Forecasting urban air quality in Paris using ensemble machine learning: A scalable framework for environmental...

Area of Science:

• Environmental Science
• Data Science
• Public Health

Background:

• Urban air pollution is a major threat to public health and sustainability in large cities.
• Accurate short-term air quality forecasting is crucial for effective management.

Purpose of the Study:

• To develop and evaluate a data-driven framework for next-hour air pollution forecasting (PM2.5, NO, CO).
• To compare the performance of ensemble machine learning models against a Long Short-Term Memory (LSTM) model and persistence baselines.

Main Methods:

• Utilized hourly meteorological data and recent pollutant history as inputs.
• Developed and hyperparameter-tuned ensemble models: Random Forest, Gradient Boosting, and Stacked Ensemble.
• Benchmarked ensemble models against an LSTM model and persistence methods using RMSE, MAE, and R² metrics.

Main Results:

• Tree-based ensembles (Random Forest, Gradient Boosting) achieved the lowest prediction errors for PM2.5 and CO.
• The Long Short-Term Memory (LSTM) model demonstrated competitive performance for NO forecasting.
• Stacked ensembles provided additional gains when base models had complementary errors, but did not universally outperform individual models.

Conclusions:

• The developed framework offers accurate, unit-consistent short-term air quality forecasts suitable for real-time deployment.
• This approach supports proactive air quality management, urban planning, and public health protection in megacities.
• The findings highlight the effectiveness of ensemble methods for specific pollutant forecasting tasks.