Koopman Forecasting package

The problem statement can be found here

Baseline solutions

For short-term forecasting, the following algorithm demonstrated the state-of-the-art. The work requires the consistency of the system, that is, the ability to make predictions in both directions. This minimizes the following loss $$ε = λ_{id}ε_{id} + λ_{fwd}ε_{fwd} + λ_{bwd}ε_{bwd} + λ_{con}ε_{con},$$ where $$ε_{id}$$ is a reconstruction loss. $$ε_{fwd}, ε_{bwd}$$ – k steps forward (backward) prediction error. $$ε_{con}$$ – consistency loss.

The implementation and instructure are in the corresponding folder koopmanAE

For long-term forecasting, there was proposed Spectral Methods usage with Koopman theory, and a comparison with Fourier transform is made.

The implementation and instructure are in the corresponding folder from_fourier_to_koopman

Repository structure

.
├── from_fourier_to_koopman
│   ├── examples.py
│   ├── fourier_koopman
│   │   ├── fourier.py
│   │   ├── __init__.py
│   │   └── koopman.py
│   ├── imgs
│   │   ├── fourier_koopman_objectives.png
│   │   └── youtube_thumb.png
│   ├── LICENSE
│   ├── README.rst
│   └── unknown_phase_problem.ipynb
├── koopmanAE
│   ├── driver.py
│   ├── model.py
│   ├── plot
│   │   └── pred_pendulum.png
│   ├── plot_pred_error.py
│   ├── read_dataset.py
│   ├── README.md
│   ├── tools.py
│   ├── training_parms.txt
│   └── train.py
├── peer-reviews
│   ├── first_peer_review_Prophet.pdf
│   ├── Firtst_peer_review_DVT.pdf
│   ├── Firtst_peer_review_Feature_selection.pdf
│   └── README.md
├── README.md
└── reports
    └── first_report.pdf

Roles for the participants (preliminary)

Nikita Balabin (50%) – Refactoring the two methods with PyTorch Lightning as the main framework for the library. Introduction of Ray to optimize all hyperparameters. + Main implementations. Oleg Maslov (50%) – Introduction the unit testing and provide a test coverage of 70% of the codebase. Creation of the necessary documentation for the API using readthedocs. Providing notebooks with examples on how to run each method. + Main implementations.