Short-Term Monitoring and Machine Learning-Based Forecasting of PM2.5 in Dhulikhel Using a Calibrated Mid-Cost Sensor and Meteorological Parameters

Abstract

This study presents a short-term monitoring and forecasting approach for fine particulate matter (PM2.5) concentrations in Dhulikhel Municipality, Kavrepalanchok, Nepal, using calibrated mid-cost sensor and meteorological parameters. A 10-day co-location calibration of the mid-tier sensor (Aeroqual AQM 65) against the reference-grade GRIMM EDM 180 monitor yielded a strong linear correlation (R² = 0.887), and a resulting correction equation was applied to all subsequent data. Meteorological data and hourly PM2.5 were collected over 42 days (April 22 to June 2, 2025). Mean PM2.5 over study period was 22.62 μg/m³, with 90% of the monitoring days exceeded the safe limit of World Health Organization (WHO) 24-hour PM2.5 guideline values of 15 μg/m³. Pollution rose analysis through software ‘R’ identified dominant pollutant transport from east-southeast and west-northwest directions, indicating contributions from road traffic corridors and regional transport from nearby cities. Four machine learning models: Random Forest, Extra Trees, eXtreme Gradient Boosting (XGBoost), and a Stacking ensemble were trained and evaluated for 1–6 hour ahead forecasting. The Stacking model demonstrated the most consistent performance across all forecast horizons, achieving the highest performance (R² = 0.844, RMSE = 3.29 μg/m³ for current hour estimation using combined inputs). Inclusion of meteorological parameters significantly enhanced model accuracy. These findings confirm the feasibility of combining calibrated mid-cost sensors with ensemble machine learning for reliable short-term air quality forecasting in resource-scare region, supporting evidence-based pollution management and public health interventions.