PPDformer: Channel-Specific Periodic Patch Division for Time Series Forecasting

This repository contains the implementation of PPDformer, a novel deep learning model designed for multivariate time series forecasting (MTS).

Overview

The PPDformer addresses critical challenges in MTS forecasting:

• Channel-Independent Denoising: It applies adaptive noise cancellation and Fourier Transform techniques to individually denoise each channel. This improves accuracy by preserving important details and removing noise more effectively.
• Periodic Patch Division: It segments time series data based on periodic patterns detected per channel. This segmentation allows for better identification of temporal dependencies and local information.

Installation

To get started, clone the repository and install the required dependencies:

git clone https://github.com/yourusername/PPDformer.git
cd PPDformer
pip install -r requirements.txt

Model Architecture

PPDformer is composed of the following major modules:

• CI Denoising: This module denoises each channel using FFT and adaptive noise cancellation.
• Periodic Patch Division: It aligns the time series data by clustering periods and then converting them into two-dimensional patches.
• Dual Attention Mechanism: Combines patch attention for fine-grained detail capture with full attention for global dependency modeling.

For detailed architecture explanations, please refer to our paper.

License

This project is licensed under the APache 2.0 License - see the LICENSE file for details.

Acknowledgements

We would like to thank all contributors and collaborators for their input and support.