Аннотация
Accurate prediction of PM2.5concentrations is essential for public health protection in heavily polluted urban areas such as Almaty, Kazakhstan. This study presents a novel comparative assessment of statistical, machine learning (ML), deep learning (DL), and hybrid models for PM2.5forecasting, addressing a gap in the literature concerning model performance under different data preprocessing scenarios. Using a comprehensive dataset spanning 2020–2024, nine models—including SARIMA, Random Forest, SVM, LSTM, CNN, CNN-ELM, and CNN-LSTM—were evaluated across three imputation techniques: mean imputation, time-based mean, and Multiple Imputation by Chained Equations (MICE). The methodology emphasizes how the quality of missing value treatment influences model accuracy. Results demonstrate that Random Forest performed best with mean imputation (R2= 0.7277), CNN-ELM achieved optimal performance under time-based imputation (R2= 0.8279), and LSTM yielded the highest accuracy when applied to MICE-imputed data (R2= 0.9570). The findings contribute to the development of robust air quality forecasting systems by highlighting the interplay between preprocessing strategies and model selection.