Notos: Multimodal Foundation Model for Wind Power Forecasting

Notos forecasts farm-level wind power from three modalities built from the same history window: the time series itself, its STFT spectrogram (CLIP ViT-B/32), and a text prompt (farm description + symbolic trend), fused by a LoRA-tuned Qwen2.5-0.5B decoder with a quantile head (21 quantiles, 0.01-0.99).

Weights	Horizon (15-min steps)	Lead time
weights/pred16	16	4 h
weights/pred32	32	8 h
weights/pred48	48	12 h
weights/pred64	64	16 h

All models share config.json and use 224 history steps (56 h @ 15 min).

Quick start

hf download gghans726/notos --include "inference/*" requirements.txt --local-dir notos && cd notos
pip install -r requirements.txt

python inference/run.py --pred_len 16 \
    --metadata my_site.yaml \
    --history my_history.csv \
    --output pred.csv

config.json and the selected weights/pred{N}/model.safetensors are fetched from this repo on first use if they are not present locally. The base backbones (Qwen/Qwen2.5-0.5B, openai/clip-vit-base-patch32) are fetched from the Hub; for offline use point NOTOS_LLM_PATH / NOTOS_CLIP_PATH to local copies.

Inputs

1. History CSV: a timestamp column and a power column (auto-detected, or set --timestamp-col / --power-col). It must cover at least 224 points after resampling to 15 min; any native interval <= 30 min is resampled automatically.
2. Site metadata YAML (template: inference/site.yaml): sampling interval, power unit (kw / mw / kwh_per_slot), rated power, optional geo / turbine description, and farm-level normalization statistics. All fields are optional; a warning is printed for every field that falls back to a default. For best accuracy, compute the statistics once from weeks-to-months of history and paste the printed block into the YAML:

python inference/calibrate.py --history long_history.csv --metadata my_site.yaml

Output

A CSV with time, predicted_power_kw (median head) and q01_kw ... q99_kw (quantile bands sorted to be monotone), all in kW.

Python API

import sys; sys.path.insert(0, 'inference')
from notos.data import build_features, load_csv
from notos.metadata import load_site
from notos.model import load, predict

model, cfg, tok = load(pred_len=32)
feat = build_features(load_csv('my_history.csv'), load_site('my_site.yaml'), tok, cfg)
df = predict(model, feat)

Data

The pre-training dataset and the test set used by Notos are available at gghans726/notos-wind-multimodal-data.

Repository layout

config.json                     shared model hyperparameters
weights/pred{16,32,48,64}/      model.safetensors
inference/run.py                CLI: history CSV + site YAML -> forecast
inference/calibrate.py          one-time farm statistics calibration
inference/site.yaml             site metadata template
inference/notos/                model and preprocessing code
requirements.txt