A temporal domain generalization method for PM2.5 concentration prediction based on adversarial training and deep variational information bottleneck

Abstract

Accurate prediction of PM_2.5 concentration is crucial for environmental pollution control and public safety. However, PM_2.5 concentration is influenced by various factors and follows a dynamic temporal distribution, potentially reducing the generalization capability of prediction models for future prediction. To address this issue, we propose a novel temporal domain generalization method modeled from a distribution perspective. This method aims to extract invariant information from dynamic and time-varying temporal distributions to improve future prediction by combining adversarial training and deep variational information bottleneck (Deep VIB). Specifically, the training data is first segmented into multiple temporal domains characterized by the largest temporal distribution differences. Deep VIB is then employed to extract fine-grained features, combined with a domain classifier to eliminate domain-specific information through adversarial training. Finally, only the trained feature extractor and predictor are applied to the prediction process. This approach enables the prediction model to simultaneously extract domain-invariant features while minimizing prediction errors, thereby enhancing its generalization capability. Experimental results from real-world data demonstrate that the proposed method outperforms state-of-the-art prediction models, achieving average improvements of 3% in mean absolute error, 7% in root mean square error, and 3% in R-square compared to the best baseline models.