This repository, a fork of DCRNN, is my response to the challege of Traffic Management.

1. Clone this repository.
2. Place the test CSV file similar to the given training.csv in the input folder in grab_data folder. (grab_data/input/test.csv)
3. Open the dcrnn_config.yaml with a text editor such as notepad++ and set the latest_timepoint as follows.

To predict the demand values just after the latest timepoint in the dataset:

latest_timepoint:
  day_hour_min_option:
    set_day_hour_min: false

To set timepoint "T" to day 61 22:30, for example, to predict demand at 22:45, 23:00, 23:15, 23:30, 23:45.

latest_timepoint:
  day_hour_min_option:
    set_day_hour_min: true
    day: 61
    hour: 22
    min: 30

4. Run dcrnn_main.py. The prediction output will be saved at grab_data/output.

Open dcrnn_config.yaml and set test_only to false.

The DCRNN methodology use both time-series and spatial dependency (relationship between geohash locations) to forecast future values.

• Refactor.
• Integrate separate modules into a single module.
• Adapt preprocessing and postprocessing to the data at Traffic Management. Compute the distances between geohash locations and generate adjacency matrix (1 for itself, 0 for far locations.)
• Sample 48 timeslots (4 x 12 slots of 15 mins: 3 hours just before "T", 3+3 hours 1 week prior to "T", and the first 3 hours in the historical 2 weeks) from the historical sequence of 2 weeks (1344 slots of 15 minutes).
• Linear cosine decay of learning rate.

• Python>=3.5.3
• scipy>=1.3.0
• numpy>=1.16.4
• pandas>=0.24.2
• pyyaml>=3.13
• tensorflow>=1.13.0
• geopy>=1.20.0
• python-geohash>=0.8.5

Dependency can be installed using the following command:

pip install -r requirements.txt

• DCRNN (https://github.com/liyaguang/DCRNN)\
• Yaguang Li, Rose Yu, Cyrus Shahabi, Yan Liu, Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting, ICLR 2018.
• Linear Cosine Decay (https://www.tensorflow.org/api_docs/python/tf/train/linear_cosine_decay)