Hybrid Prophet-XGBoost Forecasting for Urban Air Quality with Real-Time Data Integration For Chennai

This research aims to forecast short-term urban air quality by integrating meteorological variables and pollutant concentration data into a hybrid machine learning framework. The hypothesis is that using both historical pollutant data and weather features enhances predictive accuracy across different pollutant types.

Historic pollutant data (2022–2024) was collected from the WAQI (World Air Quality Index) historic dashboard, cleaned, and imputed for missing or anomalous values. This dataset includes daily values for PM₂.₅, PM₁₀, NO₂, SO₂, and O₃ from six key monitoring stations in Chennai. Supplementary weather and CAMS air quality data were sourced using the Open-Meteo APIs:

https://api.open-meteo.com/v1/forecast for daily weather

https://air-quality-api.open-meteo.com/v1/air-quality for CAMS-based pollutant history

These datasets were combined and processed to ensure daily granularity, using linear interpolation for imputation and z-score-based outlier capping.

The air quality forecast was generated for a 7-day horizon using a hybrid approach: Prophet (for seasonality), XGBoost (for lag-based learning), and ETS (for fallback modeling). AQI was computed based on Indian CPCB sub-index breakpoints, identifying both the AQI level and dominant pollutant each day.

To evaluate performance, back-testing was done on actual vs. predicted pollutant values. Results show high model accuracy, particularly for gaseous pollutants. Based on MAPE and sMAPE:

O₃: 95.22% accuracy

SO₂: 93.04% accuracy

PM₁₀: 80.28% accuracy

PM₂.₅: 76.78% accuracy

NO₂: 65.74% accuracy

The model successfully captured pollutant dynamics and generated reliable AQI forecasts, suitable for environmental monitoring and early warning applications.