Traffic Flow Forecasting Methods Considering Meteorological Factors

Currently, the majority of traffic flow prediction works based on deep learning focus on exploring the spatiotemporal correlations within traffic data itself. However, these works often neglect the influence of external factors, such as weather conditions, on traffic variations or simply incorporate these external factors as part of the model input, failing to fully consider the relationship between external factors and traffic changes. Moreover, most approaches adopt a point estimation method for prediction, which does not account for uncertainties present in the data. To address these issues, this paper proposes a selective cross-attention feature fusion approach. By dynamically assigning correlation weights between weather variables and traffic features, and integrating quantile regression to model uncertainty, the proposed model generates confidence intervals while fusing multisource information. Experiments demonstrate that this method reduces the mean absolute error on the PEMS04 dataset and decreases the prediction interval width by 23% compared to traditional conformal prediction methods, significantly enhancing the practicality and decision-support capabilities of the prediction outcomes. Compared to other benchmark models such as PDFormer, the prediction accuracy has also been improved.