Interpretable Machine Learning for PM2.5 Forecasting Using a Long-Term IoT Environmental Dataset in Tropical Thailand

This study investigates short-term PM2.5 prediction using data obtained from an Internet of Things (IoT)-based air quality monitoring network deployed in Thailand. The dataset comprised approximately 191,000 observations recorded between 2024 and 2026. The monitored variables included PM1, PM2.5, PM10, temperature, relative humidity, rainfall, wind speed, wind direction, and light intensity. The objective was to examine the capability of long-term IoT sensor data to support PM2.5 prediction and to identify the environmental factors that influence the prediction results.

Three ensemble machine learning algorithms were evaluated: Random Forest, Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). Their performance was assessed using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). SHapley Additive exPlanations (SHAP) were applied to quantify the contribution of each particulate matter and meteorological variable to the predicted PM2.5 concentration.

The main contribution of this study is the integration of long-term field measurements from an IoT air quality sensor network with comparative ensemble learning and model-level interpretation. Unlike approaches that focus mainly on prediction accuracy, this study also examines how individual sensor variables contribute to the prediction of PM2.5. The novelty lies in the use of approximately two years of continuously collected tropical environmental data together with three ensemble algorithms and SHAP-based interpretation within a single forecasting framework. The results indicate that the evaluated models can capture patterns associated with PM2.5 variation, while the SHAP analysis provides insight into the relative influence of particulate matter and meteorological variables. The proposed approach therefore provides both predictive capability and interpretable information for localized air quality monitoring and warning applications.