The importance of data splitting in combined NO_x concentration modelling

Area of Science:

• Environmental Science
• Public Health
• Data Science

Background:

• Urban air pollution poses significant health risks.
• Effective pollutant modeling can provide early warnings and inform mitigation strategies.

Purpose of the Study:

• To compare two modeling approaches (C&RT and CA) for predicting air pollutant concentrations.
• To evaluate the performance of three machine learning methods (ANN, RF, SVR) within these approaches.

Main Methods:

• Developed C&RT models based on previous hour's NOₓ concentration.
• Developed CA models incorporating meteorological factors, traffic, and past NOₓ/NO₂ concentrations.
• Applied Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Regression (SVR) to both modeling approaches.

Main Results:

• ANN and RF models achieved the best prediction accuracy, with Mean Absolute Percentage Error (MAPE) between 18.3-18.5%.
• SVR models showed poorer performance (MAPE of 23.4% for C&RT, 29.3% for CA).
• No significant difference in performance was found between the C&RT and CA modeling approaches.

Conclusions:

• Machine learning models, particularly ANN and RF, are effective for air pollution forecasting.
• The choice between C&RT and CA approaches depends on specific application needs.
• Accurate air quality prediction is vital for urban health and environmental planning.