Comparing quantile regression methods for probabilistic forecasting of NO₂ pollution levels.

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Authorities use traffic restrictions to manage high nitrogen dioxide (NO2) concentrations.
• Accurate forecasting of NO2 levels is essential for timely intervention.
• Probabilistic forecasting offers advantages over point-forecasting for predicting pollutant distributions.

Purpose of the Study:

• To compare the performance of 10 state-of-the-art quantile regression models for NO2 concentration forecasting.
• To evaluate these models for predicting the full distribution of NO2 concentrations.
• To identify optimal probabilistic models for urban air quality management.

Main Methods:

• Utilized 10 quantile regression models to predict NO2 concentration distributions.
• Applied a semi-parametric approach by deriving distribution parameters from predicted quantiles.
• Evaluated models for forecasting horizons up to 60 hours in an urban setting.

Main Results:

• Quantile gradient boosted trees demonstrated superior performance in predicting both point values and full NO2 distributions.
• Quantile k-nearest neighbors with linear regression achieved comparable results with significantly reduced training time and complexity.
• The study provides a comprehensive comparison of probabilistic models for NO2 forecasting.

Conclusions:

• Probabilistic forecasting models are effective for predicting NO2 concentration exceedances and pollution peaks.
• Quantile gradient boosted trees are highly effective, but simpler models like quantile k-nearest neighbors offer a practical alternative.
• The findings support the use of advanced statistical methods for proactive air quality management.