gghans726/notos-wind-multimodal-data

NOTOS Wind Multimodal Forecasting Dataset

A multimodal wind-power forecasting dataset built from 131 onshore and offshore wind farms. Every sample pairs a normalized power-history window with a forecasting target, a spectrogram image of the history, a symbolic trend string, and a natural-language description of the wind farm.

Overview

Part	Farms	Purpose
pretrain	99	Pre-training
test	32	Evaluation (each farm is an independent test case)

Repository layout

Each farm is stored as two archives:

pretrain/<farm>.json.tar.zst   # meta.json, meta_prompt.txt and sample JSONs (zstd-compressed tar)
pretrain/<farm>.image.tar      # spectrogram PNGs (plain tar)
test/<farm>.json.tar.zst
test/<farm>.image.tar

Uncompressed size: about 449 GiB of JSON and 544 GiB of PNG images.

Download and extraction

# Full dataset
hf download gghans726/notos-wind-multimodal-data --repo-type dataset --local-dir wind_multimodal_data_packed

# Or a single farm
hf download gghans726/notos-wind-multimodal-data --repo-type dataset --local-dir wind_multimodal_data_packed --include "test/BOWLW-1.*"

# Extract everything into one folder (requires zstd)
mkdir -p wind_multimodal_data
for f in wind_multimodal_data_packed/*/*.json.tar.zst; do tar -I zstd -xf "$f" -C wind_multimodal_data; done
for f in wind_multimodal_data_packed/*/*.image.tar; do tar -xf "$f" -C wind_multimodal_data; done

Both archive types store paths relative to the dataset root, so extracting them into the same folder restores the full layout:

wind_multimodal_data/
├── pretrain/<farm>/
│   ├── meta.json
│   ├── meta_prompt.txt
│   ├── json/<farm>_<index>.json
│   └── image/<farm>_<index>.png
└── test/<farm>/
    ├── meta.json
    ├── meta_prompt.txt
    └── {train,val,test}/
        ├── json/<farm>_<index>.json
        └── image/<farm>_<index>.png

The image of a sample has the same file stem as its JSON file and sits in the sibling image/ folder.

Sample format

{
  "description": {
    "dataset_name": "BOWLW-1",
    "history_time_range": "19/06/25 18:15 to 19/06/28 02:00",
    "prediction_steps": ...,
    "spectrogram_path": "..."
  },
  "farm_description": "Natural-language description of the wind farm ...",
  "statistics": {"farm_mean": ..., "farm_std": ..., "window_mean": ..., "window_std": ..., "max": ..., "min": ...},
  "symbolic_trend": "LsLlMnMnMs...",
  "history": [{"time": "19/06/25 18:15", "value": -0.657914}, ...],
  "prediction": [{"time": "...", "value": ...}, ...]
}

• history: normalized power history window at 15-minute resolution.
• prediction: normalized forecasting target following the history window.
• statistics: normalization statistics of the farm and of the window.
• symbolic_trend: symbolic encoding of the history trend.
• spectrogram_path: absolute path on the original build machine; use the same-stem PNG in the sibling image/ folder instead.
• Images: 224 × 224 RGB spectrograms of the history window.

Each farm folder also contains meta.json (farm metadata such as rated power, location, turbine model, climate and terrain; some keys are in Chinese) and meta_prompt.txt (the English farm description used in the prompts).

Sources and licensing

The dataset is derived from the following public data. Please also comply with the terms of the original providers and cite them:

• NREL WIND Toolkit (National Renewable Energy Laboratory): Draxl, C., Clifton, A., Hodge, B.-M. & McCaa, J. The Wind Integration National Dataset (WIND) Toolkit. Applied Energy 151, 355–366 (2015). https://doi.org/10.1016/j.apenergy.2015.03.015
• ENTSO-E Transparency Platform (European Network of Transmission System Operators for Electricity): Hirth, L., Mühlenpfordt, J. & Bulkeley, M. The ENTSO-E Transparency Platform – A review of Europe's most ambitious electricity data platform. Applied Energy 225, 1054–1067 (2018). https://doi.org/10.1016/j.apenergy.2018.04.048
• AEMO (Australian Energy Market Operator): National Electricity Market data, https://www.aemo.com.au and https://nemweb.com.au

Citation

If you use this dataset, please cite:

@misc{notos_wind_multimodal_data,
  title     = {NOTOS Wind Multimodal Forecasting Dataset},
  author    = {TBD},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {TBD},
  url       = {https://huggingface.co/datasets/gghans726/notos-wind-multimodal-data}
}

Please also cite the original data sources:

@article{draxl2015wind,
  title   = {The Wind Integration National Dataset (WIND) Toolkit},
  author  = {Draxl, Caroline and Clifton, Andrew and Hodge, Bri-Mathias and McCaa, Jim},
  journal = {Applied Energy},
  volume  = {151},
  pages   = {355--366},
  year    = {2015},
  doi     = {10.1016/j.apenergy.2015.03.015}
}

@article{hirth2018entsoe,
  title   = {The ENTSO-E Transparency Platform -- A review of Europe's most ambitious electricity data platform},
  author  = {Hirth, Lion and M{\"u}hlenpfordt, Jonathan and Bulkeley, Marisa},
  journal = {Applied Energy},
  volume  = {225},
  pages   = {1054--1067},
  year    = {2018},
  doi     = {10.1016/j.apenergy.2018.04.048}
}

@misc{aemo_nem_data,
  title        = {National Electricity Market Data},
  author       = {{Australian Energy Market Operator}},
  howpublished = {\url{https://nemweb.com.au}}
}