knl2366/NREL_Sky_Imagery

NREL SRRL Minute-Resolution Sky Imagery Dataset

Homepage	https://huggingface.co/datasets/knl2366/NREL_Sky_Imagery
Paper	Hammond & Korgel (2026), Journal of Data-centric Machine Learning Research
Contact	Joshua E. Hammond (jeh5975@utexas.edu)

Summary

A continuously growing dataset of minute-resolution sky images from the EKO ASI-16 all-sky imager at NREL's Solar Radiation Research Laboratory (SRRL) in Golden, Colorado (39.742°N, 105.180°W, 1829 m). The standard SRRL gallery saves images every 10 minutes; this dataset captures the camera's native one-minute output — 10× higher temporal frequency for training intra-hour solar irradiance forecasting models.

Each image is paired (via daily parquet files) with SRRL Baseline Measurement System (BMS) one-minute meteorological records including GHI, DNI, DHI, and a computed Clear Sky Index.

Attribute	Value
Image resolution	1920 × 1920 px
Temporal frequency	1 minute
Format	JPEG
Camera	EKO ASI-16 All-Sky Imager
Location	NREL SRRL, Golden, CO, USA
Collection start	October 2025
Status	Updated daily
Total images	~170,000+ (growing)
Days covered	170+

Repository Structure

NREL_Sky_Imagery/
├── 2025/
│   └── 10–12/
│       └── DD/
│           └── YYYYMMDD-HHMMSS.jpg      ← raw sky images
├── 2026/
│   └── 01–05/
│       └── DD/
│           └── YYYYMMDD-HHMMSS.jpg
└── parquets/
    ├── 2025/
    │   └── 10–12/
    │       └── weather_images_YYYY_DOY.parquet
    └── 2026/
        └── 01–04/
            └── weather_images_YYYY_DOY.parquet

Images are named YYYYMMDD-HHMMSS.jpg in UTC and organized as YYYY/MM/DD/. Up to 1440 images per day (all 24 hours; nighttime images are dark).

Parquets contain one file per day-of-year. Each row joins a sky image with the nearest BMS weather observation (within ±5 minutes). Only daytime rows with valid meteorological matches are included (~600–720 rows/day depending on season).

Parquet Schema (27 columns)

Column	Type	Description
filename	string	Image filename (e.g., 20260315-131533.jpg)
hf_image_path	string	Full URL to image on HuggingFace
datetime_utc	datetime	Timestamp in UTC
datetime_mst	datetime	Timestamp in MST (UTC−7)
date	datetime	Calendar date
Year	float	Year
DOY	float	Day of year (1–366)
DATE (MM/DD/YYYY)	string	Date string from BMS
MST	string	Time string from BMS (HH:MM)
Global Horizontal [W/m^2]	float	GHI — primary irradiance target
Direct Normal [W/m^2]	float	DNI
Diffuse Horizontal [W/m^2]	float	DHI
Global (secondary) [W/m^2]	float	GHI from secondary pyranometer
Global (uncorr-sec) [W/m^2]	float	Uncorrected secondary GHI
Global (uncorrected) [W/m^2]	float	Uncorrected primary GHI
Direct (uncorrected) [W/m^2]	float	Uncorrected DNI
Diffuse (uncorrected) [W/m^2]	float	Uncorrected DHI
Air Temperature [deg C]	float	Ambient temperature
Rel Humidity [%]	float	Relative humidity
Avg Wind Speed @ 19ft [m/s]	float	Mean wind speed
Peak Wind Speed @ 19ft [m/s]	float	Gust wind speed
Avg Wind Direction @ 19ft [deg from N]	float	Wind direction
Pressure [mBar]	float	Barometric pressure
Precipitation [mm]	float	Precipitation
Zenith Angle [degrees]	float	Solar zenith angle
Azimuth Angle [degrees]	float	Solar azimuth angle
clear_sky_index	float	GHI / pvlib clear-sky GHI (0 when clear-sky < 10 W/m²)

