Impact of Datasets Categorized by Weekday and Weekend on PDFormer Performance

This paper investigates the impact of weekday and weekend time classification on the performance of the traffic flow prediction model PDFormer (Propagation Delay-Aware Dynamic Long-Range Transformer). By analyzing the CHBike bike-sharing dataset, it is found that the weekday traffic shows significant morning and evening peak characteristics, and the data fluctuation is small and regular, while the weekend traffic peaks are concentrated in the midday to the afternoon, and the data is more discrete and less stable. The data set is divided into weekday (855,359), weekend (336,960) and complete data set, and the PDFormer model is used for single-step prediction. The results show that the prediction error of the model on weekday data (MAE = 3.67, MAPE = 0.30, RMSE = 5.07) is significantly better than that on weekend (MAE = 4.53, MAPE = 0.32, RMSE = 6.41) and the complete dataset (MAE = 3.85, MAPE = 0.30, RMSE = 5.45), and similar differences were found in both inflow (0-masked) and outflow (1-masked) predictions. The study confirms that the classification of datasets based on temporal patterns improves model performance and that weekday data are easier to model due to their regularity. In contrast, the complexity and sparsity of weekend data lead to increased prediction difficulty. The results provide a reference direction for optimizing the data preprocessing strategy for traffic flow prediction, and the fusion of multi-temporal granularity and multi-source data can be further explored in the future to enhance the robustness of the model.