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

Abstract

Urban air pollution poses a significant threat to public health and urban sustainability in megacities like Paris. We cast forecasting as a short-term, next-hour prediction task for PM2.5, NO, and CO, using hourly meteorology and recent pollutant history as inputs. We develop a data-driven framework based on hyperparameter-tuned ensembles (Random Forest, Gradient Boosting, and a Stacked Ensemble) and benchmark against a Long Short-Term Memory (LSTM) model, alongside persistence baselines. All evaluation metrics (RMSE/MAE) are reported in physical units (µg/m3) with R2 unitless. Results show that tree ensembles deliver the lowest errors for PM2.5 and CO, while LSTM is competitive for NO; stacking offers gains when base-model errors are complementary but does not universally dominate. The framework is designed for real-time deployment and integration into smart city pipelines, supporting proactive air quality management. By providing accurate, unit-consistent short-term forecasts, this study informs urban planning, risk mitigation, and public-health protection.