Background/Aim: Air pollution poses serious risks to human health and the economy, making accurate prediction of particulate matter (PM₁₀) exceedance events essential. This study aimed to forecast PM₁₀ concentrations in Amasya City using the Nonlinear Autoregressive with Exogenous Variables (NARX) model and to compare the effectiveness of different ANN training algorithms.Methods: Daily air quality and meteorological data from the Merzifon, Şehzade, and Suluova monitoring stations were analyzed for the 2016–2018 period. The NARX model was trained using Levenberg–Marquardt (LM), Bayesian Regularization (BR), and Scaled Conjugate Gradient (SCG) algorithms under the same dataset and model structure. Model performances were evaluated using NSE, R², KGE, and RMSE metrics.Results: The Bayesian Regularization algorithm consistently outperformed the LM and SCG algorithms, particularly during the testing phase. The BR model achieved the highest KGE values of 0.715, 0.746, and 0.774 for the Merzifon, Şehzade, and Suluova stations, respectively. It also produced the lowest RMSE values, with a minimum of 15.881 µg/m³, indicating high prediction accuracy and strong generalization capability.Conclusion: The results demonstrate that the Bayesian Regularization-based NARX model provides a robust and reliable framework for PM₁₀ forecasting in nonlinear environmental datasets. Its ability to reduce overfitting and improve generalization makes it a promising approach for air quality prediction studies.
Gülhan Özdoğan-Sarıkoç (Thu,) studied this question.