Analysis and prediction of atmospheric ozone concentrations using machine learning

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Atmospheric ozone chemistry is complex, involving numerous substances and reactions.
• Accurate prediction of ozone concentrations is crucial for air quality management.
• Machine learning (ML) offers potential for analyzing complex environmental data.

Purpose of the Study:

• To evaluate the capability of ML models for predicting daily average ozone concentrations.
• To identify relevant atmospheric parameters for ozone prediction.
• To establish a generalizable approach for ML-based air quality forecasting.

Main Methods:

• Analysis of air quality data from Switzerland's National Air Pollution Monitoring Network (NABEL).
• Application of feature selection techniques (e.g., best subset selection) to identify predictive parameters.
• Development and comparison of various ML models, including artificial neural networks, linear, and non-linear models.
• Model training and validation using data from 2016-2023 at NABEL stations in Lugano, Dübendorf, and Zürich.

Main Results:

• A non-linear ML model with 12 components, utilizing parameters like NO2, NOx, SO2, VOCs, temperature, and radiation, yielded the lowest mean absolute errors (MAE).
• Predicted ozone concentrations in Lugano achieved an MAE as low as 9 μgm⁻³.
• MAEs for stations in Zürich and Dübendorf were approximately 11 μgm⁻³ and 13 μgm⁻³, respectively.
• The accuracy of the best models approached 1 μgm⁻³, which is lower than the standard deviation of observations.

Conclusions:

• ML is a valuable tool for analyzing complex atmospheric data and predicting ozone concentrations.
• Simpler non-linear ML models can be as effective, or even more effective, than complex artificial neural networks for this task.
• The developed approach provides a framework for applying ML to air quality monitoring and prediction challenges.