Neural network and multiple regression models for PM10 prediction in Athens: a comparative assessment

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Urban particulate air pollution significantly impacts human health.
• Accurate prediction of particulate matter concentrations is vital for public health and air quality management.
• Traditional statistical models are widely used for forecasting air pollution.

Purpose of the Study:

• To evaluate artificial neural networks (ANNs) as tools for daily average particulate matter with aerodynamic diameter less than 10 micrometers (PM10) concentration forecasting.
• To compare the predictive performance of ANNs against multiple linear regression models.
• To explore the utility of meteorological variables as inputs for PM10 forecasting models.

Main Methods:

• Development and evaluation of neural network models and multiple linear regression models.
• Utilized a two-year dataset from a central site in Athens, Greece.
• Employed meteorological variables as primary input data for the models.

Main Results:

• Neural network models demonstrated lower prediction errors (8.2-9.4% reduction in root mean square error) compared to regression models.
• ANNs exhibited improved episodic prediction ability, with 7-13% lower false alarm rates.
• ANNs proved to be a viable alternative to conventional statistical methods for PM10 forecasting.

Conclusions:

• Artificial neural networks offer a promising approach for particulate pollution forecasting.
• Properly trained ANNs can effectively address the demands of particulate matter prognostic needs.
• The study highlights the potential of ANNs in enhancing air quality management strategies.