Add Temporal Fusion Transformer (TFT) for Multivariate Time Series Forecasting

[kenchiayy]

Summary

I want to suggest adding support for the Temporal Fusion Transformer (TFT) architecture to this repository. TFT is a state-of-the-art model developed by Google Research that combines the best of LSTM and attention mechanisms for interpretable multivariate time series forecasting.

Motivation

• TFT is especially useful for handling static covariates, time-varying inputs, and known future inputs in one unified framework.
• It provides interpretable attention weights, which are valuable for understanding model behavior in real-world applications.
• It is widely used in financial, energy, and operational forecasting and would complement the existing architectures in this repo.

Implementation Options

There are a few PyTorch-based implementations available:

• PyTorch Forecasting (by jdb78) – mature, well-documented
• Darts (by Unit8) – includes TFT among other forecasting models

Benefits

• Enhances the coverage of modern forecasting architectures in this repo
• Offers a strong benchmark for interpretable deep learning models
• Aligns well with existing LSTM and Transformer-based baselines in the project

Would love to hear your thoughts on whether this fits the scope of the project, and if so, I’d be happy to help contribute the integration!

[LongxingTan]

Hi @kenchiayy

That's a really nice suggestion!
Actually I once thought to consider tft, and you could even find the draft here:
https://github.com/LongxingTan/Time-series-prediction/blob/master/tfts/models/tft.py

But I never checked its correctness or tested in dataset. If you are interested, feel free to optimize or reimplement it.
If any help you need from my side, feel free to ask.

Best regards,
Longxing