Аңдатпа
Accurate short-term forecasting of fine particulate matter (PM2.5) is critical for public health protection in Almaty, Kazakhstan, where complex topography and meteorological conditions intensify pollution episodes. This study benchmarks statistical, tree-based, deep learning, and hybrid models for daily PM2.5prediction using 1,558 observations collected between February 2020 and May 2024. The evaluated methods include ARIMA, LightGBM with lag features, ARIMA+XGBoost, Wavelet+XGBoost, CNN–LSTM with wavelet smoothing, CNN– LSTM–Transformer, GRU, and a Seq2Seq GRU with attention. Missing values were handled using Multiple Imputation by Chained Equations (MICE), and model performance was assessed using MAE, RMSE, and R2under a chronological train–validation split. GRU-based architectures achieved the highest accuracy (GRU: R2= 0.8896, RMSE = 0.2768; Seq2Seq+Attention: R2= 0.8889, RMSE = 0.2776), while wavelet preprocessing enhanced stability in convolutional and boosting models. The results highlight trade-offs between predictive accuracy, interpretability, and computational cost, and suggest future directions including multi-day forecasting, spatiotemporal extensions, explainability via SHAP and attention visualization, and real-time deployment.