Performance analysis of classical time series and machine learning models for PM2.5 and PM10 forecasting: A case study at an air quality monitoring station in Udon Thani, Thailand

Air pollution forecasting is a critical component of air quality management, especially in areas experiencing persistent particulate matter issues. However, in practice, the performance of forecasting models depends on local climatic conditions and human activities in each area, resulting in varying forecasting patterns based on spatial contexts. This research aimed to compare the performance between time-series methodologies and machine learning in forecasting hourly PM2.5 and PM10 concentrations, focusing on analysis at air quality monitoring stations in Udon Thani province, Thailand, a medium-sized urban area influenced by both local emissions and regional meteorological patterns. This study used hourly datasets collected from January 1, 2023, to November 30, 2024, comprising a total of 16,800 records obtained from the pollution control department's monitoring station at Nong Phra Chak Park. Since missing values are frequently encountered in real-world data collection, a linear interpolation method was employed to handle these gaps and enhance the efficiency of the proposed methods. The tested models included three time-series methods—seasonal autoregressive integrated moving average (SARIMA), seasonal naïve, and Holt–Winters exponential smoothing (ETS)—and four machine learning methods: random forest (RF), support vector regression (SVR), artificial neural networks (ANN), and extreme gradient boosting (XGBoost). The study utilized 80% of the data for training and 20% for testing, with performance evaluated through MAE, RMSE, R², MASE, and sMAPE indicators. The findings reveal that while time-series models can effectively reflect basic seasonal structures, they possess limited capability in capturing short-term fluctuations and sudden pollution surges, with MAE ranging from 8.096 to 14.106. In contrast, machine learning models consistently demonstrated lower forecasting errors, particularly the support vector regression methodology, which provided the most accurate and stable performance for both PM2.5 (MAE = 2.4899) and PM10 (MAE = 3.7340). This demonstrates its efficiency in modeling nonlinear relationships and short-term dynamics under the environmental conditions of Udon Thani province. The experimental results suggest that forecasting models aligned with local characteristics can yield reliable predictions, which will benefit environmental agencies and policymakers in developing air quality surveillance systems and implementing mitigation measures that effectively reflect regional specificities.