Table of Contents:

• Google TiDE implementation
  • Data Pre-process
    • Patching
  • Net Framework
  • Metric Result
    • My Result
      • MSE and MAE Table
      • Traffic MSE and MAE
      • Traffic Val Curve
  • Summary
  • Reference

Google_TiDE_implementation

An unoffical pytorch implementation of (Google）Long-term Time Series Forecasting with TiDE: Time-series Dense Encoder
Link to paper: Long-term Forecasting with TiDE: Time-series Dense Encoder
Official implemention:(https://github.com/google-research/google-research/tree/master/tide)

Data Pre-process

[Dataset]

Patching

The core idea of this data pre-process is Patching, which is similar to PatchTST and many other time series forcasting jobs. Specifically, they both divide the time series into a number of time segments, each of which is considered a token .

In contrast to PatchTST and some MLP-based forecasting models, it makes use not only of past series values (LookBack) but also of information on covariates such as static covariates (Attributes, No change in relative time,use the ID of time-series usually.) and dynamic covariates (Dynamic Covariates minute of the hour, hour of the day, day of the week etc) that are known at any point in time(Like holiday, hour, day etc).

Net Framework

The overall architecture of the TiDE model is shown in the diagram above. Like PatchTST, it assumes that the channels are independent. This means that multivariate forecasting is transformed into multiple univariate forecasts with shared model parameters.

In the last hyperparameters table of Paper, each dataset has a little different network structure, I tried some other network structures also have a not bad generalisation, Maybe not so SOTA.

Metric Result

My Result

MSE and MAE Table

Predict all chanel

The result wasn't quite SOTA.

Traffic MSE and MAE

Traffic Val Curve

Predict 720 steps and 96 steps ('OT' after data scale transform)

Summary

With about the same performance as PatchTST, TiDE is very efficient. As shown below, both inference and training times are very fast, and the space complexity is low enough to handle very long sequences without out of memory (only use a T4 GPU).

Last

This study shows that self-attention might not be necessary to learn the periodicity and trend patterns at least for these long-term forecasting benchmarks.

Reference

https://github.com/google-research/google-research/tree/master/tide

https://github.com/yuqinie98/PatchTST

https://zhuanlan.zhihu.com/p/624828590

https://github.com/zhouhaoyi/Informer2020

https://github.com/ts-kim/RevIN

https://github.com/HenryLiu0820/TiDE