Forecasting task. Predicting temperatures from sensors data.
A repository which contains the implementation for a forecasting task.
The task: Given the previous 5 hours, predict the temperatures recorded from a set of sensors in in next hour.
The architecture: Youngjoo Seo, Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst, Structured Sequence Modeling With Graph Convolutional Recurrent Networks
I also implemented another architecture similar to the proposed gconvLSTM, with the LSTM cell replaced by a basic RNN cell.
The result:
| Model | k | knn | RMSE |
|---|---|---|---|
| lstm | 1.5619 | ||
| glstm | 1 | 4 | 0.1749 |
| glstm | 2 | 4 | 0.1674 |
| glstm | 3 | 4 | 0.1701 |
| glstm | 4 | 4 | 0.1756 |
| grnn | 3 | 8 | 0.1954 |
| grnn | 1 | 4 | 0.1747 |
| grnn | 2 | 4 | 0.1700 |
Sensor data, preprocessed. This should be put into datasets/japan.
Requirements:
Python 2.7Tensorflow 1.1.0+
Make sure you have the dataset ready. The settings in config.py should be handled. Run:
pip install -r requirements.txtpython gconv_main.py