CycleNetQM: Enhanced Time Series Forecasting

An improved implementation of CycleNet developed as part of the COMP4434 course project.

Contributors

• Wenchang CHAI (CCMKCCMK)
• Zitong SHEN (Qween0fPandora)
• Beichen GUO (TechnicolorGUO)
• Honghe DING (DavidDING21)

Key Improvements

Our enhanced version introduces two major improvements to the original CycleNet architecture:

1. Seasonal Scalar added to the Cycle component Q
2. Step-wise Training Strategy replacing the original joint training approach

Detailed implementation and analysis can be found in our group report.

Performance

Our model CycleNetQM achieves state-of-the-art results on the electricity dataset, outperforming several baseline models including:

• Linear
• LSTM
• GRU
• CycleNet+Linear
• CycleNet+MLP

Getting Started

Prerequisites

The conda environment requirements are identical to the original CycleNet repository.

Pre-trained Models

Pre-trained model checkpoints are available on Google Drive.

Train & Run

To train or test the model:

sh run_main.sh

Modify run_main.sh to:

Change models you want to train using model_names Toggle training/testing with is_training (1 for training & testing, 0 for testing) Change the backbone of CycleNet using model_types

Model variants available:

• CycleNetQQ: Seasonal scalar only
• CycleNetMM: Step-wise training only
• CycleNetQM: Combined approach (our main model)
• CycleNet
• Linear
• LSTM
• GRU

Acknowledgments

This work builds upon the CycleNet architecture developed by ACAT-SCUT. We thank the original authors for their foundational work in time series forecasting.