Sustech-yx / qr_forcaster

Our implementation of the paper "A Multi-Horizon Quantile Recurrent Forecaster"

Multi-Horizon Quantile Regression Forecaster

This is an implementation of the paper A Multi-Horizon Quantile Recurrent Forecaster. The paper describes an endoder decoder quantile regression model for time series prediction. The implementation was tested on electric data consumption dataset. Our model is implemented using Pytorch Lightning

Training

In order to train the model, run the following command:

python3 train.py

best trained model will be saved to "/trained_models/" and the train/test dataloaders are saved in "/dataloaders/". It is important to use the dataloaders that were created with the model to avoid data leakage in test time.

To enable forking sequence, in train.py, change forking to True. It is recommended to reduce batch size, since the forking sequence is implemented "on the fly" and requires high GPU memory.

Testing

testing the trained can be done by running:

python3 test_pred.py

the script will choose the best model from the trained models folder, will sample a series from the dataloader, predict forecasting and plot it

Prediction Results: