Journal of Electronic Science and Technology, Volume. 22, Issue 1, 100244(2024)
Multi-scale persistent spatiotemporal transformer for long-term urban traffic flow prediction
Figures & Tables(10)
Fig. 1. Structure of the proposed traffic flow prediction model with multi-head attention: (a) data embedding layer, (b) spatiotemporal encoder, and (c) spatiotemporal decoder.
Fig. 5. Results of the ablation experiments: (a) MAE on PeMSD4, (b) MAPE on PeMSD4, (c) RMSE on PeMSD4, (d) MAE on PeMSD8, (e) MAPE on PeMSD8, (f) RMSE on PeMSD8, (g) MAE on NYCTaxi, (h) MAPE on NYCTaxi, (i) RMSE on NYCTaxi, (j) MAE on CHIBike, (k) MAPE on CHIBike, and (l) RMSE on CHIBike.
Fig. 6. Loss curves of the ablation experiments or (a) PeMSD4, (b) PeMSD8, (c) NYCTaxi, and (d) CHIBike.
Table 1. Datasets information.
View tableView in ArticleTable 1. Datasets information.
Dataset Nodes Edges Time steps Time interval Time range PeMSD4 307 340 16992 5 min 01/01/2018−02/28/2018 PeMSD8 170 295 17856 5 min 07/01/2016−08/31/2016 NYCTaxi 75 484 17520 30 min 01/01/2014−12/31/2014 CHIBike 270 1966 4416 30 min 07/01/2020−09/30/2020
Table 2. Experimental results on the PeMSD4 and PeMSD8 graph datasets.
View tableView in ArticleTable 2. Experimental results on the PeMSD4 and PeMSD8 graph datasets.
Model LSTM DCRNN STGCN ASTGCN Graph WaveNet ASTGNN PDFormer MSPSTT MSPSTT_nde Dataset Index *Bold is for the best result, and underline is for the second best. PeMSD4 MAE 26.84 23.67 22.31 22.48 19.35 19.28 18.75 18.77 18.45 RMSE 40.76 37.12 35.05 34.81 31.74 31.27 30.26 30.29 30.04 MAPE (%) 22.39 16.12 14.43 15.89 13.32 13.16 13.00 12.50 12.34 PeMSD8 MAE 22.19 18.25 18.09 18.92 15.07 15.96 14.75 14.16 14.13 RMSE 33.63 28.32 27.98 28.56 23.85 24.92 24.18 23.66 23.49 MAPE (%) 18.77 11.59 11.23 12.58 9.51 10.88 9.50 9.53 9.30
Table 3. Experimental results on the grid datasets.
View tableView in ArticleTable 3. Experimental results on the grid datasets.
Model PDFormer MSPSTT MSPSTT_nde Dataset Index *Bold is for the best result, and underline is for the second best. NYCTaxi MAE 13.47 12.99 13.26 MAPE (%) 26.03 23.57 23.90 RMSE 25.33 24.55 25.44 CHIBike MAE 2.45 2.42 2.33 MAPE (%) 54.72 53.82 51.93 RMSE 4.26 4.21 4.05
Table 4. Training time of three methods (s/epoch).
View tableView in ArticleTable 4. Training time of three methods (s/epoch).
Dataset PDFormer MSPSTT MSPSTT_nde PeMSD4 313.83 2754.37 2535.21 PeMSD8 256.36 2272.16 2144.90 NYCTaxi 81.70 379.37 1770.47 CHIBike 35.87 171.80 165.75
Tools
Get Citation
Copy Citation Text
Jia-Jun Zhong, Yong Ma, Xin-Zheng Niu, Philippe Fournier-Viger, Bing Wang, Zu-kuan Wei. Multi-scale persistent spatiotemporal transformer for long-term urban traffic flow prediction[J]. Journal of Electronic Science and Technology, 2024, 22(1): 100244
Paper Information
Category:
Received: Nov. 7, 2023
Accepted: Mar. 6, 2024
Published Online: Jul. 5, 2024
The Author Email: Xin-Zheng Niu (xinzhengniu@uestc.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20