Enhancing PM2.5 Forecasts with Machine Learning: Comparative Analysis of Global and Regional Models Integrated with Low-Cost Sensor Networks

Machine Learning (ML) techniques have demonstrated significant potential in forecasting air pollution events, offering skillful predictions with relatively low computational costs. This study expands the scope of Particulate Matter (PM2.5) forecasts in the Aburra Valley, Colombia, by comparing the performance of global and regional meteorological model experiments that integrate both official monitoring stations and low-cost sensor data. We employed ML algorithms including Artificial Neural Networks, Random Forest, Gradient Boosting, and Support Vector Regression to forecast PM2.5 concentrations for the Air Quality Monitoring Network of the Aburra Valley. The models were trained for each forecast horizon up to 72 hours using data from three distinct meteorological sources: the Global Forecast System (GFS), the European Centre for Medium-Range Weather Forecasts (ECMWF), and the Weather Research and Forecasting (WRF) model implemented by the Metropolitan Area of the Aburrá Valley through its Early Warning System of Medellin and the Aburra Valley (SIATA). To investigate the added value of comprehensive air quality monitoring, we developed a novel approach incorporating PM2.5 measurements from official reference stations and real-time information from several low-cost sensors near each measurement station. This resulted in six experimental setups: GFS, ECMWF, WRF, and each of these combined with official and low-cost sensor data (GFS+Sensors, ECMWF+Sensors, WRF+Sensors). All models integrated satellite-derived data, including Aerosol Optical Depth from CAMS. The forecast performance was rigorously evaluated, with particular attention to critical air quality episodes characterized by elevated pollution levels, typically occurring during specific periods of the year. This evaluation was designed to ensure model robustness during the most challenging management periods, thus providing crucial information for environmental authorities' decision-making processes. Our results demonstrate that integrating official and low-cost sensor data significantly improved forecast accuracy across all meteorological model inputs, particularly in the first 24 hours. On the other hand, the WRF-based model consistently outperformed global models in longer-term forecasts (48–72 hours), highlighting the value of high-resolution regional modeling in complex terrain. The WRF+Sensors configuration emerged as the most skillful overall, suggesting a synergistic benefit from combining regional meteorological forecasts with dense, multi-tiered air quality observations. This study underscores the potential of integrating diverse data sources and ML techniques to enhance air quality forecasts. Our findings not only provide valuable insights for developing robust forecasting systems that leverage both official and low-cost monitoring networks but also offer a replicable framework for cities worldwide to improve air quality management. This advanced tool enables environmental authorities to make more informed and timely decisions, especially during critical pollution periods, thus contributing to more effective public health protection in urban environments.