation in OC and EC concentrations.
• Traffic emissions were identified as the primary source of OC, contributing 65-90% seasonally, with local emissions dominating overall.

Conclusions:

• Machine learning models effectively predict OC and EC concentrations in Taipei's PM2.5.
• Traffic emissions are a major contributor to OC, and local sources dominate OC and EC throughout the year.
• Long-range transport significantly impacts OC levels, particularly during spring.