Predicting the Future PM2.5
CatBoost and Transformer models are used to predict indoor PM2.5 levels 1-6 hours ahead. The models integrate 13 features from multiple sources: indoor measurements (PM2.5, humidity, pressure) from SMART Lab sensors, outdoor air quality (PM1, PM2.5, PM10, temperature, humidity, pressure) from the NYU CITIES network[1], atmospheric conditions (dew point) from the NYU ACCESS meteostation[2], and wind data (U/V components, gusts) from ERA5 reanalysis[3]. The ERA5 reanalysis data is used with a 5-day lag due to the constraint of data accessibility. Predictions are updated hourly using a rolling window of historical observations.
References
[2] ACCESS Dashboard, Mubadala Arabian Center for Climate and Environmental Sciences (ACCESS). https://dashboard.nyuadmaccess.org
[3] Hersbach, H., et al. (2023): ERA5 hourly data on single levels from 1940 to present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS). DOI: 10.24381/cds.adbb2d47
1. Hourly Forecast for Indoor PM2.5
6-Hour Forecast
This forecast displays predicted indoor PM2.5 concentrations for the next 6 hours (Horizon 1 = 1 hour ahead, Horizon 6 = 6 hours ahead). Predictions are generated hourly at the top of each hour, combining real-time indoor sensor data with outdoor air quality and meteorological conditions. Use this forecast to plan ventilation schedules, outdoor activities, or take preventive measures when elevated PM2.5 levels are expected.
Indoor PM2.5 Forecast
Next 6-hour CatBoost prediction - Predicted at 2026-04-19 17:00:00 (Dubai Time, UTC+4)
2. Performance of the Model
Actual vs Predicted Data for PM2.5
The chart below compares predicted PM2.5 values against actual sensor measurements over time. Each prediction made 1-6 hours in advance is validated against the corresponding real measurement once it becomes available. This continuous validation allows you to assess model reliability and understand how prediction accuracy varies across different forecast horizons and environmental conditions.
Horizon 1: Forecasted vs Actual (Last 48 Hours)
Comparison of forecasted and actual PM2.5 values at the first prediction horizon
Prediction Error
Root Mean Square Error across different forecast horizons (1-6 hours ahead)
