TSTransformer: Improved Transformer for Time Series Forecasting

Graduation project implementation of an improved Transformer model for multivariate long-term time series forecasting.

Model

TSTransformer incorporates:

• Multi-scale patch embedding
• Wavelet-based decomposition
• RevIN normalization
• Efficient attention mechanism

Datasets

Supported datasets: ETTh1, ETTh2, ETTm1, ETTm2, Weather, ECL.

ETT and Weather datasets are included in data/. ECL dataset can be downloaded separately.

Usage

pip install -r requirements.txt

# Example: train on ETTh1
bash scripts/run_ETTh1.sh

Structure

├── models/         # TSTransformer model definition
├── layers/         # Attention, embedding, decomp, wavelet layers
├── exp/            # Experiment runner
├── data_provider/  # Dataset loading
├── utils/          # Metrics and tools
├── scripts/        # Training shell scripts
├── data/           # ETT and Weather datasets
└── train.py        # Training entry point