Real-Time PM2.5 Forecasting Using Lightweight Ensemble Learning and Temporal Feature Fusion
Accurate and prompt prediction of PM2.5 concentration is crucial to reduce the impacts of air pollution on human health and city ecosystems. In this study, a hybrid ensemble learning model for hourly PM2.5 predictions is proposed, combining advanced data preprocessing, temporal feature engineering, and heterogeneous regression models, which can be computed efficiently. The preprocessing step cleans the data set by discarding incomplete target records, fills the missing meteorological data with median statistics, encodes categorical wind directions and normalizes numerical features. The model learns the temporal dependencies by using several lag variables and several rolling statistical features. The prediction model is an equal weight ensemble of CatBoost, Ridge, Lasso, and ElasticNet regression models; these models are added together to take advantage of their ability to learn non-linearly, regularize their coefficients, and deal with sparsity and correlated features, without the complexity of stacked meta-learning. The proposed approach has been validated by comprehensive experiments, conducted on the benchmark data from Beijing PM2.5, by using ablation analysis, seasonal evaluation, and feature importance assessment. The framework performs well in terms of RMSE (26.80), MAE (18.45), R² (0.892), IA (97.65%) and MAPE (0.535), outperforming other deep learning models such as LSTM, GRU, CNN-LSTM, DAQFF and FPHFA with significantly less computational costs. The results prove its reliability for efficient, easily interpretable and real-time air quality forecasting applications.