Quick Start

Download the Dataset

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="knl2366/NREL_Sky_Imagery",
    repo_type="dataset",
    local_dir="./nrel_sky_imagery"
)

Load Parquets

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="knl2366/NREL_Sky_Imagery",
    repo_type="dataset",
    filename="parquets/2026/03/weather_images_2026_074.parquet",
)
df = pd.read_parquet(path)
# 721 rows: one per minute of daylight on March 15, 2026
# Each row has image filename, weather data, and clear_sky_index

PyTorch DataLoader

from srrl_dataset import SRRLDataset
from torch.utils.data import DataLoader

dataset = SRRLDataset(
    image_dir="./nrel_sky_imagery",
    meteo_dir="./nrel_sky_imagery/meteorological",
    seq_length=5,          # 5 consecutive images
    forecast_horizon=15,   # predict 15 min ahead
    target_variable="Global CMP22 (vent/cor) [W/m^2]",
    image_size=(224, 224),
)
loader = DataLoader(dataset, batch_size=32, shuffle=True)

for images, meteo_features, target_ghi in loader:
    # images: (batch, 5, 3, 224, 224)
    # meteo_features: (batch, 5, num_meteo_vars)
    # target_ghi: (batch,) — GHI at t+15min
    pass

Data Collection

An automated Python scraper on Indiana University's Jetstream2 cloud (m3.tiny VM) polls the NREL MIDC real-time ASI-16 endpoint every 60 seconds. Duplicate frames are filtered via MD5 hash. Images are uploaded to HuggingFace in daily batches with SHA256 deduplication. Parquets are generated by joining image timestamps with BMS weather data using a nearest-match merge (±5 min tolerance) and computing clear sky index via pvlib.

Source: EKO ASI-16 camera operated by NREL at SRRL. BMS meteorological data from NREL's Baseline Measurement System (operational since 1981). Full BMS variable list: https://midcdmz.nrel.gov/srrl_bms/

Comparison with Related Datasets

Dataset	Resolution	Freq	Irradiance	Meteo	Period	Growing
This dataset	1920×1920	1 min	GHI/DNI/DHI	17 vars + CSI	2025–	Yes
SRRL Gallery	1536×1536	10 min	GHI/DNI/DHI	130+ vars	2017–	Yes
SKIPP'D	2048×2048	1 min	PV power only	None	2017–19	No
Folsom	~1536×1536	1 min	GHI/DNI	Limited	2014–16	No
SIRTA	768×1024	1–2 min	GHI/DHI/DNI	Yes	2017–19	No

Intended Uses

• Intra-hour solar irradiance forecasting (1–30 min horizons)
• Cloud dynamics and motion estimation
• Benchmarking 1-min vs 10-min temporal resolution
• Transfer learning across geographic sites

Limitations

• Gaps are permanent: NREL serves only the current frame; missed images cannot be recovered.
• Nighttime images: Raw image directories include dark frames. Parquets contain only daytime rows.
• Dome obstructions: Occasional moisture/frost on the ASI-16 dome produces partially obscured images.
• Single site: Golden, CO (semi-arid, 1829 m). May require transfer learning for other climates.
• Parquet coverage: 143 daily parquets (Oct 2025–Apr 2026); image directories extend through May 2026.

Maintenance

The dataset grows daily via automated pipeline. Maintained by Joshua E. Hammond at The University of Texas at Austin. Infrastructure runs on NSF-funded Jetstream2 (Indiana University).

Citation

@article{hammond2026srrl,
  title={A Minute-Resolution Sky Imagery Dataset from NREL's Solar Radiation
         Research Laboratory for Intra-Hour Solar Irradiance Forecasting},
  author={Hammond, Joshua E. and Korgel, Brian A.},
  journal={Journal of Data-centric Machine Learning Research},
  year={2026}
}

License

CC-BY 4.0

Contact

• Maintainer: Joshua E. Hammond (jeh5975@utexas.edu)
• PI: Brian A. Korgel (korgel@che.utexas.edu)
• Affiliation: McKetta Department of Chemical Engineering, The University of Texas at Austin