Abstract:A lane-level traffic flow prediction method was proposed to refine the spatial granularity of traffic flow prediction and address the limitations in traditional prediction methods which overlooked the interaction mechanisms brought by the lane differences. The soft-dynamic time warping was chosen as the measurement method to consider the variations in traffic flow time series caused by the lane positions. Then, the k-means algorithm was aggregated in the proposed algorithm to classify the lane cross-sections, which was utilized to analyze the temporal characteristics of flow at the selected lane cross-section type. The sequential variational mode decomposition was introduced to reduce the volatility of time series flow data and remove its noise, whose results can serve as inputs for the prediction model. Based on the previous results, the lane correlations can be determined by incorporating the Spearman correlation analysis in the proposed method. The heterogeneous spatiotemporal graph convolutional recurrent neural network model for predicting the flow of various lane types can be established by embedding the multi-head attention mechanism into the bidirectional gated recurrent unit and combining the lane correlations and their distance. The traffic count data from several freeways in Shanxi, China was selected to assess the effectiveness of the proposed method. The results indicate that the freeway′s lanes within the selected area can be classified to four categories: dense type, sparse type, morning peak type, and evening peak type, respectively. Compared to the autoregressive integrated moving average model, long-short term memory network model, and spatio-temporal graph convolutional network models, the proposed model reduces the mean absolute error and root mean square error by 11.21%-24.05% and 8.89%-24.43%, with the r-squared coefficient as 0.962 at the 5-minute granularity flow prediction for all the lanes. This result indicates the accuracy of the flow prediction can be further improved by considering the lane classification. When the prediction was accomplished at the 5-minute granularity with the step size within 12, the mean absolute error and root mean square error of the proposed model increased by 15.82% and 11.99% at most. The findings provide a basis for road planning and the development of intelligent transportation.
