Fig 1.
Source of ideas in this paper.
Fig 2.
Visualization of traffic flow maps for TaxiBJ, BikeNYC datasets.
Fig 3.
Urban gridding and definition of inflow and outflow.
Fig 4.
Inflow flows in four zones over a week.
Fig 5.
Comparison of traffic flow on sunny day and rainstorm.
Fig 6.
The overall framework of the ST-D3DDARN model.
Fig 7.
Comparison between 2D CNN and 3D CNN.
Fig 8.
Comparison between ordinary 3D CNN and decoupled 3D CNN.
Fig 9.
Decoupled 3D denseNet architecture.
Fig 10.
ARN unit internal structure.
Fig 11.
Spatial self-attention structure.
Fig 12.
Coordinate attention mechanism structure.
Fig 13.
External information branch structure.
Fig 14.
Flow chart of traffic flow prediction in the whole city.
Table 1.
Environment configuration.
Table 2.
Description of TaxiBJ and BikeNYC.
Fig 15.
Experimental results with different number of D3DDN layers.
(a) TaxiBJ. (b) BikeNYC.
Fig 16.
Experimental results of different ARN module layers.
(a) TaxiBJ. (b) BikeNYC.
Table 3.
Comparison of model characteristics.
Table 4.
Comparison of model prediction results.
Table 5.
Inflow and outflow prediction results in TaxiBJ.
Fig 17.
Error comparison of different models.
(a) TaxiBJ. (b) BikeNYC.
Fig 18.
RMSE and MAE changes of the test set during the training of each model.
(a) TaxiBJ of RMSE. (b) TaxiBJ of MAE. (c) BikeNYC of MAE. (d) BikeNYC of MAE.
Table 6.
Comparison of model efficiency.
Table 7.
Comparison of model variants of TaxiBJ.
Fig 19.
Performance of direct multi-step predictions for each method.
(a) Step-wise RMSE of TaxiBJ. (b) Step-wise MAE of TaxiBJ. (c) Step-wise RMSE of BikeNYC. (d) Step-wise MAE of BikeNYC.
Fig 20.
Performance of recursive multi-step predictions for each method.
(a) Step-wise RMSE of TaxiBJ. (b) Step-wise MAE of TaxiBJ. (c) Step-wise RMSE of BikeNYC. (d) Step-wise MAE of BikeNYC.