Optimization of artificial neural network models through genetic algorithms for surface ozone concentration...

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Computational Intelligence

Background:

• Predicting surface ozone (O(3)) concentrations is crucial for air quality management.
• Artificial neural networks (ANNs) offer potential for complex environmental modeling.
• Genetic algorithms (GAs) can optimize ANN architectures and parameters.

Purpose of the Study:

• To develop and evaluate three methodologies for defining ANNs using GAs to predict hourly average surface ozone concentrations.
• To explore threshold models where ozone behavior changes based on specific variables.
• To identify key predictors for ozone formation in urban environments.

Main Methods:

• Three distinct methodologies were proposed to define ANN models via GAs.
• Two methodologies employed threshold models with two and four regimes, respectively.
• Predictor variables included prior day's O(3), carbon monoxide (CO), nitrogen oxides (NOx), and meteorological data (temperature, solar radiation, humidity, wind speed).

Main Results:

• Genetic algorithms were successfully applied to determine ANN activation functions and neuron counts.
• In threshold models, temperature, CO, and nitrogen dioxide (NO(2)) were identified as key variables defining ozone regimes.
• These variables are significant due to their role in urban atmospheric ozone chemistry.

Conclusions:

• The developed threshold model with two regimes demonstrated the highest efficiency in predicting O(3) concentrations.
• The study successfully optimized ANNs for air pollution forecasting using GAs.
• The findings highlight the importance of specific chemical and meteorological factors in urban ozone dynamics.