Time series approaches for air pollution forecasting in Çanakkale: A performance comparison of ETS and Box-Jenkins models

Abstract

Air pollution is a complex process that requires statistical modeling and forecasting due to the temporal variability in the concentrations of different pollutant components in the atmosphere. This study analyzes monthly concentrations of PM10, PM2.5, SO2, NO2, and O3 measured in Çanakkale between 2019 and 2024 using time series methods to generate future forecasts. Factor analysis was first applied to identify the underlying structure among pollutants and to select the most influential variables for modeling; as a result, PM10 and SO2 were found to load most strongly on the primary factor associated with air pollution in the region. For these two pollutants, which were examined in terms of stationarity, trend, and seasonal components, forecasting models were developed based on Exponential Smoothing techniques (Simple, Holt, Damped, Holt–Winters additive and multiplicative) and the Box–Jenkins methodology (ARIMA, SARIMA). Model performance was evaluated using RMSE, MAE and MAPE for both in-sample and out-of-sample periods. The Holt–Winters multiplicative model achieved the highest forecasting accuracy, with MAPE values of 11.040% for PM₁₀ and 20.253% for SO₂ by effectively capturing trend and proportional seasonality. This study is distinctive in that it provides a comparative evaluation of ETS and Box–Jenkins approaches for multiple air pollutants and develops region-specific forecasting models for Çanakkale, thereby offering novel methodological and practical insights for air quality assessment in Türkiye.