Machine learning-enhanced air quality forecasting and trend analysis: a five-year comprehensive assessment of PM_2.5 concentrations in Bangkok, Thailand

Air pollution forecasting is crucial for protecting public health in rapidly urbanizing Asian megacities; however, comprehensive comparative studies of advanced machine learning approaches are limited in Southeast Asian urban environments. This study developed and systematically compared three state-of-the-art machine learning algorithms for operational PM_2.5 forecasting in Bangkok, Thailand, using comprehensive monitoring data from 2020 to 2024. Daily PM_2.5 concentrations and meteorological variables, including temperature, rainfall, wind speed, atmospheric pressure, and relative humidity, were collected from 12 monitoring stations across Bangkok. Three machine learning approaches were implemented and compared: Random Forest (RF), Gradient Boosting (GB), and Long Short-Term Memory (LSTM) neural networks. Advanced feature engineering incorporated temporal lags, moving averages, and cyclical encoding to capture seasonal and temporal dependencies. The dataset comprised 1,827 daily observations across all variables. PM_2.5 concentrations exhibited pronounced seasonal variations, with a mean of 21.89 ± 8.71 μg/m³, ranging from winter highs of 29.78 ± 7.88 μg/m³ to rainy-season lows of 14.58 ± 3.16 μg/m³. Strong positive correlations were observed between PM_2.5 and atmospheric pressure (r = 0.473), while negative correlations were found with rainfall (r = -0.260) and relative humidity (r = -0.237). Gradient Boosting demonstrated superior predictive performance, with an RMSE of 2.17 μg/m³ and an R² of 0.935 on an independent external validation dataset comprising 365 days of 2024 data, withheld entirely from model development, confirming genuine generalization to unseen future data. Random Forest achieved RMSE = 3.34 μg/m³ and R² = 0.845 on the same external validation set. To address overfitting identified in preliminary analyses (training R² = 0.964), hyperparameter regularization was substantially strengthened, yielding R² degradation of only 3.9% (Gradient Boosting) and 5.6% (Random Forest) from training to external validation. Feature importance analysis revealed that PM_2.5 temporal features dominated the predictions, with the 3-day moving average achieving the highest importance in Random Forest (42.17%) and Gradient Boosting (60.41%). Short-term forecasting performance (1-7 days) met operational requirements for early warning systems, but performance degraded significantly beyond 14 days. The validated Gradient Boosting framework provides immediate applicability for environmental agencies across Southeast Asian urban centers, supporting evidence-based air quality management and public health protection in rapidly developing megacities.