Solution overview

I used Encoder-Decoder Wavenet architecture similar to @sjv's awesome solution for Web Traffic Time Series Forecasting.
I implemented it by Chainer

Setup

1. Run setup_dirs.sh to make directories.
2. Download dataset and put them into data/input/.
3. Download weather data from huntermcgushion/rrv-weather-data.
   Extract 1-1-16_5-31-17_Weather.zip to a directory 1-1-16_5-31-17_Weather.
   Put the directory and weather_stations.csv into data/input/.

How to run

Preprocess the weather data.

python prepare_weather.py

Run the training script.

python seq_run.py

It calls run.py several times with some configurations and various seeds.
It saves results as one folder per one run in data/output/.
After the end of training, gather the result folders into one folder.
Then run the ensemble script.

python ensemble.py --target_dir "path/to/dir/" --without_valid

Environment

Python 3.6.3 :: Anaconda custom (64-bit)

numpy==1.14.0
chainer==3.3.0
pandas==0.20.3
matplotlib==2.1.2
geopy==1.11.0
scikit-learn==0.19.1

References

• sjv's original implementation
  https://github.com/sjvasquez/web-traffic-forecasting
• WaveNet: A Generative Model for Raw Audio https://deepmind.com/blog/wavenet-generative-model-raw-audio/