Analysis of surface ozone using a recurrent neural network

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Ozone (O₃) and nitrogen dioxide (NO₂) are key air pollutants.
• Long-term air quality data is crucial for understanding pollution trends.
• Accurate modeling is needed for effective air quality management.

Purpose of the Study:

• To compare the performance of multiple linear regression and neural network models for hourly ozone concentration prediction.
• To evaluate the impact of meteorological and photochemical parameters on ozone modeling.
• To assess the effectiveness of a recurrent neural network architecture for air quality forecasting.

Main Methods:

• Collected 16 years of hourly O₃ and NO₂ data (1998-2013).
• Developed and compared multiple linear regression and neural network models.
• Utilized meteorological and photochemical parameters as model inputs.
• Employed statistical criteria (correlation coefficient, fractional bias, NMSE, factor of two) for performance evaluation.
• Investigated recurrent neural network architecture for improved prediction accuracy.

Main Results:

• Neural network models significantly outperformed the multiple linear regression model in all scenarios.
• The recurrent neural network architecture demonstrated superior performance compared to feed-forward networks.
• The recurrent neural network using only meteorological data surpassed the regression model with both meteorological and photochemical data.
• The model successfully forecasted O₃ concentrations 1 to 48 hours ahead.

Conclusions:

• Recurrent neural networks offer a powerful and efficient tool for hourly ozone prediction.
• The developed model is particularly valuable for air quality forecasting in data-scarce regions relying solely on meteorological data.
• The study highlights the potential for operational use of this neural network model in environmental monitoring and prediction.