yryoung / power-plant-forecast

TimesNet for Forecasting Power Plant Output

This project is an forked and modified version of the state-of-the-art TimesNet model for time series analysis.

It is a temporal 2D-variation modeling approach for general time series analysis:

• It extends the analysis of temporal variations into the 2D space by transforming the 1D time series into a matrix, which encodes temporal features from multiple perspectives.
• It uses a pyramid-like architecture, TimesBlock, to hierarchically capture the complex temporal variations.
• The experiments on a variety of tasks demonstrate that TimesNet outperforms existing methods in terms of accuracy and generalization capability on datasets with diverse scales and complexities.

Original Repository is here

🛠️ Improvements

• Optimized and simplified the project structure from the original library.
• Implemented Neptune for experiment tracking.
• Enhanced dataset and data loader flexibility to handle tasks with time gaps
• Implemented parallel loading and improved data structures to enhance data processing efficiency.
• Identified and resolved a minor bug in dataset length calculation, resulting in an increased dataset size.
• Integrated additional sub-networks to process time and temperature information for prediction (e.i., y_data).
• Accelerated the training process by utilizing PyTorch built-in automatic mixed precision training and asynchronous GPU data copying when a GPU is available.

🌟 Features

• Mean Squared Error (MSE) of minute-level prediction <= 0.11 (standardized data)
• Precise and adjustable minute-level predictions
• Fast Inference: approximately 5s/32samples (each sample contains 4 power output within 1 hour) on Google Colab CPU

Installation

1. Install the required dependencies:

   pip install -r requirements.txt
   

2. (Optional) Config the neptune.yaml file for Neptune.ai tracking:

   project: your_username/your_project_name
   api_token: your_api_token
   

Usage

to view the help message:

python -u run.py --help

to run the model:

python -u run.py --args args

Visit documentation to learn more about the run.py arguments.

Alternatively, check [Data Preprocessing](Tutorials/Data Prepocessing.ipynb), [Train, Test and Predict](Tutorials/Train, Test and Predict.ipynb) and [Visualization](Tutorials/Testing Result Visualization.ipynb) notebooks in the /Tutorials folder for more examples.

Future Improvements

• Implement transfer learning to enhance inference speed and potentially gain insights into model interpretation.
• Utilize linear interpolation to augment the number of data samples.
• Enhance data preprocessing techniques:
  • Apply sliding window for further denoising.
  • Experiment with and potentially combine different data scalers.
• Incorporate the characteristic of power threshold in the model input, and if possible, use a combination of hinge loss and mean squared error (MSE) as the loss function.

Contact

If you have any questions or suggestions, feel free to describe it in Issues.

Credits

• Original TimesNet model
• Time-Series-Library