| Dataset | PeMSD4 | PeMSD7 |
|---|---|---|
| Location | San Francisco Bay Area | District 7 of California |
| Number of sensors | 3,796 | 4,817 |
| Period of time | 1st June 2017 to 30th June 2017 | 1st June 2017 to 30th June 2017 |
| Sampling interval | 5 minutes | 5 minutes |
| Raster size | (42, 34) | (20, 36) |
| Number of available time points | 8,640 | 8,640 |
Notice: The longitude_latitude.csv file contains the latitude and longitude of each sensor, while the June folder contains the raw traffic flow data.
First, the spatiotemporal traffic raster data is converted into the proposed multi-channel data structure with a shape of (B, 3, d, I, J), where B represents the batch size. Then, the input unit learns the periodic dependencies by mapping the multi-channel data structure into bigger output channels (B, C, d, I, J). Next, the backbone captures spatiotemporal correlations and channel inter-dependencies without downsampling the feature maps, which helps preserve more information. Finally, the output unit compresses the output channels into one and treats the 3D output feature map as a sequence of consecutive 2D feature maps (B, 1, d, I, J), the prediction of the next time interval is made by compressing d into a single 2D feature map (B, 1, I, J).