Adaptive regularized ensemble learning with physically informed features for interpretable PM2.5 prediction

To address the coupling challenges between atmospheric physical processes and spatiotemporal dependencies in predicting the concentration of fine particulate matter with an aerodynamic diameter of 2.5 μm or less (PM 2.5 ), this study proposes an adaptive regularized ensemble learning method with physically informed features for interpretable prediction. First, based on the maximal information coefficient (MIC), reference stations are selected for four target stations (Xuyi, Jiangdu, Ma’anshan, and Jiangyin) from 20 regional PM 2.5 monitoring sites in Jiangsu and Anhui Provinces, China. A total of 36 physically informed features are then constructed, covering five categories, including temporal periodicity, historical evolution, spatial correlation, and meteorological conditions. Subsequently, a heterogeneous model pool consisting of Random Forest, XGBoost, LightGBM, BiLSTM, and CNN–BiLSTM is established as base learners. An adaptive regularization ensemble strategy based on convex optimization is proposed, which dynamically selects L1, L2, or no regularization constraints to learn the optimal weights. Using hourly observational data from October 2023 to December 2025, the proposed ensemble model was evaluated for 1-hour-ahead PM 2.5 concentration prediction. On the test sets of the four target stations, the model achieved an average R 2 of 0.894, an MAE of 5.257 µg m⁻³, and an RMSE of 8.092 µg m⁻³, outperforming the individual base models. Process-oriented SHapley Additive exPlanations (SHAP) analysis indicates that the 1-hour lagged PM 2.5 concentration and the 1-hour change rate are the key predictive factors. This study provides an effective solution for high-accuracy and robust regional PM 2.5 prediction.