Hybrid Network Models Based on SHAP Interpretability for PM2.5 Concentration Prediction: A Case Study of Beijing

Accurate prediction of PM2.5 concentrations is essential for public health protection and environmental management. However, existing forecasting models face critical challenges in capturing complex temporal dependencies and providing interpretability for decision-making. This study proposes a novel hybrid framework integrating one-dimensional convolutional neural networks (1D CNN), extended Long Short-Term Memory (xLSTM), and attention mechanism (CNN-xLSTM-AM), combined with SHapley Additive exPlanations (SHAP) for interpretable PM2.5 forecasting. The 1D CNN unit captures local temporal relationships of multivariate sequences. The xLSTM design, based on sLSTM-mLSTM-sLSTM blocks, addresses limitations of traditional LSTM architectures through exponential gating and dual memory structures: sLSTM captures the state-tracking dynamics of scalar memory mixing, and mLSTM captures multivariate dependencies using matrix memory with larger capacity. The attention mechanism assigns dynamic weights to temporal features according to the importance of each time step. Extensive experiments on Beijing hourly data (2020–2021) show that CNN-xLSTM-AM achieves an RMSE of 9.845 μg/m3 and an R2 of 0.977 for single-step prediction. Compared with CNN-xLSTM, it reduces RMSE by approximately 8.5% in single-step prediction and by 8.9% and 12.3% at the 2 h and 10 h forecasting horizons, respectively. The SHAP interpretability analysis identifies atmospheric pollutants as the highest contributors to predictive power (57.09%), with carbon monoxide as the top feature (32.791%). SHAP dependence analyses further suggest nonlinear associations between the input factors and model predictions, with potential turning points identified for selected variables, including carbon monoxide, relative humidity, wind speed, and wind direction. These model-derived patterns indicate potential changes in predictive relationships but should not be interpreted as definitive physical thresholds or universal regime shifts in PM2.5 behavior. These results demonstrate the potential of combining predictive modeling with post hoc interpretation to support environmental decision-making and may inform the development of targeted emission-control strategies and threshold-informed early-warning approaches.