PM10 Concentration Forecasting: A Comparative Evaluation of Deep Learning and Time Series Methods

Air pollution has become an important public health issue in urban centers; therefore, PM10 is frequently utilized as one of the most common indicators of air quality due to its relationship with respiratory and cardiovascular illnesses. In this paper, we present a comparison of time series forecasting methods based on three years of daily PM10 data collected in the Kadıköy District of Istanbul (2022-2024). We also evaluate classical time series models (Prophet & SARIMA), and deep learning-based models (Bi-LSTM, LSTM, & GRU). All models were tested under similar conditions. The performance of each model was evaluated by four different metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R²), Mean Absolute Percentage Error (MAPE), and Mean Absolute Relative Error (MARE). Although all models showed very limited explanatory capacity for the reasons that can be attributed to the difficulty in forecasting the daily PM10 concentrations by only utilizing historical information, our findings show that the GRU and LSTM models achieved significantly better results than the classic models in terms of lower error values. Therefore, these two models have shown some superiority in capturing the nonlinear patterns at the short term in temporal sequences. On the other hand, the R² values of both models were very low. Therefore, these models could explain only a minor fraction of the variability of the PM10. The Bi-LSTM model performed poorly because it had increased complexity and decreased generality. In addition to having poor predictive performance, the SARIMA and Prophet models provided the highest errors in terms of predictions for the complex and nonlinear structure of the PM10 data. In summary, although deep learning-based models outperform the classic models slightly, their predictive performances remain constrained due to the constraints caused by the limitation of the data and unobserved variables affecting PM10. This study provides a realistic reference point for urban air quality forecasting and emphasizes the necessity of additional explanatory variables and advanced spatiotemporal modeling in future studies.