https://github.com/oliverguhr/transformer-time-series-prediction

proof of concept for a transformer-based time series prediction model
https://github.com/oliverguhr/transformer-time-series-prediction

Last synced: 23 days ago
JSON representation

proof of concept for a transformer-based time series prediction model

• Host: GitHub
• URL: https://github.com/oliverguhr/transformer-time-series-prediction
• Owner: oliverguhr
• License: mit
• Created: 2020-02-27T09:39:52.000Z (almost 5 years ago)
• Default Branch: master
• Last Pushed: 2023-02-20T14:48:50.000Z (almost 2 years ago)
• Last Synced: 2023-10-20T22:10:31.982Z (about 1 year ago)
• Language: Python
• Homepage:
• Size: 3.15 MB
• Stars: 977
• Watchers: 13
• Forks: 210
• Open Issues: 4
• Metadata Files:
  • Readme: readme.md
  • License: license.txt

Awesome Lists containing this project

• awesome-of-time-series-prediction - Transformer Time Series Prediction

README

        
# Transformer Time Series Prediction

This repository contains two Pytorch models for transformer-based time series prediction. Note that this is just a proof of concept and most likely not bug free nor particularly efficient. 

* transformer-singlestep.py contains a single-step prediction model

* transformer-multistep.py contains a multi-step prediction model

This is the output of a single step prediction model that has been trained for 100 epochs on the [daily minimum temperature dataset](./daily-min-temperatures.csv).

**For any serious use case, head over to the [flow-forecast](https://github.com/AIStream-Peelout/flow-forecast) package.**

![image](docs/result.gif)

The model is also capable of multi-step predictions. In this case, the model predicted 200 steps based on 100 steps input. It was able to learn long term trend from the training data:

![image](docs/transformer-future200.png)

(input blue, prediction red)