Traffic Flow Prediction

Traffic Flow Prediction with Neural Networks(SAEs、LSTM、GRU).

Requirement

• Python 3.6
• Tensorflow-gpu 1.5.0
• Keras 2.1.3
• scikit-learn 0.19

Train the model

Run command below to train the model:

python train.py --model model_name

You can choose "lstm", "gru" or "saes" as arguments. The .h5 weight file was saved at model folder.

Experiment

Data are obtained from the Caltrans Performance Measurement System (PeMS). Data are collected in real-time from individual detectors spanning the freeway system across all major metropolitan areas of the State of California.

device: Tesla K80
dataset: PeMS 5min-interval traffic flow data
optimizer: RMSprop(lr=0.001, rho=0.9, epsilon=1e-06)
batch_szie: 256 

Run command below to run the program:

python main.py

These are the details for the traffic flow prediction experiment.

| Metrics | MAE | MSE | RMSE | MAPE | R2 | Explained variance score | | ------- |:---:| :--:| :--: | :--: | :--: | :----------------------: | | LSTM | 7.21 | 98.05 | 9.90 | 16.56% | 0.9396 | 0.9419 | | GRU | 7.20 | 99.32 | 9.97| 16.78% | 0.9389 | 0.9389| | SAEs | 7.06 | 92.08 | 9.60 | 17.80% | 0.9433 | 0.9442 |

Reference

@article{SAEs,  
  title={Traffic Flow Prediction With Big Data: A Deep Learning Approach},  
  author={Y Lv, Y Duan, W Kang, Z Li, FY Wang},
  journal={IEEE Transactions on Intelligent Transportation Systems, 2015, 16(2):865-873},
  year={2015}
}

@article{RNN,  
  title={Using LSTM and GRU neural network methods for traffic flow prediction},  
  author={R Fu, Z Zhang, L Li},
  journal={Chinese Association of Automation, 2017:324-328},
  year={2017}
}

Copyright

See LICENSE for details.