ILoveStudying / AM-RGCN

A deep learning model for traffic flow forecasting

Augmented Multi-component Recurrent Graph Convolustional Network for traffic flow forecasting (AM-RGCN)

Dataset

The public traffic datasets, PEMSD4 and PEMSD8, are the real highway traffic datasets in California released by Guo (ASTGCN). The observations of the sensors are aggregated into 5-minute windows, including three dimensions of time-stamped total traffic flow, average speed, and average occupancy. Moreover, the geographic information of the sensors is also contained.

• PeMSD4 records two months of statistics on traffic flow, ranging from Jan 1st 2018 to Feb 28th 2018, including 307 sensors on the highways of San Francisco Bay. We choose data on the first 50 days as the training set and valid set, and the remaining 9 days as the test set.

• PeMSD8 contains two months of statistics on traffic flow, ranging from July 1st 2016 to Aug 31st 2016, including 170 sensors on the highways of San Bernardino. We select data on the first 50 days as the training set and valid set, and the remaining 12 days as the test set.

• visual/showdata.npy. Didi’s real-world traffic flow data ranging from 31 October 2019 to 30 November 2019 in Beijing (a small district which has been masked).

Parameter Setting

The detail setting of our experiment refers to our paper.

CUDA memory-usage: >7GB for PEMSD8; >12GB for PEMSD4. You can reduce the batch_size if necessary.

Usage

You need edit the options in opt.py:

• dataset: str, choose pems04 or pems08
• save_path: str, checkpoint path for model
• adj: str, the path of adjacency matrix, distance08.csv or distance04.csv
• Multidataset: str, whether there exists Multidataset, create one automaticly if not
• process_method: str, MultiComponent or SlideWindow
• hdwps: str, hour(h), day(d), week(w), and shift(s) are multiples of prediction(p) length
• model: AM-RGCN, Baseline_LSTM, Baseline_GRU, MCSTGCN, ASTGCN, DM_LSTM_GCN. We provide the realizations of six models. More details of these models can be found in the paper.

if model is in (AM-RGCN, MCSTGCN, ASTGCN, DM_LSTM_GCN)

 python Multi_train.py
 
 python Multi_test.py

if model is in (Baseline_LSTM, Baseline_GRU) the process method is SlideWindow

python lstm_gru_train.py

python lstm_gru_test.py

Citation

Zhang, C.; Zhou, H.-Y.; Qiu, Q.; Jian, Z.; Zhu, D.; Cheng, C.; He, L.; Liu, G.; Wen, X.; Hu, R. Augmented Multi-Component Recurrent Graph Convolutional Network for Traffic Flow Forecasting. ISPRS Int. J. Geo-Inf. 2022, 11, 88. https://doi.org/10.3390/ijgi11020088

@Article{ijgi11020088,
AUTHOR = {Zhang, Chi and Zhou, Hong-Yu and Qiu, Qiang and Jian, Zhichun and Zhu, Daoye and Cheng, Chengqi and He, Liesong and Liu, Guoping and Wen, Xiang and Hu, Runbo},
TITLE = {Augmented Multi-Component Recurrent Graph Convolutional Network for Traffic Flow Forecasting},
JOURNAL = {ISPRS International Journal of Geo-Information},
VOLUME = {11},
YEAR = {2022},
NUMBER = {2},
ARTICLE-NUMBER = {88},
URL = {https://www.mdpi.com/2220-9964/11/2/88},
ISSN = {2220-9964},
DOI = {10.3390/ijgi11020088}
}