PM2.5/PM10 Forecasting System with Benchmarking of 44 Machine Learning Algorithms and Ensemble Learning Approaches

Air pollution from particulate matter (PM2.5 and PM10) poses a serious public health risk in urban environments, particularly in areas with heavy vehicular traffic. Against this backdrop, the present study proposes an Internet of Things (IoT)-based system designed to support air quality monitoring and evidence-based decision-making regarding PM2.5 and PM10 concentrations, integrating low-cost sensors with a machine learning prediction module. The study follows an experimental-applied design with a quantitative-comparative approach. Its scientific contribution is organized around an integrated IoT-ML framework addressing a concrete gap in the literature: the lack of local empirical evidence regarding which family of machine learning algorithms delivers the greatest accuracy, stability, and computational efficiency for particulate matter forecasting in mid-altitude urban environments using low-cost sensors. On one hand, the framework proposes and deploys a four-node IoT network for continuous PM2.5 and PM10 monitoring in high-traffic urban microenvironments-representing one of the first sustained deployments with low-cost, high-temporal-resolution sensors (10-min intervals) in Arequipa, Peru. On the other hand, the study presents the most extensive benchmarking reported in the local literature: a systematic evaluation of 44 machine learning algorithms under homogeneous experimental conditions, covering classical statistical models, traditional machine learning techniques, deep learning architectures, and hybrid approaches, along with an analysis of ensemble learning strategies using Ridge stacking and K-Fold cross-validation. This unified comparative analysis-applying consistent metrics (MAE, RMSE, R², and MAPE), the same prediction horizon, and a shared dataset-provides replicable empirical evidence that had not previously been reported for the urban context of Arequipa. The results show that traditional statistical models perform poorly overall, while tree-based and boosting algorithms consistently achieve R² values above 0.90 for both pollutants. Ensemble models, particularly stacking with Ridge regression and cross-validation, yielded the strongest overall performance, demonstrating greater robustness and prediction stability. Explainability criteria were also incorporated, enabling an assessment of each base model's individual contribution and identifying the variables most relevant to the prediction process. The methodological contribution provides future researchers with a rigorous reference framework for algorithm selection in environmental IoT systems. Taken together, the findings demonstrate that combining low-cost IoT networks with advanced machine learning and ensemble learning techniques constitutes an effective, scalable, and cost-efficient alternative for air quality monitoring, predictive analysis, and the support of informed mitigation strategies in urban environments.