Flipout Bayesian LSTM with Residual Attention for Uncertainty-Aware PM2.5 Forecasting and Anomaly Detection

Accurate PM2.5 prediction and reliable uncertainty assessments are essential for effective early warnings and public health protection. However, most existing deep learning models only provide deterministic predictions, with limited treatment of predictive uncertainty, which may reduce the reliability under noisy or abrupt pollution conditions. This study presents a flipout Bayesian LSTM with residual attention (FBA-LSTM), which integrates Bayesian flipout inference, residual connections, and temporal attention to jointly improve the predictive accuracy and uncertainty estimation. Unlike MC dropout, our model explicitly represents weight distributions through variational flipout inference, yielding more comprehensive and stable uncertainty estimates with a lower computational cost. A lightweight calibration module based on standard-deviation scaling further aligns the confidence intervals with empirical coverage. Experiments on hourly PM2.5 data from four Nanjing stations (2020) showed that FBA-LSTM improves the accuracy, noise robustness, and exceedance warnings (F1 = 0.996) and achieves a higher anomaly-detection performance (F1 = 0.691) than baseline models and methods, thereby facilitating the realization of urban environmental sustainability and public health security.