Benchmarking Transformers vs. CatBoost on Hourly Loads

What it does This repository packages a comparative analysis pipeline for hourly load forecasting, bundling data preprocessing, model training, validation, and visualization around four algorithms: PatchTST, TFT, N-HiTS, and CatBoost. The authors describe it as an information system aimed at energy-market decision support, though the README offers only a high-level summary without exposing architecture diagrams or sample outputs.

The interesting bit The authors are unusually direct about wanting industrial adoption, openly inviting energy companies and data centers to pilot the models in production. That ambition is admirable, but the README’s dead English-localization links and absence of concrete metrics make it hard to tell whether the codebase is pilot-ready or still a research prototype.

Key highlights

• Pits three deep-learning time-series architectures (PatchTST, TFT, N-HiTS) against CatBoost gradient boosting.
• Claims to cover the full workflow: preprocessing, training, validation, and visualization.
• Explicitly targets energy-market decision support rather than pure algorithmic benchmarking.
• Authors solicit real-world pilots with utilities, universities, and manufacturing firms.
• Badges indicate a Python 3.11 / PyTorch 2.0 / Pandas stack, plus Darts-ML.

Caveats

• The README is almost entirely in Russian; English and Chinese localization links are dead placeholders.
• No benchmark results, dataset descriptions, or methodological details are shown in the README.
• It is unclear from the summary how much of the repository is reusable framework versus a single experiment.

Verdict Energy-sector data scientists looking for a structured comparison template might find a starting point here, but researchers needing reproducible methodology or English documentation should look elsewhere for now.