ESG

Most of this README is about data. It walks through fetching and converting Solar-Energy, Electricity, Exchange-rate, Wind, NYC-Bike, and NYC-Taxi sets for both single-step and multi-step forecasting.

ESG is a PyTorch implementation, and the README spends nearly all of its space on dataset preparation for forecasting experiments. The data is split into two categories, single-step and multi-step, with separate instructions for each. The implementation leans on resources from a repository that is acknowledged, and the original authors get thanks for open-sourcing their work.

For single-step forecasting, the Solar-Energy, Electricity, and Exchange-rate datasets come from a multivariate time series data repository. They get uncompressed, converted into h5 files, and moved into the data folder. The Wind dataset is different. It comes from a Kaggle page, and the instructions say to delete a column of zero values, then sum 24 hours per day to arrive at daily energy potential estimates.

For multi-step forecasting, the NYC-Bike and NYC-Taxi datasets download from Google Drive or Baidu Yun. The h5 files move into the data folder, the train, test, and val splits go into the data directory, and commands generate the npz files from the h5 files. As a shortcut, the h5 files for the single-step datasets can be downloaded directly from Google Drive or Baidu Yun and dropped into the data folder, skipping the conversion steps entirely. The acknowledgements close the README with thanks to the repository whose resources made the work possible.

What the README does well is remove the guesswork from data prep. It says where each dataset lives and gives the commands that turn raw files into usable splits, which is often the hardest part of reproducing a forecasting paper.

• Source 1 · Project repository
• Source 2 · README